diff --git a/.gitattributes b/.gitattributes
new file mode 100644
index 0000000000000000000000000000000000000000..a6344aac8c09253b3b630fb776ae94478aa0275b
--- /dev/null
+++ b/.gitattributes
@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
diff --git a/1280ms/.DS_Store b/1280ms/.DS_Store
new file mode 100644
index 0000000000000000000000000000000000000000..6c0472dd5f7cb37790e42abfe607b3ecf31df01e
Binary files /dev/null and b/1280ms/.DS_Store differ
diff --git a/1280ms/convert_parakeet_eou.py b/1280ms/convert_parakeet_eou.py
new file mode 100644
index 0000000000000000000000000000000000000000..9c865f90c8132840b5755dcd62fcfac15e3f0352
--- /dev/null
+++ b/1280ms/convert_parakeet_eou.py
@@ -0,0 +1,740 @@
+#!/usr/bin/env python3
+"""CLI for exporting Parakeet Realtime EOU 120M components to CoreML.
+
+This model is a cache-aware streaming FastConformer-RNNT model optimized for
+low-latency speech recognition with end-of-utterance detection.
+
+Key differences from Parakeet TDT v3:
+- Smaller model (120M vs 600M params)
+- No duration outputs (standard RNNT, not TDT)
+- Cache-aware streaming encoder (17 layers, attention context [70,1])
+- Special <EOU> token for end-of-utterance detection
+- Optimized for 80-160ms latency
+
+Reference: https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1
+"""
+from __future__ import annotations
+
+import json
+from dataclasses import asdict
+from pathlib import Path
+from typing import Dict, Optional, Tuple
+
+import coremltools as ct
+import numpy as np
+import soundfile as sf
+import torch
+import typer
+
+import nemo.collections.asr as nemo_asr
+
+from individual_components import (
+    DecoderWrapper,
+    EncoderWrapper,
+    ExportSettings,
+    JointWrapper,
+    JointDecisionWrapper,
+    JointDecisionSingleStep,
+    PreprocessorWrapper,
+    MelEncoderWrapper,
+    _coreml_convert,
+)
+
+def apply_stft_patch():
+    # Monkey patch coremltools.stft to handle extra arguments from newer torch versions
+    try:
+        import coremltools.converters.mil.frontend.torch.ops as torch_ops
+        _original_stft = torch_ops.stft
+
+        def patched_stft(context, node):
+            if len(node.inputs) > 8:
+                node.inputs = node.inputs[:8]
+            return _original_stft(context, node)
+
+        torch_ops.stft = patched_stft
+        if "stft" in torch_ops._TORCH_OPS_REGISTRY:
+            torch_ops._TORCH_OPS_REGISTRY["stft"] = patched_stft
+        print("Monkey patched coremltools.stft for compatibility.")
+    except Exception as e:
+        print(f"Warning: Could not monkey patch stft: {e}")
+
+DEFAULT_MODEL_ID = "nvidia/parakeet_realtime_eou_120m-v1"
+AUTHOR = "Fluid Inference"
+
+
+def _compute_length(seconds: float, sample_rate: int) -> int:
+    return int(round(seconds * sample_rate))
+
+
+def _prepare_audio(
+    validation_audio: Optional[Path],
+    sample_rate: int,
+    max_samples: int,
+    seed: Optional[int],
+) -> torch.Tensor:
+    if validation_audio is None:
+        if seed is not None:
+            torch.manual_seed(seed)
+        audio = torch.randn(1, max_samples, dtype=torch.float32)
+        return audio
+
+    data, sr = sf.read(str(validation_audio), dtype="float32")
+    if sr != sample_rate:
+        raise typer.BadParameter(
+            f"Validation audio sample rate {sr} does not match model rate {sample_rate}"
+        )
+
+    if data.ndim > 1:
+        data = data[:, 0]
+
+    if data.size == 0:
+        raise typer.BadParameter("Validation audio is empty")
+
+    if data.size < max_samples:
+        pad_width = max_samples - data.size
+        data = np.pad(data, (0, pad_width))
+    elif data.size > max_samples:
+        data = data[:max_samples]
+
+    audio = torch.from_numpy(data).unsqueeze(0).to(dtype=torch.float32)
+    return audio
+
+
+def _save_mlpackage(model: ct.models.MLModel, path: Path, description: str) -> None:
+    try:
+        model.minimum_deployment_target = ct.target.iOS17
+    except Exception:
+        pass
+    model.short_description = description
+    model.author = AUTHOR
+    path.parent.mkdir(parents=True, exist_ok=True)
+    model.save(str(path))
+
+
+def _tensor_shape(tensor: torch.Tensor) -> Tuple[int, ...]:
+    return tuple(int(dim) for dim in tensor.shape)
+
+
+def _parse_compute_units(name: str) -> ct.ComputeUnit:
+    """Parse a human-friendly compute units string into ct.ComputeUnit."""
+    normalized = str(name).strip().upper()
+    mapping = {
+        "ALL": ct.ComputeUnit.ALL,
+        "CPU_ONLY": ct.ComputeUnit.CPU_ONLY,
+        "CPU_AND_GPU": ct.ComputeUnit.CPU_AND_GPU,
+        "CPU_AND_NE": ct.ComputeUnit.CPU_AND_NE,
+        "CPU_AND_NEURALENGINE": ct.ComputeUnit.CPU_AND_NE,
+    }
+    if normalized not in mapping:
+        raise typer.BadParameter(
+            f"Unknown compute units '{name}'. Choose from: " + ", ".join(mapping.keys())
+        )
+    return mapping[normalized]
+
+
+def _parse_compute_precision(name: Optional[str]) -> Optional[ct.precision]:
+    """Parse compute precision string into ct.precision or None."""
+    if name is None:
+        return None
+    normalized = str(name).strip().upper()
+    if normalized == "":
+        return None
+    mapping = {
+        "FLOAT32": ct.precision.FLOAT32,
+        "FLOAT16": ct.precision.FLOAT16,
+    }
+    if normalized not in mapping:
+        raise typer.BadParameter(
+            f"Unknown compute precision '{name}'. Choose from: "
+            + ", ".join(mapping.keys())
+        )
+    return mapping[normalized]
+
+
+app = typer.Typer(add_completion=False, pretty_exceptions_show_locals=False)
+
+
+@app.command()
+def convert(
+    nemo_path: Optional[Path] = typer.Option(
+        None,
+        "--nemo-path",
+        exists=True,
+        resolve_path=True,
+        help="Path to parakeet_realtime_eou_120m-v1.nemo checkpoint (skip to auto-download)",
+    ),
+    model_id: str = typer.Option(
+        DEFAULT_MODEL_ID,
+        "--model-id",
+        help="Model identifier to download when --nemo-path is omitted",
+    ),
+    output_dir: Path = typer.Option(
+        Path("parakeet_eou_coreml"),
+        help="Directory where mlpackages and metadata will be written",
+    ),
+    preprocessor_cu: str = typer.Option(
+        "CPU_ONLY",
+        "--preprocessor-cu",
+        help="Compute units for preprocessor (default CPU_ONLY)",
+    ),
+    mel_encoder_cu: str = typer.Option(
+        "CPU_ONLY",
+        "--mel-encoder-cu",
+        help="Compute units for fused mel+encoder (default CPU_ONLY)",
+    ),
+    compute_precision: Optional[str] = typer.Option(
+        None,
+        "--compute-precision",
+        help="Export precision: FLOAT32 (default) or FLOAT16 to shrink non-quantized weights.",
+    ),
+    max_audio_seconds: float = typer.Option(
+        15.0,
+        "--max-audio-seconds",
+        help="Maximum audio duration in seconds for the fixed window export",
+    ),
+    validation_audio: Optional[Path] = typer.Option(
+        None,
+        "--validation-audio",
+        exists=True,
+        resolve_path=True,
+        help="Path to a 16kHz WAV file for tracing (uses random if not provided)",
+    ),
+) -> None:
+    """Export all Parakeet Realtime EOU sub-modules to CoreML.
+
+    This exports the cache-aware streaming FastConformer-RNNT model for
+    low-latency speech recognition with end-of-utterance detection.
+    """
+    export_settings = ExportSettings(
+        output_dir=output_dir,
+        compute_units=ct.ComputeUnit.CPU_ONLY,
+        deployment_target=ct.target.iOS17,
+        compute_precision=_parse_compute_precision(compute_precision),
+        max_audio_seconds=max_audio_seconds,
+        max_symbol_steps=1,
+    )
+
+    typer.echo("Export configuration:")
+    typer.echo(asdict(export_settings))
+
+    output_dir.mkdir(parents=True, exist_ok=True)
+    pre_cu = _parse_compute_units(preprocessor_cu)
+    melenc_cu = _parse_compute_units(mel_encoder_cu)
+
+    if nemo_path is not None:
+        typer.echo(f"Loading NeMo model from {nemo_path}…")
+        # Try loading as generic ASRModel first, then specific class
+        try:
+            asr_model = nemo_asr.models.ASRModel.restore_from(
+                str(nemo_path), map_location="cpu"
+            )
+        except Exception:
+            # Fallback to EncDecRNNTBPEModel
+            asr_model = nemo_asr.models.EncDecRNNTBPEModel.restore_from(
+                str(nemo_path), map_location="cpu"
+            )
+        checkpoint_meta = {
+            "type": "file",
+            "path": str(nemo_path),
+        }
+    else:
+        typer.echo(f"Downloading NeMo model via {model_id}…")
+        # Use ASRModel.from_pretrained as recommended for this model
+        try:
+            asr_model = nemo_asr.models.ASRModel.from_pretrained(
+                model_id, map_location="cpu"
+            )
+        except Exception:
+            # Fallback to EncDecRNNTBPEModel
+            asr_model = nemo_asr.models.EncDecRNNTBPEModel.from_pretrained(
+                model_id, map_location="cpu"
+            )
+        checkpoint_meta = {
+            "type": "pretrained",
+            "model_id": model_id,
+        }
+    asr_model.eval()
+
+    # Print model info
+    typer.echo(f"Model class: {type(asr_model).__name__}")
+    typer.echo(f"Encoder class: {type(asr_model.encoder).__name__}")
+
+    sample_rate = int(asr_model.cfg.preprocessor.sample_rate)
+    max_samples = _compute_length(export_settings.max_audio_seconds, sample_rate)
+
+    # Prepare audio for tracing
+    if validation_audio is not None:
+        typer.echo(f"Using validation audio: {validation_audio}")
+        audio_tensor = _prepare_audio(validation_audio, sample_rate, max_samples, seed=None)
+    else:
+        typer.echo("Using random audio for tracing (seed=42)")
+        audio_tensor = _prepare_audio(None, sample_rate, max_samples, seed=42)
+
+    audio_length = torch.tensor([max_samples], dtype=torch.int32)
+
+    preprocessor = PreprocessorWrapper(asr_model.preprocessor.eval())
+    encoder = EncoderWrapper(asr_model.encoder.eval())
+    decoder = DecoderWrapper(asr_model.decoder.eval())
+    joint = JointWrapper(asr_model.joint.eval())
+
+    decoder_export_flag = getattr(asr_model.decoder, "_rnnt_export", False)
+    asr_model.decoder._rnnt_export = True
+
+    try:
+        with torch.no_grad():
+            mel_ref, mel_length_ref = preprocessor(audio_tensor, audio_length)
+            mel_length_ref = mel_length_ref.to(dtype=torch.int32)
+            encoder_ref, encoder_length_ref, frame_times_ref = encoder(
+                mel_ref, mel_length_ref
+            )
+            encoder_length_ref = encoder_length_ref.to(dtype=torch.int32)
+
+            # Clone tensors to drop inference flags
+            mel_ref = mel_ref.clone().detach()
+            mel_length_ref = mel_length_ref.clone().detach()
+            encoder_ref = encoder_ref.clone().detach()
+            encoder_length_ref = encoder_length_ref.clone().detach()
+            frame_times_ref = frame_times_ref.clone().detach()
+
+        vocab_size = int(asr_model.tokenizer.vocab_size)
+        decoder_hidden = int(asr_model.decoder.pred_hidden)
+        decoder_layers = int(asr_model.decoder.pred_rnn_layers)
+
+        # Check if model has extra outputs (TDT-style duration)
+        num_extra = getattr(asr_model.joint, "num_extra_outputs", 0)
+        typer.echo(f"Vocab size: {vocab_size}, num_extra_outputs: {num_extra}")
+
+        targets = torch.full(
+            (1, export_settings.max_symbol_steps),
+            fill_value=asr_model.decoder.blank_idx,
+            dtype=torch.int32,
+        )
+        target_lengths = torch.tensor(
+            [export_settings.max_symbol_steps], dtype=torch.int32
+        )
+        zero_state = torch.zeros(
+            decoder_layers,
+            1,
+            decoder_hidden,
+            dtype=torch.float32,
+        )
+
+        with torch.no_grad():
+            decoder_ref, h_ref, c_ref = decoder(
+                targets, target_lengths, zero_state, zero_state
+            )
+            joint_ref = joint(encoder_ref, decoder_ref)
+
+        decoder_ref = decoder_ref.clone()
+        h_ref = h_ref.clone()
+        c_ref = c_ref.clone()
+        joint_ref = joint_ref.clone()
+
+        typer.echo(f"Encoder output shape: {encoder_ref.shape}")
+        typer.echo(f"Decoder output shape: {decoder_ref.shape}")
+        typer.echo(f"Joint output shape: {joint_ref.shape}")
+
+        # === Export Preprocessor ===
+        typer.echo("Tracing and converting preprocessor…")
+        preprocessor = preprocessor.cpu()
+        audio_tensor = audio_tensor.cpu()
+        audio_length = audio_length.cpu()
+        traced_preprocessor = torch.jit.trace(
+            preprocessor, (audio_tensor, audio_length), strict=False
+        )
+        traced_preprocessor.eval()
+        preprocessor_inputs = [
+            ct.TensorType(
+                name="audio_signal",
+                shape=(1, ct.RangeDim(1, max_samples)),
+                dtype=np.float32,
+            ),
+            ct.TensorType(name="audio_length", shape=(1,), dtype=np.int32),
+        ]
+        preprocessor_outputs = [
+            ct.TensorType(name="mel", dtype=np.float32),
+            ct.TensorType(name="mel_length", dtype=np.int32),
+        ]
+        preprocessor_model = _coreml_convert(
+            traced_preprocessor,
+            preprocessor_inputs,
+            preprocessor_outputs,
+            export_settings,
+            compute_units_override=pre_cu,
+        )
+        preprocessor_path = output_dir / "parakeet_eou_preprocessor.mlpackage"
+        _save_mlpackage(
+            preprocessor_model,
+            preprocessor_path,
+            f"Parakeet EOU preprocessor ({max_audio_seconds}s window)",
+        )
+
+        # === Export Encoder ===
+        typer.echo("Tracing and converting encoder…")
+        traced_encoder = torch.jit.trace(
+            encoder, (mel_ref, mel_length_ref), strict=False
+        )
+        traced_encoder.eval()
+        encoder_inputs = [
+            ct.TensorType(
+                name="mel", shape=_tensor_shape(mel_ref), dtype=np.float32
+            ),
+            ct.TensorType(name="mel_length", shape=(1,), dtype=np.int32),
+        ]
+        encoder_outputs = [
+            ct.TensorType(name="encoder", dtype=np.float32),
+            ct.TensorType(name="encoder_length", dtype=np.int32),
+            ct.TensorType(name="frame_times", dtype=np.float32),
+        ]
+        encoder_model = _coreml_convert(
+            traced_encoder,
+            encoder_inputs,
+            encoder_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        encoder_path = output_dir / "parakeet_eou_encoder.mlpackage"
+        _save_mlpackage(
+            encoder_model,
+            encoder_path,
+            f"Parakeet EOU encoder ({max_audio_seconds}s window)",
+        )
+
+        # === Export Fused Mel+Encoder ===
+        typer.echo("Tracing and converting fused mel+encoder…")
+        mel_encoder = MelEncoderWrapper(preprocessor, encoder)
+        traced_mel_encoder = torch.jit.trace(
+            mel_encoder, (audio_tensor, audio_length), strict=False
+        )
+        traced_mel_encoder.eval()
+        mel_encoder_inputs = [
+            ct.TensorType(
+                name="audio_signal", shape=(1, max_samples), dtype=np.float32
+            ),
+            ct.TensorType(name="audio_length", shape=(1,), dtype=np.int32),
+        ]
+        mel_encoder_outputs = [
+            ct.TensorType(name="encoder", dtype=np.float32),
+            ct.TensorType(name="encoder_length", dtype=np.int32),
+            ct.TensorType(name="frame_times", dtype=np.float32),
+        ]
+        mel_encoder_model = _coreml_convert(
+            traced_mel_encoder,
+            mel_encoder_inputs,
+            mel_encoder_outputs,
+            export_settings,
+            compute_units_override=melenc_cu,
+        )
+        mel_encoder_path = output_dir / "parakeet_eou_mel_encoder.mlpackage"
+        _save_mlpackage(
+            mel_encoder_model,
+            mel_encoder_path,
+            f"Parakeet EOU fused Mel+Encoder ({max_audio_seconds}s window)",
+        )
+
+        # === Export Decoder ===
+        typer.echo("Tracing and converting decoder…")
+        traced_decoder = torch.jit.trace(
+            decoder,
+            (targets, target_lengths, zero_state, zero_state),
+            strict=False,
+        )
+        traced_decoder.eval()
+        decoder_inputs = [
+            ct.TensorType(
+                name="targets", shape=_tensor_shape(targets), dtype=np.int32
+            ),
+            ct.TensorType(name="target_length", shape=(1,), dtype=np.int32),
+            ct.TensorType(
+                name="h_in", shape=_tensor_shape(zero_state), dtype=np.float32
+            ),
+            ct.TensorType(
+                name="c_in", shape=_tensor_shape(zero_state), dtype=np.float32
+            ),
+        ]
+        decoder_outputs = [
+            ct.TensorType(name="decoder", dtype=np.float32),
+            ct.TensorType(name="h_out", dtype=np.float32),
+            ct.TensorType(name="c_out", dtype=np.float32),
+        ]
+        decoder_model = _coreml_convert(
+            traced_decoder,
+            decoder_inputs,
+            decoder_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        decoder_path = output_dir / "parakeet_eou_decoder.mlpackage"
+        _save_mlpackage(
+            decoder_model,
+            decoder_path,
+            "Parakeet EOU decoder (RNNT prediction network)",
+        )
+
+        # === Export Joint ===
+        typer.echo("Tracing and converting joint…")
+        traced_joint = torch.jit.trace(
+            joint,
+            (encoder_ref, decoder_ref),
+            strict=False,
+        )
+        traced_joint.eval()
+        joint_inputs = [
+            ct.TensorType(
+                name="encoder", shape=_tensor_shape(encoder_ref), dtype=np.float32
+            ),
+            ct.TensorType(
+                name="decoder", shape=_tensor_shape(decoder_ref), dtype=np.float32
+            ),
+        ]
+        joint_outputs = [
+            ct.TensorType(name="logits", dtype=np.float32),
+        ]
+        joint_model = _coreml_convert(
+            traced_joint,
+            joint_inputs,
+            joint_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        joint_path = output_dir / "parakeet_eou_joint.mlpackage"
+        _save_mlpackage(
+            joint_model,
+            joint_path,
+            "Parakeet EOU joint network (RNNT)",
+        )
+
+        # === Export Joint Decision Head ===
+        typer.echo("Tracing and converting joint decision head…")
+        joint_decision = JointDecisionWrapper(joint, vocab_size=vocab_size)
+        traced_joint_decision = torch.jit.trace(
+            joint_decision,
+            (encoder_ref, decoder_ref),
+            strict=False,
+        )
+        traced_joint_decision.eval()
+        joint_decision_inputs = [
+            ct.TensorType(
+                name="encoder", shape=_tensor_shape(encoder_ref), dtype=np.float32
+            ),
+            ct.TensorType(
+                name="decoder", shape=_tensor_shape(decoder_ref), dtype=np.float32
+            ),
+        ]
+        joint_decision_outputs = [
+            ct.TensorType(name="token_id", dtype=np.int32),
+            ct.TensorType(name="token_prob", dtype=np.float32),
+        ]
+        joint_decision_model = _coreml_convert(
+            traced_joint_decision,
+            joint_decision_inputs,
+            joint_decision_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        joint_decision_path = output_dir / "parakeet_eou_joint_decision.mlpackage"
+        _save_mlpackage(
+            joint_decision_model,
+            joint_decision_path,
+            "Parakeet EOU joint + decision head (softmax, argmax)",
+        )
+
+        # === Export Single-Step Joint Decision ===
+        typer.echo("Tracing and converting single-step joint decision…")
+        jd_single = JointDecisionSingleStep(joint, vocab_size=vocab_size)
+        # Create single-step slices from refs
+        enc_step = encoder_ref[:, :, :1].contiguous()
+        dec_step = decoder_ref[:, :, :1].contiguous()
+        traced_jd_single = torch.jit.trace(
+            jd_single,
+            (enc_step, dec_step),
+            strict=False,
+        )
+        traced_jd_single.eval()
+        jd_single_inputs = [
+            ct.TensorType(
+                name="encoder_step",
+                shape=(1, enc_step.shape[1], 1),
+                dtype=np.float32,
+            ),
+            ct.TensorType(
+                name="decoder_step",
+                shape=(1, dec_step.shape[1], 1),
+                dtype=np.float32,
+            ),
+        ]
+        jd_single_outputs = [
+            ct.TensorType(name="token_id", dtype=np.int32),
+            ct.TensorType(name="token_prob", dtype=np.float32),
+            ct.TensorType(name="top_k_ids", dtype=np.int32),
+            ct.TensorType(name="top_k_logits", dtype=np.float32),
+        ]
+        jd_single_model = _coreml_convert(
+            traced_jd_single,
+            jd_single_inputs,
+            jd_single_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        jd_single_path = output_dir / "parakeet_eou_joint_decision_single_step.mlpackage"
+        _save_mlpackage(
+            jd_single_model,
+            jd_single_path,
+            "Parakeet EOU single-step joint decision (current frame)",
+        )
+
+        # === Save Metadata ===
+        metadata: Dict[str, object] = {
+            "model_id": model_id,
+            "model_name": "parakeet_realtime_eou_120m-v1",
+            "model_class": type(asr_model).__name__,
+            "encoder_class": type(asr_model.encoder).__name__,
+            "sample_rate": sample_rate,
+            "max_audio_seconds": export_settings.max_audio_seconds,
+            "max_audio_samples": max_samples,
+            "max_symbol_steps": export_settings.max_symbol_steps,
+            "vocab_size": vocab_size,
+            "vocab_with_blank": vocab_size + 1,
+            "decoder_hidden": decoder_hidden,
+            "decoder_layers": decoder_layers,
+            "num_extra_outputs": num_extra,
+            "has_eou_token": True,
+            "checkpoint": checkpoint_meta,
+            "coreml": {
+                "compute_units": export_settings.compute_units.name,
+                "compute_precision": (
+                    export_settings.compute_precision.name
+                    if export_settings.compute_precision is not None
+                    else "FLOAT32"
+                ),
+            },
+            "components": {
+                "preprocessor": {
+                    "inputs": {
+                        "audio_signal": [1, max_samples],
+                        "audio_length": [1],
+                    },
+                    "outputs": {
+                        "mel": list(_tensor_shape(mel_ref)),
+                        "mel_length": [1],
+                    },
+                    "path": preprocessor_path.name,
+                },
+                "encoder": {
+                    "inputs": {
+                        "mel": list(_tensor_shape(mel_ref)),
+                        "mel_length": [1],
+                    },
+                    "outputs": {
+                        "encoder": list(_tensor_shape(encoder_ref)),
+                        "encoder_length": [1],
+                        "frame_times": [1, _tensor_shape(encoder_ref)[2]],
+                    },
+                    "path": encoder_path.name,
+                },
+                "mel_encoder": {
+                    "inputs": {
+                        "audio_signal": [1, max_samples],
+                        "audio_length": [1],
+                    },
+                    "outputs": {
+                        "encoder": list(_tensor_shape(encoder_ref)),
+                        "encoder_length": [1],
+                        "frame_times": [1, _tensor_shape(encoder_ref)[2]],
+                    },
+                    "path": mel_encoder_path.name,
+                },
+                "decoder": {
+                    "inputs": {
+                        "targets": list(_tensor_shape(targets)),
+                        "target_length": [1],
+                        "h_in": list(_tensor_shape(zero_state)),
+                        "c_in": list(_tensor_shape(zero_state)),
+                    },
+                    "outputs": {
+                        "decoder": list(_tensor_shape(decoder_ref)),
+                        "h_out": list(_tensor_shape(h_ref)),
+                        "c_out": list(_tensor_shape(c_ref)),
+                    },
+                    "path": decoder_path.name,
+                },
+                "joint": {
+                    "inputs": {
+                        "encoder": list(_tensor_shape(encoder_ref)),
+                        "decoder": list(_tensor_shape(decoder_ref)),
+                    },
+                    "outputs": {
+                        "logits": list(_tensor_shape(joint_ref)),
+                    },
+                    "path": joint_path.name,
+                },
+                "joint_decision": {
+                    "inputs": {
+                        "encoder": list(_tensor_shape(encoder_ref)),
+                        "decoder": list(_tensor_shape(decoder_ref)),
+                    },
+                    "outputs": {
+                        "token_id": [
+                            _tensor_shape(encoder_ref)[0],
+                            _tensor_shape(encoder_ref)[2],
+                            _tensor_shape(decoder_ref)[2],
+                        ],
+                        "token_prob": [
+                            _tensor_shape(encoder_ref)[0],
+                            _tensor_shape(encoder_ref)[2],
+                            _tensor_shape(decoder_ref)[2],
+                        ],
+                    },
+                    "path": joint_decision_path.name,
+                },
+                "joint_decision_single_step": {
+                    "inputs": {
+                        "encoder_step": [1, _tensor_shape(encoder_ref)[1], 1],
+                        "decoder_step": [1, _tensor_shape(decoder_ref)[1], 1],
+                    },
+                    "outputs": {
+                        "token_id": [1, 1, 1],
+                        "token_prob": [1, 1, 1],
+                        "top_k_ids": [1, 1, 1, 64],
+                        "top_k_logits": [1, 1, 1, 64],
+                    },
+                    "path": jd_single_path.name,
+                },
+            },
+        }
+
+        # Export tokenizer vocab if available
+        try:
+            tokenizer = asr_model.tokenizer
+            vocab = {
+                "blank_id": int(asr_model.decoder.blank_idx),
+                "vocab_size": vocab_size,
+            }
+            # Try to get special tokens
+            if hasattr(tokenizer, "tokenizer"):
+                inner_tokenizer = tokenizer.tokenizer
+                if hasattr(inner_tokenizer, "get_vocab"):
+                    full_vocab = inner_tokenizer.get_vocab()
+                    # Find EOU token
+                    eou_token = None
+                    for token, idx in full_vocab.items():
+                        if "<EOU>" in token.upper() or "eou" in token.lower():
+                            eou_token = {"token": token, "id": idx}
+                            break
+                    if eou_token:
+                        vocab["eou_token"] = eou_token
+            metadata["tokenizer"] = vocab
+        except Exception as e:
+            typer.echo(f"Warning: Could not export tokenizer info: {e}")
+
+        metadata_path = output_dir / "metadata.json"
+        metadata_path.write_text(json.dumps(metadata, indent=2))
+        typer.echo(f"\nExport complete. Metadata written to {metadata_path}")
+        typer.echo(f"Output directory: {output_dir}")
+
+    finally:
+        asr_model.decoder._rnnt_export = decoder_export_flag
+
+
+if __name__ == "__main__":
+    app()
diff --git a/1280ms/convert_streaming_encoder.py b/1280ms/convert_streaming_encoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..a1c199c8a6c540dab48b973ab4232db2901b8e72
--- /dev/null
+++ b/1280ms/convert_streaming_encoder.py
@@ -0,0 +1,193 @@
+
+import torch
+import torch.nn as nn
+import coremltools as ct
+import numpy as np
+import typer
+from pathlib import Path
+from typing import Tuple, List, Optional
+import json
+import shutil
+
+# Iimport torch
+import coremltools as ct
+import numpy as np
+import argparse
+from nemo.collections.asr.models import EncDecRNNTBPEModel
+
+app = typer.Typer()
+
+class LoopbackEncoderWrapper(nn.Module):
+    """
+    Wraps the entire Parakeet Encoder (PreEncode + Conformer) for CoreML Loopback Streaming.
+    
+    Inputs:
+      - audio_signal: [B, D, T] (Mel spectrogram chunk)
+      - audio_length: [B]
+      - pre_cache: [B, D, pre_cache_size] (Previous audio context)
+      - cache_last_channel: [layers, B, cache_size, hidden]
+      - cache_last_time: [layers, B, hidden, time_cache]
+      - cache_last_channel_len: [B]
+      
+    Outputs:
+      - encoded_output: [B, D_out, T_out]
+      - encoded_length: [B]
+      - new_pre_cache: [B, D, pre_cache_size]
+      - new_cache_last_channel
+      - new_cache_last_time
+      - new_cache_last_channel_len
+    """
+    def __init__(self, encoder, pre_cache_size=16):
+        super().__init__()
+        self.encoder = encoder
+        self.pre_cache_size = pre_cache_size
+        
+    def forward(
+        self, 
+        audio_signal: torch.Tensor, 
+        audio_length: torch.Tensor,
+        pre_cache: torch.Tensor,
+        cache_last_channel: torch.Tensor,
+        cache_last_time: torch.Tensor,
+        cache_last_channel_len: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        
+        # 1. Prepend pre_cache to audio_signal
+        # audio_signal: [B, D, T]
+        # pre_cache: [B, D, T_cache]
+        full_input = torch.cat([pre_cache, audio_signal], dim=2)
+        full_length = audio_length + self.pre_cache_size
+        
+        # 2. Extract NEW pre_cache (last N frames of full_input)
+        # Note: We do this BEFORE processing because we want the raw audio context
+        new_pre_cache = full_input[:, :, -self.pre_cache_size:]
+        
+        # 3. Process with Encoder
+        # Reconstruct NeMo cache object
+        current_cache = [cache_last_channel, cache_last_time, cache_last_channel_len]
+        
+        encoded, encoded_len, new_cache_channel, new_cache_time, new_cache_len = self.encoder.cache_aware_stream_step(
+            processed_signal=full_input,
+            processed_signal_length=full_length,
+            cache_last_channel=cache_last_channel,
+            cache_last_time=cache_last_time,
+            cache_last_channel_len=cache_last_channel_len
+        )
+        
+        # 4. Drop the first few frames corresponding to pre_cache?
+        # NeMo's cache_aware_stream_step usually handles the "valid" output frames.
+        # But since we manually prepended, we might get extra output frames.
+        # However, for streaming, we usually want the model to see the context but only output the new tokens.
+        # Let's trust NeMo's streaming logic for now, or check if we need to slice.
+        # Given we are using 'cache_aware_stream_step', it expects the full context window?
+        # Actually, standard usage is: input IS the new chunk, but internal convolution looks at past.
+        # But since we are stateless, we MUST provide the past.
+        # So passing (pre_cache + chunk) is correct.
+        
+        # Cast lengths to Int32 for CoreML
+        encoded_len_32 = encoded_len.to(dtype=torch.int32)
+        new_channel_len_32 = new_cache_len.to(dtype=torch.int32)
+        
+        return encoded, encoded_len_32, new_pre_cache, new_cache_channel, new_cache_time, new_channel_len_32
+
+def _coreml_convert(
+    traced_model,
+    inputs,
+    outputs,
+    compute_units=ct.ComputeUnit.CPU_ONLY
+):
+    return ct.convert(
+        traced_model,
+        inputs=inputs,
+        outputs=outputs,
+        compute_units=compute_units,
+        minimum_deployment_target=ct.target.macOS14,
+    )
+
+def main():
+    model_id: str = "nvidia/parakeet_realtime_eou_120m-v1"
+    output_dir: str = "temp_swift_models/StreamingLoopback"
+    output_path = Path(output_dir)
+    output_path.mkdir(parents=True, exist_ok=True)
+    
+    print(f"Loading model: {model_id}...")
+    asr_model = EncDecRNNTBPEModel.from_pretrained(model_name=model_id)
+    asr_model.eval()
+    
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--chunk-frames", type=int, default=17, help="Number of frames in the input chunk (e.g. 17 for 160ms, 129 for 1.28s)")
+    args = parser.parse_args()
+
+    encoder = asr_model.encoder
+    
+    # --- Configuration ---
+    # 160ms chunk = 16 frames (but preprocessor produces 17 with padding/centering)
+    # 1.28s chunk = 128 frames (preprocessor produces 129)
+    chunk_size_in = args.chunk_frames 
+    mel_dim = 128
+    hidden_dim = encoder.d_model # 512
+    num_layers = len(encoder.layers) # 17
+    
+    # Cache sizes
+    cache_channel_size = 70
+    cache_time_size = 8
+    pre_cache_size = 16
+    
+    print(f"Config: Chunk={chunk_size_in}, Mel={mel_dim}, Hidden={hidden_dim}, Layers={num_layers}")
+    print(f"Cache: Channel={cache_channel_size}, Time={cache_time_size}, Pre={pre_cache_size}")
+
+    # --- Wrapper ---
+    wrapper = LoopbackEncoderWrapper(encoder, pre_cache_size=pre_cache_size)
+    wrapper.eval()
+    
+    # --- Test Inputs (for Tracing) ---
+    batch_size = 1
+    test_mel = torch.randn(batch_size, mel_dim, chunk_size_in)
+    test_mel_len = torch.tensor([chunk_size_in], dtype=torch.int32)
+    test_pre_cache = torch.zeros(batch_size, mel_dim, pre_cache_size)
+    
+    # Initial Cache (Zeros)
+    test_cache_channel = torch.zeros(num_layers, batch_size, cache_channel_size, hidden_dim)
+    test_cache_time = torch.zeros(num_layers, batch_size, hidden_dim, cache_time_size)
+    test_cache_len = torch.zeros(batch_size, dtype=torch.int32) 
+    
+    print("Tracing model...")
+    traced_model = torch.jit.trace(
+        wrapper,
+        (test_mel, test_mel_len, test_pre_cache, test_cache_channel, test_cache_time, test_cache_len),
+        strict=False
+    )
+    
+    # --- CoreML Conversion ---
+    print("Converting to CoreML...")
+    
+    inputs = [
+        ct.TensorType(name="audio_signal", shape=(1, 128, chunk_size_in), dtype=np.float32),
+        ct.TensorType(name="audio_length", shape=(1,), dtype=np.int32),
+        ct.TensorType(name="pre_cache", shape=(1, 128, pre_cache_size), dtype=np.float32),
+        ct.TensorType(name="cache_last_channel", shape=(num_layers, 1, cache_channel_size, hidden_dim), dtype=np.float32),
+        ct.TensorType(name="cache_last_time", shape=(num_layers, 1, hidden_dim, cache_time_size), dtype=np.float32),
+        ct.TensorType(name="cache_last_channel_len", shape=(1,), dtype=np.int32),
+    ]
+    
+    outputs = [
+        ct.TensorType(name="encoded_output", dtype=np.float32),
+        ct.TensorType(name="encoded_length", dtype=np.int32),
+        ct.TensorType(name="new_pre_cache", dtype=np.float32),
+        ct.TensorType(name="new_cache_last_channel", dtype=np.float32),
+        ct.TensorType(name="new_cache_last_time", dtype=np.float32),
+        ct.TensorType(name="new_cache_last_channel_len", dtype=np.int32),
+    ]
+    
+    mlmodel = _coreml_convert(traced_model, inputs, outputs)
+    
+    save_path = output_path / "streaming_encoder.mlpackage"
+    mlmodel.save(str(save_path))
+    print(f"Saved: {save_path}")
+    
+    # Also export Preprocessor, Decoder, Joint for completeness? 
+    # For now, let's assume we reuse the existing ones or export them separately if needed.
+    # But the user asked specifically for the Encoder loopback.
+
+if __name__ == "__main__":
+    main()
diff --git a/1280ms/decoder.mlmodelc/analytics/coremldata.bin b/1280ms/decoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..c6f51dedd2638184c5a2d71512339ceea086abc0
--- /dev/null
+++ b/1280ms/decoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3996975a8cbc1949159c55605b3132b39b2484f51acbd55d796d93c70de02b49
+size 243
diff --git a/1280ms/decoder.mlmodelc/coremldata.bin b/1280ms/decoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..71936e61ffe7697e4af55bae8870d95c1fbb66c5
--- /dev/null
+++ b/1280ms/decoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3ccbff963d8cf07e2be2bd56ea3384a89ea49628922c6bd95ff62e2ae57dc34
+size 497
diff --git a/1280ms/decoder.mlmodelc/metadata.json b/1280ms/decoder.mlmodelc/metadata.json
new file mode 100644
index 0000000000000000000000000000000000000000..3581673bcb062f7e0e1e0efa9f59edc5f218d50e
--- /dev/null
+++ b/1280ms/decoder.mlmodelc/metadata.json
@@ -0,0 +1,118 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Parakeet EOU decoder (RNNT prediction network)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 640 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 640, 1]",
+        "name" : "decoder",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 640]",
+        "name" : "h_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 640]",
+        "name" : "c_out",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.squeeze" : 2,
+      "Ios17.gather" : 1,
+      "Ios17.cast" : 6,
+      "Ios17.lstm" : 1,
+      "Ios17.transpose" : 2,
+      "Identity" : 1,
+      "Ios17.expandDims" : 2
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int16, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1]",
+        "name" : "targets",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "target_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 640]",
+        "name" : "h_in",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 640]",
+        "name" : "c_in",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.4.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "parakeet_eou_decoder",
+    "method" : "predict"
+  }
+]
\ No newline at end of file
diff --git a/1280ms/decoder.mlmodelc/model.mil b/1280ms/decoder.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..5ead417fbdc13f8bf0c05a26039b2828503d4eca
--- /dev/null
+++ b/1280ms/decoder.mlmodelc/model.mil
@@ -0,0 +1,45 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.4.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<fp32, [1, 1, 640]> c_in, tensor<fp32, [1, 1, 640]> h_in, tensor<int32, [1]> target_length, tensor<int32, [1, 1]> targets) {
+            tensor<int32, []> y_axis_0 = const()[name = tensor<string, []>("y_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> y_batch_dims_0 = const()[name = tensor<string, []>("y_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> y_validate_indices_0 = const()[name = tensor<string, []>("y_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1027, 640]> module_prediction_embed_weight_to_fp16 = const()[name = tensor<string, []>("module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [1027, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<string, []> targets_to_int16_dtype_0 = const()[name = tensor<string, []>("targets_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 1]> targets_to_int16 = cast(dtype = targets_to_int16_dtype_0, x = targets)[name = tensor<string, []>("cast_8")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = targets_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = tensor<string, []>("y_cast_fp16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [1]> input_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<string, []> h_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("h_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = tensor<string, []>("cast_7")];
+            tensor<fp16, [1, 640]> input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = h_in_to_fp16)[name = tensor<string, []>("input_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<string, []> c_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("c_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = tensor<string, []>("cast_6")];
+            tensor<fp16, [1, 640]> input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = c_in_to_fp16)[name = tensor<string, []>("input_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> input_direction_0 = const()[name = tensor<string, []>("input_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_output_sequence_0 = const()[name = tensor<string, []>("input_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_recurrent_activation_0 = const()[name = tensor<string, []>("input_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_cell_activation_0 = const()[name = tensor<string, []>("input_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_activation_0 = const()[name = tensor<string, []>("input_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = tensor<string, []>("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1314688)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = tensor<string, []>("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4591552)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = tensor<string, []>("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7868416)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_2")];
+            tensor<fp16, [1, 1, 640]> input_cast_fp16_0, tensor<fp16, [1, 640]> input_cast_fp16_1, tensor<fp16, [1, 640]> input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_0_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [1]> obj_3_axes_0 = const()[name = tensor<string, []>("obj_3_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 640]> obj_3_cast_fp16 = expand_dims(axes = obj_3_axes_0, x = input_cast_fp16_1)[name = tensor<string, []>("obj_3_cast_fp16")];
+            tensor<string, []> obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_3_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> obj_axes_0 = const()[name = tensor<string, []>("obj_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 640]> obj_cast_fp16 = expand_dims(axes = obj_axes_0, x = input_cast_fp16_2)[name = tensor<string, []>("obj_cast_fp16")];
+            tensor<string, []> obj_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = tensor<string, []>("transpose_0_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
+            tensor<string, []> transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("transpose_0_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp16, [1, 640, 1]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = tensor<string, []>("transpose_1")];
+            tensor<fp32, [1, 640, 1]> decoder = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = tensor<string, []>("cast_3")];
+            tensor<fp32, [1, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = tensor<string, []>("cast_4")];
+            tensor<fp32, [1, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = tensor<string, []>("cast_5")];
+            tensor<int32, [1]> target_length_tmp = identity(x = target_length)[name = tensor<string, []>("target_length_tmp")];
+        } -> (decoder, h_out, c_out);
+}
\ No newline at end of file
diff --git a/1280ms/decoder.mlmodelc/weights/weight.bin b/1280ms/decoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..cde1618cbed235b421984a302289e3bdd7e3df02
--- /dev/null
+++ b/1280ms/decoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b4cacecdcd9df79ab1e56de67230baf5a8664d2afe0bb8f3408eefa972cb2f4
+size 7873600
diff --git a/1280ms/individual_components.py b/1280ms/individual_components.py
new file mode 100644
index 0000000000000000000000000000000000000000..47271397bc8d9d17cc0fabcf4bb63be7e7c2109c
--- /dev/null
+++ b/1280ms/individual_components.py
@@ -0,0 +1,250 @@
+#!/usr/bin/env python3
+"""Export Parakeet Realtime EOU 120M RNNT components into CoreML."""
+from __future__ import annotations
+
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Optional, Tuple
+
+import coremltools as ct
+import torch
+
+
+@dataclass
+class ExportSettings:
+    output_dir: Path
+    compute_units: ct.ComputeUnit
+    deployment_target: Optional[ct.target]
+    compute_precision: Optional[ct.precision]
+    max_audio_seconds: float
+    max_symbol_steps: int
+
+
+class PreprocessorWrapper(torch.nn.Module):
+    """Wrapper for the audio preprocessor (mel spectrogram extraction)."""
+
+    def __init__(self, module: torch.nn.Module) -> None:
+        super().__init__()
+        self.module = module
+
+    def forward(
+        self, audio_signal: torch.Tensor, length: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        mel, mel_length = self.module(
+            input_signal=audio_signal, length=length.to(dtype=torch.long)
+        )
+        return mel, mel_length
+
+
+class EncoderWrapper(torch.nn.Module):
+    """Wrapper for the cache-aware FastConformer encoder.
+
+    Note: For the realtime EOU model, the encoder is cache-aware which means
+    it can operate in a streaming fashion. For CoreML export, we export
+    without cache state for simplicity (full-context mode).
+    """
+
+    def __init__(self, module: torch.nn.Module) -> None:
+        super().__init__()
+        self.module = module
+
+    def forward(
+        self, features: torch.Tensor, length: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        encoded, encoded_lengths = self.module(
+            audio_signal=features, length=length.to(dtype=torch.long)
+        )
+        # Synthesize per-frame timestamps (seconds) using the 80 ms encoder stride.
+        # Shape: [B, T_enc]
+        frame_times = (
+            torch.arange(encoded.shape[-1], device=encoded.device, dtype=torch.float32)
+            * 0.08
+        )
+        return encoded, encoded_lengths, frame_times
+
+
+class DecoderWrapper(torch.nn.Module):
+    """Wrapper for the RNNT prediction network (decoder)."""
+
+    def __init__(self, module: torch.nn.Module) -> None:
+        super().__init__()
+        self.module = module
+
+    def forward(
+        self,
+        targets: torch.Tensor,
+        target_lengths: torch.Tensor,
+        h_in: torch.Tensor,
+        c_in: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        state = [h_in, c_in]
+        decoder_output, _, new_state = self.module(
+            targets=targets.to(dtype=torch.long),
+            target_length=target_lengths.to(dtype=torch.long),
+            states=state,
+        )
+        return decoder_output, new_state[0], new_state[1]
+
+
+class JointWrapper(torch.nn.Module):
+    """Wrapper for the RNNT joint network.
+
+    Note: Unlike Parakeet TDT v3, the realtime EOU model does NOT have
+    duration outputs (num_extra_outputs). The joint network outputs only
+    token logits over the vocabulary + blank.
+    """
+
+    def __init__(self, module: torch.nn.Module) -> None:
+        super().__init__()
+        self.module = module
+
+    def forward(
+        self, encoder_outputs: torch.Tensor, decoder_outputs: torch.Tensor
+    ) -> torch.Tensor:
+        # Input: encoder_outputs [B, D, T], decoder_outputs [B, D, U]
+        # Transpose to match what projection layers expect
+        encoder_outputs = encoder_outputs.transpose(1, 2)  # [B, T, D]
+        decoder_outputs = decoder_outputs.transpose(1, 2)  # [B, U, D]
+
+        # Apply projections
+        enc_proj = self.module.enc(encoder_outputs)  # [B, T, joint_hidden]
+        dec_proj = self.module.pred(decoder_outputs)  # [B, U, joint_hidden]
+
+        # Explicit broadcasting along T and U to avoid converter ambiguity
+        x = enc_proj.unsqueeze(2) + dec_proj.unsqueeze(1)  # [B, T, U, joint_hidden]
+        x = self.module.joint_net[0](x)  # ReLU
+        x = self.module.joint_net[1](x)  # Dropout (no-op in eval)
+        out = self.module.joint_net[2](x)  # Linear -> logits [B, T, U, vocab+blank]
+        return out
+
+
+class MelEncoderWrapper(torch.nn.Module):
+    """Fused wrapper: waveform -> mel -> encoder.
+
+    Inputs:
+      - audio_signal: [B, S]
+      - audio_length: [B]
+
+    Outputs:
+      - encoder: [B, D, T_enc]
+      - encoder_length: [B]
+      - frame_times: [T_enc]
+    """
+
+    def __init__(
+        self, preprocessor: PreprocessorWrapper, encoder: EncoderWrapper
+    ) -> None:
+        super().__init__()
+        self.preprocessor = preprocessor
+        self.encoder = encoder
+
+    def forward(
+        self, audio_signal: torch.Tensor, audio_length: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        mel, mel_length = self.preprocessor(audio_signal, audio_length)
+        encoded, enc_len, frame_times = self.encoder(mel, mel_length.to(dtype=torch.int32))
+        return encoded, enc_len, frame_times
+
+
+class JointDecisionWrapper(torch.nn.Module):
+    """Joint + decision head: outputs label id and label prob.
+
+    Unlike Parakeet TDT v3, this model does NOT have duration outputs.
+
+    Inputs:
+      - encoder_outputs: [B, D, T]
+      - decoder_outputs: [B, D, U]
+
+    Returns:
+      - token_id: [B, T, U] int32
+      - token_prob: [B, T, U] float32
+    """
+
+    def __init__(self, joint: JointWrapper, vocab_size: int) -> None:
+        super().__init__()
+        self.joint = joint
+        self.vocab_with_blank = int(vocab_size) + 1
+
+    def forward(self, encoder_outputs: torch.Tensor, decoder_outputs: torch.Tensor):
+        logits = self.joint(encoder_outputs, decoder_outputs)
+
+        # Token selection
+        token_ids = torch.argmax(logits, dim=-1).to(dtype=torch.int32)
+        token_probs_all = torch.softmax(logits, dim=-1)
+        # gather expects int64 (long) indices; cast only for gather
+        token_prob = torch.gather(
+            token_probs_all, dim=-1, index=token_ids.long().unsqueeze(-1)
+        ).squeeze(-1)
+
+        return token_ids, token_prob
+
+
+class JointDecisionSingleStep(torch.nn.Module):
+    """Single-step variant for streaming: encoder_step -> token decision.
+
+    Inputs:
+      - encoder_step: [B=1, D, T=1]
+      - decoder_step: [B=1, D, U=1]
+
+    Returns:
+      - token_id: [1, 1, 1] int32
+      - token_prob: [1, 1, 1] float32
+      - top_k_ids: [1, 1, 1, K] int32
+      - top_k_logits: [1, 1, 1, K] float32
+    """
+
+    def __init__(self, joint: JointWrapper, vocab_size: int, top_k: int = 64) -> None:
+        super().__init__()
+        self.joint = joint
+        self.vocab_with_blank = int(vocab_size) + 1
+        self.top_k = int(top_k)
+
+    def forward(self, encoder_step: torch.Tensor, decoder_step: torch.Tensor):
+        # Reuse JointWrapper which expects [B, D, T] and [B, D, U]
+        logits = self.joint(encoder_step, decoder_step)  # [1, 1, 1, V+blank]
+
+        token_ids = torch.argmax(logits, dim=-1, keepdim=False).to(dtype=torch.int32)
+        token_probs_all = torch.softmax(logits, dim=-1)
+        token_prob = torch.gather(
+            token_probs_all, dim=-1, index=token_ids.long().unsqueeze(-1)
+        ).squeeze(-1)
+
+        # Also expose top-K candidates for host-side processing
+        topk_logits, topk_ids_long = torch.topk(
+            logits, k=min(self.top_k, logits.shape[-1]), dim=-1
+        )
+        topk_ids = topk_ids_long.to(dtype=torch.int32)
+        return token_ids, token_prob, topk_ids, topk_logits
+
+
+def _coreml_convert(
+    traced: torch.jit.ScriptModule,
+    inputs,
+    outputs,
+    settings: ExportSettings,
+    compute_units_override: Optional[ct.ComputeUnit] = None,
+    compute_precision: Optional[ct.precision] = None,
+) -> ct.models.MLModel:
+    cu = (
+        compute_units_override
+        if compute_units_override is not None
+        else settings.compute_units
+    )
+    kwargs = {
+        "convert_to": "mlprogram",
+        "inputs": inputs,
+        "outputs": outputs,
+        "compute_units": cu,
+    }
+    print("Converting:", traced.__class__.__name__)
+    print("Conversion kwargs:", kwargs)
+    if settings.deployment_target is not None:
+        kwargs["minimum_deployment_target"] = settings.deployment_target
+    
+    # Priority: explicit argument > settings
+    if compute_precision is not None:
+        kwargs["compute_precision"] = compute_precision
+    elif settings.compute_precision is not None:
+        kwargs["compute_precision"] = settings.compute_precision
+        
+    return ct.convert(traced, **kwargs)
diff --git a/1280ms/joint_decision.mlmodelc/analytics/coremldata.bin b/1280ms/joint_decision.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..ff9929e12bae25c509655bda0ebda6b9fe13fade
--- /dev/null
+++ b/1280ms/joint_decision.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5bca32ad130dcad6605cc00044c752aa5b45ef57d14c17f2d1a2fa49d6cf55b5
+size 243
diff --git a/1280ms/joint_decision.mlmodelc/coremldata.bin b/1280ms/joint_decision.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..1aa5afe7c02e93f9baba2757cad5c21f563157bf
--- /dev/null
+++ b/1280ms/joint_decision.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22d4abc4625b935ee035b5f8ce7cb28d1041b9b01c12173e287bf4b5f5d99625
+size 493
diff --git a/1280ms/joint_decision.mlmodelc/metadata.json b/1280ms/joint_decision.mlmodelc/metadata.json
new file mode 100644
index 0000000000000000000000000000000000000000..a0a319857f429f5ec4b106397cd52ec354ef2e96
--- /dev/null
+++ b/1280ms/joint_decision.mlmodelc/metadata.json
@@ -0,0 +1,112 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Parakeet EOU single-step joint decision",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1]",
+        "name" : "token_id",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1]",
+        "name" : "token_prob",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1 × 1 × 64)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1, 64]",
+        "name" : "top_k_ids",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 1 × 64)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1, 64]",
+        "name" : "top_k_logits",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.reduceArgmax" : 1,
+      "Ios17.squeeze" : 1,
+      "Ios17.cast" : 6,
+      "Ios17.linear" : 3,
+      "Ios17.transpose" : 2,
+      "Ios17.add" : 1,
+      "Ios16.relu" : 1,
+      "Ios16.softmax" : 1,
+      "Ios17.gatherAlongAxis" : 1,
+      "Ios17.topk" : 1,
+      "Ios17.expandDims" : 3
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int16, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 1]",
+        "name" : "encoder_step",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 640 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 640, 1]",
+        "name" : "decoder_step",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.4.0"
+    },
+    "generatedClassName" : "parakeet_eou_joint_decision_single_step",
+    "method" : "predict"
+  }
+]
\ No newline at end of file
diff --git a/1280ms/joint_decision.mlmodelc/model.mil b/1280ms/joint_decision.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..172a7579866c84407b32b5f746b7f1ae132599d8
--- /dev/null
+++ b/1280ms/joint_decision.mlmodelc/model.mil
@@ -0,0 +1,57 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.4.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<fp32, [1, 640, 1]> decoder_step, tensor<fp32, [1, 512, 1]> encoder_step) {
+            tensor<int32, [3]> input_1_perm_0 = const()[name = tensor<string, []>("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> encoder_step_to_fp16_dtype_0 = const()[name = tensor<string, []>("encoder_step_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> decoder_step_to_fp16_dtype_0 = const()[name = tensor<string, []>("decoder_step_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [640, 512]> joint_module_enc_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_enc_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [640]> joint_module_enc_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(655488)))];
+            tensor<fp16, [1, 512, 1]> encoder_step_to_fp16 = cast(dtype = encoder_step_to_fp16_dtype_0, x = encoder_step)[name = tensor<string, []>("cast_8")];
+            tensor<fp16, [1, 1, 512]> input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_step_to_fp16)[name = tensor<string, []>("transpose_1")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(656832)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1476096)))];
+            tensor<fp16, [1, 640, 1]> decoder_step_to_fp16 = cast(dtype = decoder_step_to_fp16_dtype_0, x = decoder_step)[name = tensor<string, []>("cast_7")];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_step_to_fp16)[name = tensor<string, []>("transpose_0")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_23_axes_0 = const()[name = tensor<string, []>("op_23_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_23_cast_fp16")];
+            tensor<int32, [1]> var_24_axes_0 = const()[name = tensor<string, []>("op_24_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_24_cast_fp16 = expand_dims(axes = var_24_axes_0, x = linear_1_cast_fp16)[name = tensor<string, []>("op_24_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_24_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<fp16, [1027, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [1027, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1477440)))];
+            tensor<fp16, [1027]> joint_module_joint_net_2_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [1027]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2792064)))];
+            tensor<fp16, [1, 1, 1, 1027]> linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<int32, []> var_38_axis_0 = const()[name = tensor<string, []>("op_38_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> var_38_keep_dims_0 = const()[name = tensor<string, []>("op_38_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> var_38_output_dtype_0 = const()[name = tensor<string, []>("op_38_output_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 1, 1]> token_id = reduce_argmax(axis = var_38_axis_0, keep_dims = var_38_keep_dims_0, output_dtype = var_38_output_dtype_0, x = linear_2_cast_fp16)[name = tensor<string, []>("op_38_cast_fp16")];
+            tensor<int32, []> var_44 = const()[name = tensor<string, []>("op_44"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [1, 1, 1, 1027]> token_probs_all_cast_fp16 = softmax(axis = var_44, x = linear_2_cast_fp16)[name = tensor<string, []>("token_probs_all_cast_fp16")];
+            tensor<int32, [1]> var_53_axes_0 = const()[name = tensor<string, []>("op_53_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 1, 1, 1]> var_53 = expand_dims(axes = var_53_axes_0, x = token_id)[name = tensor<string, []>("op_53")];
+            tensor<int32, []> var_54 = const()[name = tensor<string, []>("op_54"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> var_56_validate_indices_0 = const()[name = tensor<string, []>("op_56_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> var_53_to_int16_dtype_0 = const()[name = tensor<string, []>("op_53_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 1, 1, 1]> var_53_to_int16 = cast(dtype = var_53_to_int16_dtype_0, x = var_53)[name = tensor<string, []>("cast_6")];
+            tensor<fp16, [1, 1, 1, 1]> var_56_cast_fp16_cast_int16 = gather_along_axis(axis = var_54, indices = var_53_to_int16, validate_indices = var_56_validate_indices_0, x = token_probs_all_cast_fp16)[name = tensor<string, []>("op_56_cast_fp16_cast_int16")];
+            tensor<int32, [1]> var_58_axes_0 = const()[name = tensor<string, []>("op_58_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 1, 1]> var_58_cast_fp16 = squeeze(axes = var_58_axes_0, x = var_56_cast_fp16_cast_int16)[name = tensor<string, []>("op_58_cast_fp16")];
+            tensor<string, []> var_58_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_58_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, []> var_59 = const()[name = tensor<string, []>("op_59"), val = tensor<int32, []>(64)];
+            tensor<int32, []> var_63_axis_0 = const()[name = tensor<string, []>("op_63_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> var_63_ascending_0 = const()[name = tensor<string, []>("op_63_ascending_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_63_sort_0 = const()[name = tensor<string, []>("op_63_sort_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_63_return_indices_0 = const()[name = tensor<string, []>("op_63_return_indices_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> var_63_cast_fp16_cast_int16_output_indices_dtype_0 = const()[name = tensor<string, []>("op_63_cast_fp16_cast_int16_output_indices_dtype_0"), val = tensor<string, []>("uint16")];
+            tensor<fp16, [1, 1, 1, 64]> var_63_cast_fp16_cast_int16_0, tensor<uint16, [1, 1, 1, 64]> var_63_cast_fp16_cast_int16_1 = topk(ascending = var_63_ascending_0, axis = var_63_axis_0, k = var_59, output_indices_dtype = var_63_cast_fp16_cast_int16_output_indices_dtype_0, return_indices = var_63_return_indices_0, sort = var_63_sort_0, x = linear_2_cast_fp16)[name = tensor<string, []>("op_63_cast_fp16_cast_int16")];
+            tensor<string, []> var_63_cast_fp16_cast_int16_1_to_int32_dtype_0 = const()[name = tensor<string, []>("op_63_cast_fp16_cast_int16_1_to_int32_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<string, []> var_63_cast_fp16_0_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_63_cast_fp16_0_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 1, 1, 64]> top_k_logits = cast(dtype = var_63_cast_fp16_0_to_fp32_dtype_0, x = var_63_cast_fp16_cast_int16_0)[name = tensor<string, []>("cast_3")];
+            tensor<int32, [1, 1, 1, 64]> top_k_ids = cast(dtype = var_63_cast_fp16_cast_int16_1_to_int32_dtype_0, x = var_63_cast_fp16_cast_int16_1)[name = tensor<string, []>("cast_4")];
+            tensor<fp32, [1, 1, 1]> token_prob = cast(dtype = var_58_cast_fp16_to_fp32_dtype_0, x = var_58_cast_fp16)[name = tensor<string, []>("cast_5")];
+        } -> (token_id, token_prob, top_k_ids, top_k_logits);
+}
\ No newline at end of file
diff --git a/1280ms/joint_decision.mlmodelc/weights/weight.bin b/1280ms/joint_decision.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..ec3d2c1dea18b332fa1a3cd8d31981bae4e3f649
--- /dev/null
+++ b/1280ms/joint_decision.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7039b2010a269153f5a96edf28637f921a86ef8822f248f2d6712f7a6bce84b4
+size 2794182
diff --git a/1280ms/parakeet_eou_preprocessor.mlmodelc/analytics/coremldata.bin b/1280ms/parakeet_eou_preprocessor.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..ca93efecf9fbf12e37943f7e192cd324a39e9f94
--- /dev/null
+++ b/1280ms/parakeet_eou_preprocessor.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4ada8b0b99ac1d2ba7acbffacfbbf1a06cb69d30e9410d237ee0aa4c2b0ad63
+size 243
diff --git a/1280ms/parakeet_eou_preprocessor.mlmodelc/coremldata.bin b/1280ms/parakeet_eou_preprocessor.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..ca69399b73b039b35bb1b4967b25b7a7a12e8178
--- /dev/null
+++ b/1280ms/parakeet_eou_preprocessor.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc7252fa47622fe39577361233627062019a3bb740fdbb5366a7bae09df0ec5e
+size 422
diff --git a/1280ms/parakeet_eou_preprocessor.mlmodelc/metadata.json b/1280ms/parakeet_eou_preprocessor.mlmodelc/metadata.json
new file mode 100644
index 0000000000000000000000000000000000000000..98b70ffcae507184b68f87d7adaad706e5c68c9a
--- /dev/null
+++ b/1280ms/parakeet_eou_preprocessor.mlmodelc/metadata.json
@@ -0,0 +1,105 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Parakeet EOU preprocessor",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "mel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "mel_length",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float32",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Range1d" : 1,
+      "Ios17.reshape" : 2,
+      "Identity" : 1,
+      "Ios17.matmul" : 1,
+      "Ios17.expandDims" : 5,
+      "Select" : 1,
+      "Ios17.add" : 3,
+      "Ios17.sliceByIndex" : 3,
+      "Ios16.reduceSum" : 1,
+      "Shape" : 1,
+      "Ios17.gather" : 1,
+      "Pad" : 1,
+      "Ios17.log" : 1,
+      "Ios17.conv" : 2,
+      "Ios17.sub" : 2,
+      "Ios17.pow" : 1,
+      "Ios17.cast" : 2,
+      "Stack" : 1,
+      "Ios17.concat" : 1,
+      "Ios17.floorDiv" : 1,
+      "Ios17.greaterEqual" : 1,
+      "Ios17.mul" : 1
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "dataType" : "Float32",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 1...32000",
+        "shapeRange" : "[[1, 1], [1, 32000]]",
+        "formattedType" : "MultiArray (Float32 1 × 1)",
+        "type" : "MultiArray",
+        "shape" : "[1, 1]",
+        "name" : "audio_signal",
+        "shortDescription" : ""
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "audio_length",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.4.0",
+      "com.github.apple.coremltools.version" : "8.3.0"
+    },
+    "generatedClassName" : "parakeet_eou_preprocessor",
+    "method" : "predict"
+  }
+]
\ No newline at end of file
diff --git a/1280ms/parakeet_eou_preprocessor.mlmodelc/model.mil b/1280ms/parakeet_eou_preprocessor.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..25db6a1374c1718340459f374949bfba2bff85bb
--- /dev/null
+++ b/1280ms/parakeet_eou_preprocessor.mlmodelc/model.mil
@@ -0,0 +1,96 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.4.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<int32, [1]> audio_length, tensor<fp32, [1, ?]> audio_signal) [FlexibleShapeInformation = tuple<tuple<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>, tuple<tensor<string, []>, dict<tensor<string, []>, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"audio_signal", [1, 1]}}), ("RangeDims", {{"audio_signal", [[1, 1], [1, 32000]]}})))] {
+            tensor<int32, []> var_9 = const()[name = tensor<string, []>("op_9"), val = tensor<int32, []>(1)];
+            tensor<int32, []> var_10 = const()[name = tensor<string, []>("op_10"), val = tensor<int32, []>(160)];
+            tensor<int32, []> var_32 = const()[name = tensor<string, []>("op_32"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> var_33 = add(x = audio_length, y = var_32)[name = tensor<string, []>("op_33")];
+            tensor<int32, []> var_34 = const()[name = tensor<string, []>("op_34"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> var_35 = sub(x = var_33, y = var_34)[name = tensor<string, []>("op_35")];
+            tensor<int32, [1]> floor_div_0 = floor_div(x = var_35, y = var_10)[name = tensor<string, []>("floor_div_0")];
+            tensor<string, []> var_36_dtype_0 = const()[name = tensor<string, []>("op_36_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, []> var_37_promoted = const()[name = tensor<string, []>("op_37_promoted"), val = tensor<fp32, []>(0x1p+0)];
+            tensor<fp32, [1]> var_36 = cast(dtype = var_36_dtype_0, x = floor_div_0)[name = tensor<string, []>("cast_11")];
+            tensor<fp32, [1]> seq_len_1 = add(x = var_36, y = var_37_promoted)[name = tensor<string, []>("seq_len_1")];
+            tensor<string, []> cast_2_dtype_0 = const()[name = tensor<string, []>("cast_2_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [2]> var_41_begin_0 = const()[name = tensor<string, []>("op_41_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_41_end_0 = const()[name = tensor<string, []>("op_41_end_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<bool, [2]> var_41_end_mask_0 = const()[name = tensor<string, []>("op_41_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<bool, [2]> var_41_squeeze_mask_0 = const()[name = tensor<string, []>("op_41_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
+            tensor<fp32, [1]> var_41 = slice_by_index(begin = var_41_begin_0, end = var_41_end_0, end_mask = var_41_end_mask_0, squeeze_mask = var_41_squeeze_mask_0, x = audio_signal)[name = tensor<string, []>("op_41")];
+            tensor<int32, [1]> var_42_axes_0 = const()[name = tensor<string, []>("op_42_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1]> var_42 = expand_dims(axes = var_42_axes_0, x = var_41)[name = tensor<string, []>("op_42")];
+            tensor<int32, [2]> var_44_begin_0 = const()[name = tensor<string, []>("op_44_begin_0"), val = tensor<int32, [2]>([0, 1])];
+            tensor<int32, [2]> var_44_end_0 = const()[name = tensor<string, []>("op_44_end_0"), val = tensor<int32, [2]>([1, 0])];
+            tensor<bool, [2]> var_44_end_mask_0 = const()[name = tensor<string, []>("op_44_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<fp32, [1, ?]> var_44 = slice_by_index(begin = var_44_begin_0, end = var_44_end_0, end_mask = var_44_end_mask_0, x = audio_signal)[name = tensor<string, []>("op_44")];
+            tensor<int32, [2]> var_46_begin_0 = const()[name = tensor<string, []>("op_46_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_46_end_0 = const()[name = tensor<string, []>("op_46_end_0"), val = tensor<int32, [2]>([1, -1])];
+            tensor<bool, [2]> var_46_end_mask_0 = const()[name = tensor<string, []>("op_46_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp32, [1, ?]> var_46 = slice_by_index(begin = var_46_begin_0, end = var_46_end_0, end_mask = var_46_end_mask_0, x = audio_signal)[name = tensor<string, []>("op_46")];
+            tensor<fp32, []> var_47 = const()[name = tensor<string, []>("op_47"), val = tensor<fp32, []>(0x1.f0a3d8p-1)];
+            tensor<fp32, [1, ?]> var_48 = mul(x = var_46, y = var_47)[name = tensor<string, []>("op_48")];
+            tensor<fp32, [1, ?]> var_49 = sub(x = var_44, y = var_48)[name = tensor<string, []>("op_49")];
+            tensor<bool, []> input_1_interleave_0 = const()[name = tensor<string, []>("input_1_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, ?]> input_1 = concat(axis = var_9, interleave = input_1_interleave_0, values = (var_42, var_49))[name = tensor<string, []>("input_1")];
+            tensor<int32, [3]> concat_0x = const()[name = tensor<string, []>("concat_0x"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp32, [1, 1, ?]> input_3 = reshape(shape = concat_0x, x = input_1)[name = tensor<string, []>("input_3")];
+            tensor<fp32, []> const_1 = const()[name = tensor<string, []>("const_1"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [6]> input_5_pad_0 = const()[name = tensor<string, []>("input_5_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 256, 256])];
+            tensor<string, []> input_5_mode_0 = const()[name = tensor<string, []>("input_5_mode_0"), val = tensor<string, []>("reflect")];
+            tensor<fp32, [1, 1, ?]> input_5 = pad(constant_val = const_1, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3)[name = tensor<string, []>("input_5")];
+            tensor<int32, [2]> concat_1x = const()[name = tensor<string, []>("concat_1x"), val = tensor<int32, [2]>([1, -1])];
+            tensor<fp32, [1, ?]> input = reshape(shape = concat_1x, x = input_5)[name = tensor<string, []>("input")];
+            tensor<fp32, [257, 1, 512]> expand_dims_1 = const()[name = tensor<string, []>("expand_dims_1"), val = tensor<fp32, [257, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp32, [257, 1, 512]> expand_dims_2 = const()[name = tensor<string, []>("expand_dims_2"), val = tensor<fp32, [257, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526464)))];
+            tensor<int32, [1]> expand_dims_3 = const()[name = tensor<string, []>("expand_dims_3"), val = tensor<int32, [1]>([160])];
+            tensor<int32, [1]> expand_dims_4_axes_0 = const()[name = tensor<string, []>("expand_dims_4_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1, ?]> expand_dims_4 = expand_dims(axes = expand_dims_4_axes_0, x = input)[name = tensor<string, []>("expand_dims_4")];
+            tensor<string, []> conv_0_pad_type_0 = const()[name = tensor<string, []>("conv_0_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> conv_0_pad_0 = const()[name = tensor<string, []>("conv_0_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_0_dilations_0 = const()[name = tensor<string, []>("conv_0_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> conv_0_groups_0 = const()[name = tensor<string, []>("conv_0_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 257, ?]> conv_0 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1, x = expand_dims_4)[name = tensor<string, []>("conv_0")];
+            tensor<string, []> conv_1_pad_type_0 = const()[name = tensor<string, []>("conv_1_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> conv_1_pad_0 = const()[name = tensor<string, []>("conv_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_1_dilations_0 = const()[name = tensor<string, []>("conv_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> conv_1_groups_0 = const()[name = tensor<string, []>("conv_1_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 257, ?]> conv_1 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2, x = expand_dims_4)[name = tensor<string, []>("conv_1")];
+            tensor<int32, []> stack_0_axis_0 = const()[name = tensor<string, []>("stack_0_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<fp32, [1, 257, ?, 2]> stack_0 = stack(axis = stack_0_axis_0, values = (conv_0, conv_1))[name = tensor<string, []>("stack_0")];
+            tensor<fp32, []> var_17_promoted = const()[name = tensor<string, []>("op_17_promoted"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1, 257, ?, 2]> var_65 = pow(x = stack_0, y = var_17_promoted)[name = tensor<string, []>("op_65")];
+            tensor<int32, [1]> var_67_axes_0 = const()[name = tensor<string, []>("op_67_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> var_67_keep_dims_0 = const()[name = tensor<string, []>("op_67_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 257, ?]> var_67 = reduce_sum(axes = var_67_axes_0, keep_dims = var_67_keep_dims_0, x = var_65)[name = tensor<string, []>("op_67")];
+            tensor<fp32, [1, 257, ?]> x_9 = identity(x = var_67)[name = tensor<string, []>("x_9")];
+            tensor<fp32, [1, 128, 257]> const_2 = const()[name = tensor<string, []>("const_2"), val = tensor<fp32, [1, 128, 257]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1052864)))];
+            tensor<bool, []> x_11_transpose_x_0 = const()[name = tensor<string, []>("x_11_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_11_transpose_y_0 = const()[name = tensor<string, []>("x_11_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 128, ?]> x_11 = matmul(transpose_x = x_11_transpose_x_0, transpose_y = x_11_transpose_y_0, x = const_2, y = x_9)[name = tensor<string, []>("x_11")];
+            tensor<fp32, []> var_74 = const()[name = tensor<string, []>("op_74"), val = tensor<fp32, []>(0x1p-24)];
+            tensor<fp32, [1, 128, ?]> var_75 = add(x = x_11, y = var_74)[name = tensor<string, []>("op_75")];
+            tensor<fp32, []> x_epsilon_0 = const()[name = tensor<string, []>("x_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
+            tensor<fp32, [1, 128, ?]> x = log(epsilon = x_epsilon_0, x = var_75)[name = tensor<string, []>("x")];
+            tensor<int32, [3]> var_77_shape = shape(x = x)[name = tensor<string, []>("op_77_shape")];
+            tensor<int32, []> select_4 = const()[name = tensor<string, []>("select_4"), val = tensor<int32, []>(2)];
+            tensor<int32, []> gather_4_axis_0 = const()[name = tensor<string, []>("gather_4_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> gather_4_batch_dims_0 = const()[name = tensor<string, []>("gather_4_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_4_validate_indices_0 = const()[name = tensor<string, []>("gather_4_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<int32, []> gather_4 = gather(axis = gather_4_axis_0, batch_dims = gather_4_batch_dims_0, indices = select_4, validate_indices = gather_4_validate_indices_0, x = var_77_shape)[name = tensor<string, []>("gather_4")];
+            tensor<int32, []> const_3 = const()[name = tensor<string, []>("const_3"), val = tensor<int32, []>(0)];
+            tensor<int32, []> const_4 = const()[name = tensor<string, []>("const_4"), val = tensor<int32, []>(1)];
+            tensor<int32, [?]> mask_1 = range_1d(end = gather_4, start = const_3, step = const_4)[name = tensor<string, []>("mask_1")];
+            tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = tensor<string, []>("expand_dims_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = tensor<string, []>("expand_dims_0")];
+            tensor<int32, [1]> var_82_axes_0 = const()[name = tensor<string, []>("op_82_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> mel_length = cast(dtype = cast_2_dtype_0, x = seq_len_1)[name = tensor<string, []>("cast_10")];
+            tensor<int32, [1, 1]> var_82 = expand_dims(axes = var_82_axes_0, x = mel_length)[name = tensor<string, []>("op_82")];
+            tensor<bool, [1, ?]> mask = greater_equal(x = expand_dims_0, y = var_82)[name = tensor<string, []>("mask")];
+            tensor<int32, [1]> var_84_axes_0 = const()[name = tensor<string, []>("op_84_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, ?]> var_84 = expand_dims(axes = var_84_axes_0, x = mask)[name = tensor<string, []>("op_84")];
+            tensor<fp32, []> cast_7 = const()[name = tensor<string, []>("cast_7"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<fp32, [1, 128, ?]> mel = select(a = cast_7, b = x, cond = var_84)[name = tensor<string, []>("processed_signal")];
+        } -> (mel, mel_length);
+}
\ No newline at end of file
diff --git a/1280ms/parakeet_eou_preprocessor.mlmodelc/weights/weight.bin b/1280ms/parakeet_eou_preprocessor.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..99f6476377d3e7474aa570db365b7429f9d7b5b8
--- /dev/null
+++ b/1280ms/parakeet_eou_preprocessor.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:009bba4fde82dc55db9b55d77cf3ba5f791ce366c49f079285fe25a3b6e2291d
+size 1184512
diff --git a/1280ms/streaming_encoder.mlmodelc/analytics/coremldata.bin b/1280ms/streaming_encoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..e22bce589c25d1e6413bee12465c6eb2d68c6262
--- /dev/null
+++ b/1280ms/streaming_encoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0a3c84022a9d2dc769d38cf8f45e93423e20734d092e3c16db11fbf6dca4004
+size 243
diff --git a/1280ms/streaming_encoder.mlmodelc/coremldata.bin b/1280ms/streaming_encoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..7e43604cbe391d85b29d5826bb5cee4f987575d2
--- /dev/null
+++ b/1280ms/streaming_encoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:41ce3f96c3d6b3333796fc4ed82cb0c9b4ea99396b88f8eec3ba24394ba2bb78
+size 671
diff --git a/1280ms/streaming_encoder.mlmodelc/metadata.json b/1280ms/streaming_encoder.mlmodelc/metadata.json
new file mode 100644
index 0000000000000000000000000000000000000000..bd9ca5c8491ca0588ddec260eb9c532ffb9ef560
--- /dev/null
+++ b/1280ms/streaming_encoder.mlmodelc/metadata.json
@@ -0,0 +1,187 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512 × 17)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 17]",
+        "name" : "encoded_output",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "encoded_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128 × 16)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 16]",
+        "name" : "new_pre_cache",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 17 × 1 × 70 × 512)",
+        "shortDescription" : "",
+        "shape" : "[17, 1, 70, 512]",
+        "name" : "new_cache_last_channel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 17 × 1 × 512 × 8)",
+        "shortDescription" : "",
+        "shape" : "[17, 1, 512, 8]",
+        "name" : "new_cache_last_time",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "new_cache_last_channel_len",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.floor" : 3,
+      "Ios17.logicalAnd" : 3,
+      "Ios17.reshape" : 103,
+      "Ios16.softmax" : 17,
+      "Ios17.matmul" : 51,
+      "Ios17.transpose" : 157,
+      "Split" : 17,
+      "Ios17.expandDims" : 6,
+      "Select" : 51,
+      "Ios17.add" : 126,
+      "Tile" : 1,
+      "Ios17.sliceByIndex" : 106,
+      "Ios16.sigmoid" : 17,
+      "Pad" : 20,
+      "Ios17.logicalNot" : 2,
+      "Ios17.layerNorm" : 102,
+      "Ios17.less" : 1,
+      "Ios17.sub" : 1,
+      "Ios17.conv" : 56,
+      "Ios17.clip" : 2,
+      "Ios16.relu" : 3,
+      "Ios17.linear" : 137,
+      "Ios17.concat" : 52,
+      "Ios17.greaterEqual" : 1,
+      "Ios17.cast" : 14,
+      "Ios16.silu" : 51,
+      "Stack" : 2,
+      "Ios17.mul" : 72
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.4.0",
+      "com.github.apple.coremltools.version" : "8.3.0"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128 × 129)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 129]",
+        "name" : "audio_signal",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "audio_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128 × 16)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 16]",
+        "name" : "pre_cache",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 17 × 1 × 70 × 512)",
+        "shortDescription" : "",
+        "shape" : "[17, 1, 70, 512]",
+        "name" : "cache_last_channel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 17 × 1 × 512 × 8)",
+        "shortDescription" : "",
+        "shape" : "[17, 1, 512, 8]",
+        "name" : "cache_last_time",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "cache_last_channel_len",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "streaming_encoder",
+    "method" : "predict"
+  }
+]
\ No newline at end of file
diff --git a/1280ms/streaming_encoder.mlmodelc/model.mil b/1280ms/streaming_encoder.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..e87da5042d394a857c4bd11850c4ec70d6577d86
--- /dev/null
+++ b/1280ms/streaming_encoder.mlmodelc/model.mil
@@ -0,0 +1,2977 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.4.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<int32, [1]> audio_length, tensor<fp32, [1, 128, 129]> audio_signal, tensor<fp32, [17, 1, 70, 512]> cache_last_channel, tensor<int32, [1]> cache_last_channel_len, tensor<fp32, [17, 1, 512, 8]> cache_last_time, tensor<fp32, [1, 128, 16]> pre_cache) {
+            tensor<int32, []> var_9 = const()[name = tensor<string, []>("op_9"), val = tensor<int32, []>(2)];
+            tensor<bool, []> full_input_interleave_0 = const()[name = tensor<string, []>("full_input_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> pre_cache_to_fp16_dtype_0 = const()[name = tensor<string, []>("pre_cache_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<string, []> audio_signal_to_fp16_dtype_0 = const()[name = tensor<string, []>("audio_signal_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 128, 129]> audio_signal_to_fp16 = cast(dtype = audio_signal_to_fp16_dtype_0, x = audio_signal)[name = tensor<string, []>("cast_193")];
+            tensor<fp16, [1, 128, 16]> pre_cache_to_fp16 = cast(dtype = pre_cache_to_fp16_dtype_0, x = pre_cache)[name = tensor<string, []>("cast_194")];
+            tensor<fp16, [1, 128, 145]> full_input_cast_fp16 = concat(axis = var_9, interleave = full_input_interleave_0, values = (pre_cache_to_fp16, audio_signal_to_fp16))[name = tensor<string, []>("full_input_cast_fp16")];
+            tensor<int32, []> var_12 = const()[name = tensor<string, []>("op_12"), val = tensor<int32, []>(16)];
+            tensor<int32, [1]> value_1 = add(x = audio_length, y = var_12)[name = tensor<string, []>("value_1")];
+            tensor<int32, [3]> var_28_begin_0 = const()[name = tensor<string, []>("op_28_begin_0"), val = tensor<int32, [3]>([0, 0, 129])];
+            tensor<int32, [3]> var_28_end_0 = const()[name = tensor<string, []>("op_28_end_0"), val = tensor<int32, [3]>([1, 128, 145])];
+            tensor<bool, [3]> var_28_end_mask_0 = const()[name = tensor<string, []>("op_28_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 128, 16]> var_28_cast_fp16 = slice_by_index(begin = var_28_begin_0, end = var_28_end_0, end_mask = var_28_end_mask_0, x = full_input_cast_fp16)[name = tensor<string, []>("op_28_cast_fp16")];
+            tensor<string, []> var_28_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_28_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, []> var_62 = const()[name = tensor<string, []>("op_62"), val = tensor<int32, []>(-1)];
+            tensor<int32, []> var_64 = const()[name = tensor<string, []>("op_64"), val = tensor<int32, []>(1)];
+            tensor<int32, [3]> x_1_perm_0 = const()[name = tensor<string, []>("x_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> cast_0_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_0_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> _inversed_108_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_108_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1]> value_1_to_fp16 = cast(dtype = cast_0_to_fp16_dtype_0, x = value_1)[name = tensor<string, []>("cast_191")];
+            tensor<fp16, [1]> _inversed_108_cast_fp16 = mul(x = value_1_to_fp16, y = _inversed_108_y_0_to_fp16)[name = tensor<string, []>("_inversed_108_cast_fp16")];
+            tensor<fp16, []> var_109_to_fp16 = const()[name = tensor<string, []>("op_109_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> lengths_1_cast_fp16 = add(x = _inversed_108_cast_fp16, y = var_109_to_fp16)[name = tensor<string, []>("lengths_1_cast_fp16")];
+            tensor<fp16, [1]> lengths_3_cast_fp16 = floor(x = lengths_1_cast_fp16)[name = tensor<string, []>("lengths_3_cast_fp16")];
+            tensor<fp16, []> _inversed_116_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_116_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1]> _inversed_116_cast_fp16 = mul(x = lengths_3_cast_fp16, y = _inversed_116_y_0_to_fp16)[name = tensor<string, []>("_inversed_116_cast_fp16")];
+            tensor<fp16, []> var_117_to_fp16 = const()[name = tensor<string, []>("op_117_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> lengths_7_cast_fp16 = add(x = _inversed_116_cast_fp16, y = var_117_to_fp16)[name = tensor<string, []>("lengths_7_cast_fp16")];
+            tensor<fp16, [1]> lengths_9_cast_fp16 = floor(x = lengths_7_cast_fp16)[name = tensor<string, []>("lengths_9_cast_fp16")];
+            tensor<fp16, []> _inversed_124_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_124_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1]> _inversed_124_cast_fp16 = mul(x = lengths_9_cast_fp16, y = _inversed_124_y_0_to_fp16)[name = tensor<string, []>("_inversed_124_cast_fp16")];
+            tensor<fp16, []> var_125_to_fp16 = const()[name = tensor<string, []>("op_125_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> lengths_13_cast_fp16 = add(x = _inversed_124_cast_fp16, y = var_125_to_fp16)[name = tensor<string, []>("lengths_13_cast_fp16")];
+            tensor<fp16, [1]> lengths_cast_fp16 = floor(x = lengths_13_cast_fp16)[name = tensor<string, []>("lengths_cast_fp16")];
+            tensor<string, []> cast_9_dtype_0 = const()[name = tensor<string, []>("cast_9_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1]> input_1_axes_0 = const()[name = tensor<string, []>("input_1_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 145, 128]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = full_input_cast_fp16)[name = tensor<string, []>("transpose_241")];
+            tensor<fp16, [1, 1, 145, 128]> input_1_cast_fp16 = expand_dims(axes = input_1_axes_0, x = x_1_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
+            tensor<int32, [8]> input_3_pad_0 = const()[name = tensor<string, []>("input_3_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            tensor<string, []> input_3_mode_0 = const()[name = tensor<string, []>("input_3_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_0_to_fp16 = const()[name = tensor<string, []>("const_0_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 1, 148, 131]> input_3_cast_fp16 = pad(constant_val = const_0_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<string, []> input_5_pad_type_0 = const()[name = tensor<string, []>("input_5_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_5_strides_0 = const()[name = tensor<string, []>("input_5_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, [4]> input_5_pad_0 = const()[name = tensor<string, []>("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_5_dilations_0 = const()[name = tensor<string, []>("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_5_groups_0 = const()[name = tensor<string, []>("input_5_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_0_weight_to_fp16"), val = tensor<fp16, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4736)))];
+            tensor<fp16, [1, 256, 73, 65]> input_5_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [1, 256, 73, 65]> input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<int32, [8]> input_9_pad_0 = const()[name = tensor<string, []>("input_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            tensor<string, []> input_9_mode_0 = const()[name = tensor<string, []>("input_9_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_1_to_fp16 = const()[name = tensor<string, []>("const_1_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 256, 76, 68]> input_9_cast_fp16 = pad(constant_val = const_1_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
+            tensor<string, []> input_11_pad_type_0 = const()[name = tensor<string, []>("input_11_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_11_strides_0 = const()[name = tensor<string, []>("input_11_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, []> input_11_groups_0 = const()[name = tensor<string, []>("input_11_groups_0"), val = tensor<int32, []>(256)];
+            tensor<int32, [4]> input_11_pad_0 = const()[name = tensor<string, []>("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_11_dilations_0 = const()[name = tensor<string, []>("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_2_weight_to_fp16"), val = tensor<fp16, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5312)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9984)))];
+            tensor<fp16, [1, 256, 37, 33]> input_11_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<string, []> input_13_pad_type_0 = const()[name = tensor<string, []>("input_13_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_13_strides_0 = const()[name = tensor<string, []>("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_13_pad_0 = const()[name = tensor<string, []>("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_13_dilations_0 = const()[name = tensor<string, []>("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_13_groups_0 = const()[name = tensor<string, []>("input_13_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_3_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_3_weight_to_fp16"), val = tensor<fp16, [256, 256, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10560)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141696)))];
+            tensor<fp16, [1, 256, 37, 33]> input_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
+            tensor<fp16, [1, 256, 37, 33]> input_15_cast_fp16 = relu(x = input_13_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
+            tensor<int32, [8]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            tensor<string, []> input_17_mode_0 = const()[name = tensor<string, []>("input_17_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_2_to_fp16 = const()[name = tensor<string, []>("const_2_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 256, 40, 36]> input_17_cast_fp16 = pad(constant_val = const_2_to_fp16, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
+            tensor<string, []> input_19_pad_type_0 = const()[name = tensor<string, []>("input_19_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_19_strides_0 = const()[name = tensor<string, []>("input_19_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, []> input_19_groups_0 = const()[name = tensor<string, []>("input_19_groups_0"), val = tensor<int32, []>(256)];
+            tensor<int32, [4]> input_19_pad_0 = const()[name = tensor<string, []>("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_19_dilations_0 = const()[name = tensor<string, []>("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_5_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_5_weight_to_fp16"), val = tensor<fp16, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142272)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146944)))];
+            tensor<fp16, [1, 256, 19, 17]> input_19_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
+            tensor<string, []> input_21_pad_type_0 = const()[name = tensor<string, []>("input_21_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_21_strides_0 = const()[name = tensor<string, []>("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_21_pad_0 = const()[name = tensor<string, []>("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_21_dilations_0 = const()[name = tensor<string, []>("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_21_groups_0 = const()[name = tensor<string, []>("input_21_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_6_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_6_weight_to_fp16"), val = tensor<fp16, [256, 256, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147520)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278656)))];
+            tensor<fp16, [1, 256, 19, 17]> input_21_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
+            tensor<fp16, [1, 256, 19, 17]> x_3_cast_fp16 = relu(x = input_21_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
+            tensor<int32, [4]> var_181_perm_0 = const()[name = tensor<string, []>("op_181_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_182 = const()[name = tensor<string, []>("op_182"), val = tensor<int32, [3]>([1, 19, -1])];
+            tensor<fp16, [1, 19, 256, 17]> var_181_cast_fp16 = transpose(perm = var_181_perm_0, x = x_3_cast_fp16)[name = tensor<string, []>("transpose_240")];
+            tensor<fp16, [1, 19, 4352]> input_23_cast_fp16 = reshape(shape = var_182, x = var_181_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
+            tensor<fp16, [512, 4352]> encoder_pre_encode_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_out_weight_to_fp16"), val = tensor<fp16, [512, 4352]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279232)))];
+            tensor<fp16, [512]> encoder_pre_encode_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4735744)))];
+            tensor<fp16, [1, 19, 512]> linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<int32, [3]> var_192_begin_0 = const()[name = tensor<string, []>("op_192_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_192_end_0 = const()[name = tensor<string, []>("op_192_end_0"), val = tensor<int32, [3]>([1, 19, 512])];
+            tensor<bool, [3]> var_192_end_mask_0 = const()[name = tensor<string, []>("op_192_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 17, 512]> var_192_cast_fp16 = slice_by_index(begin = var_192_begin_0, end = var_192_end_0, end_mask = var_192_end_mask_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_192_cast_fp16")];
+            tensor<int32, []> var_194 = const()[name = tensor<string, []>("op_194"), val = tensor<int32, []>(2)];
+            tensor<int32, [1]> lengths_cast_fp16_to_int32 = cast(dtype = cast_9_dtype_0, x = lengths_cast_fp16)[name = tensor<string, []>("cast_190")];
+            tensor<int32, [1]> var_195 = sub(x = lengths_cast_fp16_to_int32, y = var_194)[name = tensor<string, []>("op_195")];
+            tensor<string, []> var_195_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_195_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> var_60_promoted_to_fp16 = const()[name = tensor<string, []>("op_60_promoted_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, []> const_5_to_fp16 = const()[name = tensor<string, []>("const_5_to_fp16"), val = tensor<fp16, []>(inf)];
+            tensor<fp16, [1]> var_195_to_fp16 = cast(dtype = var_195_promoted_to_fp16_dtype_0, x = var_195)[name = tensor<string, []>("cast_189")];
+            tensor<fp16, [1]> clip_0_cast_fp16 = clip(alpha = var_60_promoted_to_fp16, beta = const_5_to_fp16, x = var_195_to_fp16)[name = tensor<string, []>("clip_0_cast_fp16")];
+            tensor<int32, [1]> max_audio_length_1 = const()[name = tensor<string, []>("max_audio_length_1"), val = tensor<int32, [1]>([17])];
+            tensor<fp16, []> var_211_promoted_to_fp16 = const()[name = tensor<string, []>("op_211_promoted_to_fp16"), val = tensor<fp16, []>(0x1.18p+6)];
+            tensor<fp16, [1]> padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_211_promoted_to_fp16)[name = tensor<string, []>("padding_length_cast_fp16")];
+            tensor<int32, []> const_7 = const()[name = tensor<string, []>("const_7"), val = tensor<int32, []>(-1)];
+            tensor<int32, [1]> var_213 = mul(x = cache_last_channel_len, y = const_7)[name = tensor<string, []>("op_213")];
+            tensor<int32, []> var_214 = const()[name = tensor<string, []>("op_214"), val = tensor<int32, []>(70)];
+            tensor<int32, [1]> offset = add(x = var_213, y = var_214)[name = tensor<string, []>("offset")];
+            tensor<int32, [1]> var_254_axes_0 = const()[name = tensor<string, []>("op_254_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 1]> var_254_cast_fp16 = expand_dims(axes = var_254_axes_0, x = padding_length_cast_fp16)[name = tensor<string, []>("op_254_cast_fp16")];
+            tensor<fp16, [1, 87]> var_253_promoted_to_fp16 = const()[name = tensor<string, []>("op_253_promoted_to_fp16"), val = tensor<fp16, [1, 87]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4736832)))];
+            tensor<bool, [1, 87]> pad_mask_1_cast_fp16 = less(x = var_253_promoted_to_fp16, y = var_254_cast_fp16)[name = tensor<string, []>("pad_mask_1_cast_fp16")];
+            tensor<int32, [1, 87]> expand_dims_1 = const()[name = tensor<string, []>("expand_dims_1"), val = tensor<int32, [1, 87]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86]])];
+            tensor<int32, [1]> var_260_axes_0 = const()[name = tensor<string, []>("op_260_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 1]> var_260 = expand_dims(axes = var_260_axes_0, x = offset)[name = tensor<string, []>("op_260")];
+            tensor<bool, [1, 87]> pad_mask_off = greater_equal(x = expand_dims_1, y = var_260)[name = tensor<string, []>("pad_mask_off")];
+            tensor<bool, [1, 87]> pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = tensor<string, []>("pad_mask_3")];
+            tensor<int32, [1]> var_263_axes_0 = const()[name = tensor<string, []>("op_263_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 87]> var_263 = expand_dims(axes = var_263_axes_0, x = pad_mask_3)[name = tensor<string, []>("op_263")];
+            tensor<int32, [3]> var_264 = const()[name = tensor<string, []>("op_264"), val = tensor<int32, [3]>([1, 87, 1])];
+            tensor<bool, [1, 87, 87]> pad_mask_for_att_mask_1 = tile(reps = var_264, x = var_263)[name = tensor<string, []>("pad_mask_for_att_mask_1")];
+            tensor<int32, [3]> var_266_perm_0 = const()[name = tensor<string, []>("op_266_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<bool, [1, 87, 87]> var_266 = transpose(perm = var_266_perm_0, x = pad_mask_for_att_mask_1)[name = tensor<string, []>("transpose_239")];
+            tensor<bool, [1, 87, 87]> pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_266)[name = tensor<string, []>("pad_mask_for_att_mask")];
+            tensor<bool, [1, 87, 87]> const_15 = const()[name = tensor<string, []>("const_15"), val = tensor<bool, [1, 87, 87]>([[[true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])];
+            tensor<bool, [1, 87, 87]> att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_15)[name = tensor<string, []>("att_mask_9")];
+            tensor<bool, [1, 87, 87]> att_mask = logical_not(x = att_mask_9)[name = tensor<string, []>("att_mask")];
+            tensor<bool, [1, 87]> pad_mask_5 = logical_not(x = pad_mask_3)[name = tensor<string, []>("pad_mask_5")];
+            tensor<int32, [2]> pad_mask_begin_0 = const()[name = tensor<string, []>("pad_mask_begin_0"), val = tensor<int32, [2]>([0, 70])];
+            tensor<int32, [2]> pad_mask_end_0 = const()[name = tensor<string, []>("pad_mask_end_0"), val = tensor<int32, [2]>([1, 87])];
+            tensor<bool, [2]> pad_mask_end_mask_0 = const()[name = tensor<string, []>("pad_mask_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<bool, [1, 17]> pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = tensor<string, []>("pad_mask")];
+            tensor<int32, [3]> mask_1_begin_0 = const()[name = tensor<string, []>("mask_1_begin_0"), val = tensor<int32, [3]>([0, 70, 0])];
+            tensor<int32, [3]> mask_1_end_0 = const()[name = tensor<string, []>("mask_1_end_0"), val = tensor<int32, [3]>([1, 87, 87])];
+            tensor<bool, [3]> mask_1_end_mask_0 = const()[name = tensor<string, []>("mask_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<bool, [1, 17, 87]> mask_1 = slice_by_index(begin = mask_1_begin_0, end = mask_1_end_0, end_mask = mask_1_end_mask_0, x = att_mask)[name = tensor<string, []>("mask_1")];
+            tensor<int32, [4]> cache_1_begin_0 = const()[name = tensor<string, []>("cache_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_1_end_0 = const()[name = tensor<string, []>("cache_1_end_0"), val = tensor<int32, [4]>([1, 1, 70, 512])];
+            tensor<bool, [4]> cache_1_end_mask_0 = const()[name = tensor<string, []>("cache_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_1_squeeze_mask_0 = const()[name = tensor<string, []>("cache_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<string, []> cache_last_channel_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_last_channel_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [17, 1, 70, 512]> cache_last_channel_to_fp16 = cast(dtype = cache_last_channel_to_fp16_dtype_0, x = cache_last_channel)[name = tensor<string, []>("cast_188")];
+            tensor<fp16, [1, 70, 512]> cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_1_cast_fp16")];
+            tensor<int32, [4]> cache_3_begin_0 = const()[name = tensor<string, []>("cache_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_3_end_0 = const()[name = tensor<string, []>("cache_3_end_0"), val = tensor<int32, [4]>([1, 1, 512, 8])];
+            tensor<bool, [4]> cache_3_end_mask_0 = const()[name = tensor<string, []>("cache_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_3_squeeze_mask_0 = const()[name = tensor<string, []>("cache_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<string, []> cache_last_time_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_last_time_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [17, 1, 512, 8]> cache_last_time_to_fp16 = cast(dtype = cache_last_time_to_fp16_dtype_0, x = cache_last_time)[name = tensor<string, []>("cast_187")];
+            tensor<fp16, [1, 512, 8]> cache_3_cast_fp16 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_3_cast_fp16")];
+            tensor<int32, [1]> input_27_axes_0 = const()[name = tensor<string, []>("input_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4737088)))];
+            tensor<fp16, [512]> encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4738176)))];
+            tensor<fp16, []> var_38_to_fp16 = const()[name = tensor<string, []>("op_38_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 17, 512]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_192_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_0_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4739264)))];
+            tensor<fp16, [2048]> linear_1_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_1_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6836480)))];
+            tensor<fp16, [1, 17, 2048]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_0_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6840640)))];
+            tensor<fp16, [512]> linear_2_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_2_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8937856)))];
+            tensor<fp16, [1, 17, 512]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<fp16, []> var_303_to_fp16 = const()[name = tensor<string, []>("op_303_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_304_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_303_to_fp16)[name = tensor<string, []>("op_304_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_37_cast_fp16 = add(x = var_192_cast_fp16, y = var_304_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
+            tensor<int32, [1]> key_1_axes_0 = const()[name = tensor<string, []>("key_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8938944)))];
+            tensor<fp16, [512]> encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8940032)))];
+            tensor<fp16, [1, 17, 512]> key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")];
+            tensor<bool, []> input_39_interleave_0 = const()[name = tensor<string, []>("input_39_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 87, 512]> input_39_cast_fp16 = concat(axis = var_64, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = tensor<string, []>("input_39_cast_fp16")];
+            tensor<int32, [3]> var_326_begin_0 = const()[name = tensor<string, []>("op_326_begin_0"), val = tensor<int32, [3]>([0, 17, 0])];
+            tensor<int32, [3]> var_326_end_0 = const()[name = tensor<string, []>("op_326_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_326_end_mask_0 = const()[name = tensor<string, []>("op_326_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 53, 512]> var_326_cast_fp16 = slice_by_index(begin = var_326_begin_0, end = var_326_end_0, end_mask = var_326_end_mask_0, x = cache_1_cast_fp16)[name = tensor<string, []>("op_326_cast_fp16")];
+            tensor<bool, []> var_332_interleave_0 = const()[name = tensor<string, []>("op_332_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_332_cast_fp16 = concat(axis = var_64, interleave = var_332_interleave_0, values = (var_326_cast_fp16, key_1_cast_fp16))[name = tensor<string, []>("op_332_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_0_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8941120)))];
+            tensor<fp16, [1, 17, 512]> linear_3_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16, x = key_1_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
+            tensor<int32, [4]> var_336 = const()[name = tensor<string, []>("op_336"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 17, 8, 64]> q_1_cast_fp16 = reshape(shape = var_336, x = linear_3_cast_fp16)[name = tensor<string, []>("q_1_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_0_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9465472)))];
+            tensor<fp16, [1, 87, 512]> linear_4_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
+            tensor<int32, [4]> var_340 = const()[name = tensor<string, []>("op_340"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> k_1_cast_fp16 = reshape(shape = var_340, x = linear_4_cast_fp16)[name = tensor<string, []>("k_1_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_0_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9989824)))];
+            tensor<fp16, [1, 87, 512]> linear_5_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
+            tensor<int32, [4]> var_344 = const()[name = tensor<string, []>("op_344"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> v_1_cast_fp16 = reshape(shape = var_344, x = linear_5_cast_fp16)[name = tensor<string, []>("v_1_cast_fp16")];
+            tensor<int32, [4]> value_3_perm_0 = const()[name = tensor<string, []>("value_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10514176)))];
+            tensor<fp16, [1, 17, 8, 64]> var_356_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_356_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10515264)))];
+            tensor<fp16, [1, 17, 8, 64]> var_358_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_358_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_1_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_7_transpose_x_0 = const()[name = tensor<string, []>("x_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_7_transpose_y_0 = const()[name = tensor<string, []>("x_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 173]> var_360_to_fp16 = const()[name = tensor<string, []>("op_360_to_fp16"), val = tensor<fp16, [1, 8, 64, 173]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10516352)))];
+            tensor<fp16, [1, 8, 17, 64]> q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_358_cast_fp16)[name = tensor<string, []>("transpose_237")];
+            tensor<fp16, [1, 8, 17, 173]> x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = var_360_to_fp16)[name = tensor<string, []>("x_7_cast_fp16")];
+            tensor<int32, [8]> x_9_pad_0 = const()[name = tensor<string, []>("x_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_9_mode_0 = const()[name = tensor<string, []>("x_9_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_23_to_fp16 = const()[name = tensor<string, []>("const_23_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 17, 174]> x_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7_cast_fp16)[name = tensor<string, []>("x_9_cast_fp16")];
+            tensor<int32, [4]> var_368 = const()[name = tensor<string, []>("op_368"), val = tensor<int32, [4]>([1, 8, -1, 17])];
+            tensor<fp16, [1, 8, 174, 17]> x_11_cast_fp16 = reshape(shape = var_368, x = x_9_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
+            tensor<int32, [4]> var_372_begin_0 = const()[name = tensor<string, []>("op_372_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_372_end_0 = const()[name = tensor<string, []>("op_372_end_0"), val = tensor<int32, [4]>([1, 8, 174, 17])];
+            tensor<bool, [4]> var_372_end_mask_0 = const()[name = tensor<string, []>("op_372_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 173, 17]> var_372_cast_fp16 = slice_by_index(begin = var_372_begin_0, end = var_372_end_0, end_mask = var_372_end_mask_0, x = x_11_cast_fp16)[name = tensor<string, []>("op_372_cast_fp16")];
+            tensor<int32, [4]> var_373 = const()[name = tensor<string, []>("op_373"), val = tensor<int32, [4]>([1, 8, 17, 173])];
+            tensor<fp16, [1, 8, 17, 173]> matrix_bd_1_cast_fp16 = reshape(shape = var_373, x = var_372_cast_fp16)[name = tensor<string, []>("matrix_bd_1_cast_fp16")];
+            tensor<bool, []> matrix_ac_1_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_1_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_51_perm_0 = const()[name = tensor<string, []>("transpose_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_52_perm_0 = const()[name = tensor<string, []>("transpose_52_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 87]> transpose_52 = transpose(perm = transpose_52_perm_0, x = k_1_cast_fp16)[name = tensor<string, []>("transpose_235")];
+            tensor<fp16, [1, 8, 17, 64]> transpose_51 = transpose(perm = transpose_51_perm_0, x = var_356_cast_fp16)[name = tensor<string, []>("transpose_236")];
+            tensor<fp16, [1, 8, 17, 87]> matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_51, y = transpose_52)[name = tensor<string, []>("matrix_ac_1_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_3_begin_0 = const()[name = tensor<string, []>("matrix_bd_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_3_end_0 = const()[name = tensor<string, []>("matrix_bd_3_end_0"), val = tensor<int32, [4]>([1, 8, 17, 87])];
+            tensor<bool, [4]> matrix_bd_3_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 17, 87]> matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1_cast_fp16)[name = tensor<string, []>("matrix_bd_3_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_382_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = tensor<string, []>("op_382_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_1_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_1_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 17, 87]> _inversed_scores_1_cast_fp16 = mul(x = var_382_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_1_cast_fp16")];
+            tensor<int32, [1]> mask_3_axes_0 = const()[name = tensor<string, []>("mask_3_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 17, 87]> mask_3 = expand_dims(axes = mask_3_axes_0, x = mask_1)[name = tensor<string, []>("mask_3")];
+            tensor<fp16, []> var_41_to_fp16 = const()[name = tensor<string, []>("op_41_to_fp16"), val = tensor<fp16, []>(-0x1.388p+13)];
+            tensor<fp16, [1, 8, 17, 87]> scores_3_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_3_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_388_cast_fp16 = softmax(axis = var_62, x = scores_3_cast_fp16)[name = tensor<string, []>("op_388_cast_fp16")];
+            tensor<fp16, []> var_40_to_fp16 = const()[name = tensor<string, []>("op_40_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 17, 87]> input_41_cast_fp16 = select(a = var_40_to_fp16, b = var_388_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_41_cast_fp16")];
+            tensor<bool, []> x_13_transpose_x_0 = const()[name = tensor<string, []>("x_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_13_transpose_y_0 = const()[name = tensor<string, []>("x_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 87, 64]> value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = v_1_cast_fp16)[name = tensor<string, []>("transpose_238")];
+            tensor<fp16, [1, 8, 17, 64]> x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41_cast_fp16, y = value_3_cast_fp16)[name = tensor<string, []>("x_13_cast_fp16")];
+            tensor<int32, [4]> var_392_perm_0 = const()[name = tensor<string, []>("op_392_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_393 = const()[name = tensor<string, []>("op_393"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 17, 8, 64]> var_392_cast_fp16 = transpose(perm = var_392_perm_0, x = x_13_cast_fp16)[name = tensor<string, []>("transpose_234")];
+            tensor<fp16, [1, 17, 512]> input_43_cast_fp16 = reshape(shape = var_393, x = var_392_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_0_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10693568)))];
+            tensor<fp16, [1, 17, 512]> linear_7_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
+            tensor<int32, [1]> x_17_axes_0 = const()[name = tensor<string, []>("x_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11217920)))];
+            tensor<fp16, [512]> encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11219008)))];
+            tensor<fp16, [1, 17, 512]> x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("x_17_cast_fp16")];
+            tensor<int32, [3]> input_49_perm_0 = const()[name = tensor<string, []>("input_49_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_51_pad_type_0 = const()[name = tensor<string, []>("input_51_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_51_strides_0 = const()[name = tensor<string, []>("input_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_51_pad_0 = const()[name = tensor<string, []>("input_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_51_dilations_0 = const()[name = tensor<string, []>("input_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_51_groups_0 = const()[name = tensor<string, []>("input_51_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_0_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11220096)))];
+            tensor<fp16, [1, 512, 17]> input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = tensor<string, []>("transpose_233")];
+            tensor<fp16, [1, 1024, 17]> input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
+            tensor<int32, []> x_19_split_num_splits_0 = const()[name = tensor<string, []>("x_19_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_19_split_axis_0 = const()[name = tensor<string, []>("x_19_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 17]> x_19_split_cast_fp16_0, tensor<fp16, [1, 512, 17]> x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51_cast_fp16)[name = tensor<string, []>("x_19_split_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = tensor<string, []>("x_19_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")];
+            tensor<int32, [1]> var_418_axes_0 = const()[name = tensor<string, []>("op_418_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 17]> var_418 = expand_dims(axes = var_418_axes_0, x = pad_mask)[name = tensor<string, []>("op_418")];
+            tensor<fp16, [1, 512, 17]> input_53_cast_fp16 = select(a = var_40_to_fp16, b = x_19_cast_fp16, cond = var_418)[name = tensor<string, []>("input_53_cast_fp16")];
+            tensor<bool, []> new_x_3_interleave_0 = const()[name = tensor<string, []>("new_x_3_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 25]> new_x_3_cast_fp16 = concat(axis = var_62, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = tensor<string, []>("new_x_3_cast_fp16")];
+            tensor<int32, [3]> var_431_begin_0 = const()[name = tensor<string, []>("op_431_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
+            tensor<int32, [3]> var_431_end_0 = const()[name = tensor<string, []>("op_431_end_0"), val = tensor<int32, [3]>([1, 512, 25])];
+            tensor<bool, [3]> var_431_end_mask_0 = const()[name = tensor<string, []>("op_431_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_431_cast_fp16 = slice_by_index(begin = var_431_begin_0, end = var_431_end_0, end_mask = var_431_end_mask_0, x = new_x_3_cast_fp16)[name = tensor<string, []>("op_431_cast_fp16")];
+            tensor<string, []> x_21_pad_type_0 = const()[name = tensor<string, []>("x_21_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_21_groups_0 = const()[name = tensor<string, []>("x_21_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_21_strides_0 = const()[name = tensor<string, []>("x_21_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_21_pad_0 = const()[name = tensor<string, []>("x_21_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_21_dilations_0 = const()[name = tensor<string, []>("x_21_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_0_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12268736)))];
+            tensor<fp16, [1, 512, 17]> x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16, x = new_x_3_cast_fp16)[name = tensor<string, []>("x_21_cast_fp16")];
+            tensor<int32, [3]> input_55_perm_0 = const()[name = tensor<string, []>("input_55_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_23_axes_0 = const()[name = tensor<string, []>("x_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12278016)))];
+            tensor<fp16, [512]> encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12279104)))];
+            tensor<fp16, [1, 17, 512]> input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = tensor<string, []>("transpose_232")];
+            tensor<fp16, [1, 17, 512]> x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")];
+            tensor<int32, [3]> input_57_perm_0 = const()[name = tensor<string, []>("input_57_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 17]> input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = tensor<string, []>("transpose_231")];
+            tensor<fp16, [1, 512, 17]> input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
+            tensor<string, []> x_25_pad_type_0 = const()[name = tensor<string, []>("x_25_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_25_strides_0 = const()[name = tensor<string, []>("x_25_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_25_pad_0 = const()[name = tensor<string, []>("x_25_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_25_dilations_0 = const()[name = tensor<string, []>("x_25_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_25_groups_0 = const()[name = tensor<string, []>("x_25_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_0_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12280192)))];
+            tensor<fp16, [1, 512, 17]> x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("x_25_cast_fp16")];
+            tensor<int32, [3]> input_61_perm_0 = const()[name = tensor<string, []>("input_61_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 17, 512]> input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = tensor<string, []>("transpose_230")];
+            tensor<fp16, [1, 17, 512]> input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
+            tensor<int32, [1]> input_65_axes_0 = const()[name = tensor<string, []>("input_65_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12804544)))];
+            tensor<fp16, [512]> encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12805632)))];
+            tensor<fp16, [1, 17, 512]> input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_0_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12806720)))];
+            tensor<fp16, [1, 17, 2048]> linear_8_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_0_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14903936)))];
+            tensor<fp16, [1, 17, 512]> linear_9_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
+            tensor<fp16, []> var_472_to_fp16 = const()[name = tensor<string, []>("op_472_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_473_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_472_to_fp16)[name = tensor<string, []>("op_473_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_473_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
+            tensor<int32, [1]> input_77_axes_0 = const()[name = tensor<string, []>("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_0_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17001152)))];
+            tensor<fp16, [512]> encoder_layers_0_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17002240)))];
+            tensor<fp16, [1, 17, 512]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
+            tensor<int32, [4]> cache_5_begin_0 = const()[name = tensor<string, []>("cache_5_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_5_end_0 = const()[name = tensor<string, []>("cache_5_end_0"), val = tensor<int32, [4]>([2, 1, 70, 512])];
+            tensor<bool, [4]> cache_5_end_mask_0 = const()[name = tensor<string, []>("cache_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_5_squeeze_mask_0 = const()[name = tensor<string, []>("cache_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_5_cast_fp16 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_5_cast_fp16")];
+            tensor<int32, [4]> cache_7_begin_0 = const()[name = tensor<string, []>("cache_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_7_end_0 = const()[name = tensor<string, []>("cache_7_end_0"), val = tensor<int32, [4]>([2, 1, 512, 8])];
+            tensor<bool, [4]> cache_7_end_mask_0 = const()[name = tensor<string, []>("cache_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_7_squeeze_mask_0 = const()[name = tensor<string, []>("cache_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_7_cast_fp16 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_7_cast_fp16")];
+            tensor<int32, [1]> input_79_axes_0 = const()[name = tensor<string, []>("input_79_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17003328)))];
+            tensor<fp16, [512]> encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17004416)))];
+            tensor<fp16, [1, 17, 512]> input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_1_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17005504)))];
+            tensor<fp16, [1, 17, 2048]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_1_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19102720)))];
+            tensor<fp16, [1, 17, 512]> linear_11_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
+            tensor<fp16, []> var_507_to_fp16 = const()[name = tensor<string, []>("op_507_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_508_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_507_to_fp16)[name = tensor<string, []>("op_508_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_508_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
+            tensor<int32, [1]> key_3_axes_0 = const()[name = tensor<string, []>("key_3_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21199936)))];
+            tensor<fp16, [512]> encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21201024)))];
+            tensor<fp16, [1, 17, 512]> key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("key_3_cast_fp16")];
+            tensor<bool, []> input_91_interleave_0 = const()[name = tensor<string, []>("input_91_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 87, 512]> input_91_cast_fp16 = concat(axis = var_64, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = tensor<string, []>("input_91_cast_fp16")];
+            tensor<int32, [3]> var_530_begin_0 = const()[name = tensor<string, []>("op_530_begin_0"), val = tensor<int32, [3]>([0, 17, 0])];
+            tensor<int32, [3]> var_530_end_0 = const()[name = tensor<string, []>("op_530_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_530_end_mask_0 = const()[name = tensor<string, []>("op_530_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 53, 512]> var_530_cast_fp16 = slice_by_index(begin = var_530_begin_0, end = var_530_end_0, end_mask = var_530_end_mask_0, x = cache_5_cast_fp16)[name = tensor<string, []>("op_530_cast_fp16")];
+            tensor<bool, []> var_536_interleave_0 = const()[name = tensor<string, []>("op_536_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_536_cast_fp16 = concat(axis = var_64, interleave = var_536_interleave_0, values = (var_530_cast_fp16, key_3_cast_fp16))[name = tensor<string, []>("op_536_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_1_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21202112)))];
+            tensor<fp16, [1, 17, 512]> linear_12_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16, x = key_3_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
+            tensor<int32, [4]> var_540 = const()[name = tensor<string, []>("op_540"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 17, 8, 64]> q_7_cast_fp16 = reshape(shape = var_540, x = linear_12_cast_fp16)[name = tensor<string, []>("q_7_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_1_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21726464)))];
+            tensor<fp16, [1, 87, 512]> linear_13_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
+            tensor<int32, [4]> var_544 = const()[name = tensor<string, []>("op_544"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> k_5_cast_fp16 = reshape(shape = var_544, x = linear_13_cast_fp16)[name = tensor<string, []>("k_5_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_1_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22250816)))];
+            tensor<fp16, [1, 87, 512]> linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
+            tensor<int32, [4]> var_548 = const()[name = tensor<string, []>("op_548"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> v_3_cast_fp16 = reshape(shape = var_548, x = linear_14_cast_fp16)[name = tensor<string, []>("v_3_cast_fp16")];
+            tensor<int32, [4]> value_5_perm_0 = const()[name = tensor<string, []>("value_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22775168)))];
+            tensor<fp16, [1, 17, 8, 64]> var_560_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_560_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22776256)))];
+            tensor<fp16, [1, 17, 8, 64]> var_562_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_562_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_3_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_33_transpose_x_0 = const()[name = tensor<string, []>("x_33_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_33_transpose_y_0 = const()[name = tensor<string, []>("x_33_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 173]> var_564_to_fp16 = const()[name = tensor<string, []>("op_564_to_fp16"), val = tensor<fp16, [1, 8, 64, 173]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22777344)))];
+            tensor<fp16, [1, 8, 17, 64]> q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_562_cast_fp16)[name = tensor<string, []>("transpose_228")];
+            tensor<fp16, [1, 8, 17, 173]> x_33_cast_fp16 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = var_564_to_fp16)[name = tensor<string, []>("x_33_cast_fp16")];
+            tensor<int32, [8]> x_35_pad_0 = const()[name = tensor<string, []>("x_35_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_35_mode_0 = const()[name = tensor<string, []>("x_35_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_36_to_fp16 = const()[name = tensor<string, []>("const_36_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 17, 174]> x_35_cast_fp16 = pad(constant_val = const_36_to_fp16, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")];
+            tensor<int32, [4]> var_572 = const()[name = tensor<string, []>("op_572"), val = tensor<int32, [4]>([1, 8, -1, 17])];
+            tensor<fp16, [1, 8, 174, 17]> x_37_cast_fp16 = reshape(shape = var_572, x = x_35_cast_fp16)[name = tensor<string, []>("x_37_cast_fp16")];
+            tensor<int32, [4]> var_576_begin_0 = const()[name = tensor<string, []>("op_576_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_576_end_0 = const()[name = tensor<string, []>("op_576_end_0"), val = tensor<int32, [4]>([1, 8, 174, 17])];
+            tensor<bool, [4]> var_576_end_mask_0 = const()[name = tensor<string, []>("op_576_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 173, 17]> var_576_cast_fp16 = slice_by_index(begin = var_576_begin_0, end = var_576_end_0, end_mask = var_576_end_mask_0, x = x_37_cast_fp16)[name = tensor<string, []>("op_576_cast_fp16")];
+            tensor<int32, [4]> var_577 = const()[name = tensor<string, []>("op_577"), val = tensor<int32, [4]>([1, 8, 17, 173])];
+            tensor<fp16, [1, 8, 17, 173]> matrix_bd_5_cast_fp16 = reshape(shape = var_577, x = var_576_cast_fp16)[name = tensor<string, []>("matrix_bd_5_cast_fp16")];
+            tensor<bool, []> matrix_ac_3_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_3_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_3_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_3_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_53_perm_0 = const()[name = tensor<string, []>("transpose_53_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_54_perm_0 = const()[name = tensor<string, []>("transpose_54_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 87]> transpose_54 = transpose(perm = transpose_54_perm_0, x = k_5_cast_fp16)[name = tensor<string, []>("transpose_226")];
+            tensor<fp16, [1, 8, 17, 64]> transpose_53 = transpose(perm = transpose_53_perm_0, x = var_560_cast_fp16)[name = tensor<string, []>("transpose_227")];
+            tensor<fp16, [1, 8, 17, 87]> matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_53, y = transpose_54)[name = tensor<string, []>("matrix_ac_3_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_7_begin_0 = const()[name = tensor<string, []>("matrix_bd_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_7_end_0 = const()[name = tensor<string, []>("matrix_bd_7_end_0"), val = tensor<int32, [4]>([1, 8, 17, 87])];
+            tensor<bool, [4]> matrix_bd_7_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 17, 87]> matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5_cast_fp16)[name = tensor<string, []>("matrix_bd_7_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_586_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = tensor<string, []>("op_586_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_5_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_5_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 17, 87]> _inversed_scores_5_cast_fp16 = mul(x = var_586_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_5_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> scores_7_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_7_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_592_cast_fp16 = softmax(axis = var_62, x = scores_7_cast_fp16)[name = tensor<string, []>("op_592_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> input_93_cast_fp16 = select(a = var_40_to_fp16, b = var_592_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_93_cast_fp16")];
+            tensor<bool, []> x_39_transpose_x_0 = const()[name = tensor<string, []>("x_39_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_39_transpose_y_0 = const()[name = tensor<string, []>("x_39_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 87, 64]> value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = v_3_cast_fp16)[name = tensor<string, []>("transpose_229")];
+            tensor<fp16, [1, 8, 17, 64]> x_39_cast_fp16 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93_cast_fp16, y = value_5_cast_fp16)[name = tensor<string, []>("x_39_cast_fp16")];
+            tensor<int32, [4]> var_596_perm_0 = const()[name = tensor<string, []>("op_596_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_597 = const()[name = tensor<string, []>("op_597"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 17, 8, 64]> var_596_cast_fp16 = transpose(perm = var_596_perm_0, x = x_39_cast_fp16)[name = tensor<string, []>("transpose_225")];
+            tensor<fp16, [1, 17, 512]> input_95_cast_fp16 = reshape(shape = var_597, x = var_596_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_1_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22954560)))];
+            tensor<fp16, [1, 17, 512]> linear_16_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
+            tensor<int32, [1]> x_43_axes_0 = const()[name = tensor<string, []>("x_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23478912)))];
+            tensor<fp16, [512]> encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23480000)))];
+            tensor<fp16, [1, 17, 512]> x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")];
+            tensor<int32, [3]> input_101_perm_0 = const()[name = tensor<string, []>("input_101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_103_pad_type_0 = const()[name = tensor<string, []>("input_103_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_103_strides_0 = const()[name = tensor<string, []>("input_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_103_pad_0 = const()[name = tensor<string, []>("input_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_103_dilations_0 = const()[name = tensor<string, []>("input_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_103_groups_0 = const()[name = tensor<string, []>("input_103_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_1_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23481088)))];
+            tensor<fp16, [1, 512, 17]> input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = tensor<string, []>("transpose_224")];
+            tensor<fp16, [1, 1024, 17]> input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
+            tensor<int32, []> x_45_split_num_splits_0 = const()[name = tensor<string, []>("x_45_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_45_split_axis_0 = const()[name = tensor<string, []>("x_45_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 17]> x_45_split_cast_fp16_0, tensor<fp16, [1, 512, 17]> x_45_split_cast_fp16_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103_cast_fp16)[name = tensor<string, []>("x_45_split_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = tensor<string, []>("x_45_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_45_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> input_105_cast_fp16 = select(a = var_40_to_fp16, b = x_45_cast_fp16, cond = var_418)[name = tensor<string, []>("input_105_cast_fp16")];
+            tensor<bool, []> new_x_7_interleave_0 = const()[name = tensor<string, []>("new_x_7_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 25]> new_x_7_cast_fp16 = concat(axis = var_62, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = tensor<string, []>("new_x_7_cast_fp16")];
+            tensor<int32, [3]> var_635_begin_0 = const()[name = tensor<string, []>("op_635_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
+            tensor<int32, [3]> var_635_end_0 = const()[name = tensor<string, []>("op_635_end_0"), val = tensor<int32, [3]>([1, 512, 25])];
+            tensor<bool, [3]> var_635_end_mask_0 = const()[name = tensor<string, []>("op_635_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_635_cast_fp16 = slice_by_index(begin = var_635_begin_0, end = var_635_end_0, end_mask = var_635_end_mask_0, x = new_x_7_cast_fp16)[name = tensor<string, []>("op_635_cast_fp16")];
+            tensor<string, []> x_47_pad_type_0 = const()[name = tensor<string, []>("x_47_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_47_groups_0 = const()[name = tensor<string, []>("x_47_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_47_strides_0 = const()[name = tensor<string, []>("x_47_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_47_pad_0 = const()[name = tensor<string, []>("x_47_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_47_dilations_0 = const()[name = tensor<string, []>("x_47_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_1_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24529728)))];
+            tensor<fp16, [1, 512, 17]> x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16, x = new_x_7_cast_fp16)[name = tensor<string, []>("x_47_cast_fp16")];
+            tensor<int32, [3]> input_107_perm_0 = const()[name = tensor<string, []>("input_107_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_49_axes_0 = const()[name = tensor<string, []>("x_49_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24539008)))];
+            tensor<fp16, [512]> encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24540096)))];
+            tensor<fp16, [1, 17, 512]> input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = tensor<string, []>("transpose_223")];
+            tensor<fp16, [1, 17, 512]> x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = tensor<string, []>("x_49_cast_fp16")];
+            tensor<int32, [3]> input_109_perm_0 = const()[name = tensor<string, []>("input_109_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 17]> input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = tensor<string, []>("transpose_222")];
+            tensor<fp16, [1, 512, 17]> input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
+            tensor<string, []> x_51_pad_type_0 = const()[name = tensor<string, []>("x_51_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_51_strides_0 = const()[name = tensor<string, []>("x_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_51_pad_0 = const()[name = tensor<string, []>("x_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_51_dilations_0 = const()[name = tensor<string, []>("x_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_51_groups_0 = const()[name = tensor<string, []>("x_51_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_1_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24541184)))];
+            tensor<fp16, [1, 512, 17]> x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("x_51_cast_fp16")];
+            tensor<int32, [3]> input_113_perm_0 = const()[name = tensor<string, []>("input_113_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 17, 512]> input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = tensor<string, []>("transpose_221")];
+            tensor<fp16, [1, 17, 512]> input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
+            tensor<int32, [1]> input_117_axes_0 = const()[name = tensor<string, []>("input_117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25065536)))];
+            tensor<fp16, [512]> encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25066624)))];
+            tensor<fp16, [1, 17, 512]> input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_1_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25067712)))];
+            tensor<fp16, [1, 17, 2048]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16, x = input_117_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_1_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27164928)))];
+            tensor<fp16, [1, 17, 512]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
+            tensor<fp16, []> var_676_to_fp16 = const()[name = tensor<string, []>("op_676_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_677_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_676_to_fp16)[name = tensor<string, []>("op_677_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_677_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
+            tensor<int32, [1]> input_129_axes_0 = const()[name = tensor<string, []>("input_129_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_1_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29262144)))];
+            tensor<fp16, [512]> encoder_layers_1_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29263232)))];
+            tensor<fp16, [1, 17, 512]> input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
+            tensor<int32, [4]> cache_9_begin_0 = const()[name = tensor<string, []>("cache_9_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_9_end_0 = const()[name = tensor<string, []>("cache_9_end_0"), val = tensor<int32, [4]>([3, 1, 70, 512])];
+            tensor<bool, [4]> cache_9_end_mask_0 = const()[name = tensor<string, []>("cache_9_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_9_squeeze_mask_0 = const()[name = tensor<string, []>("cache_9_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_9_cast_fp16 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_9_cast_fp16")];
+            tensor<int32, [4]> cache_11_begin_0 = const()[name = tensor<string, []>("cache_11_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_11_end_0 = const()[name = tensor<string, []>("cache_11_end_0"), val = tensor<int32, [4]>([3, 1, 512, 8])];
+            tensor<bool, [4]> cache_11_end_mask_0 = const()[name = tensor<string, []>("cache_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_11_squeeze_mask_0 = const()[name = tensor<string, []>("cache_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_11_cast_fp16 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_11_cast_fp16")];
+            tensor<int32, [1]> input_131_axes_0 = const()[name = tensor<string, []>("input_131_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29264320)))];
+            tensor<fp16, [512]> encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29265408)))];
+            tensor<fp16, [1, 17, 512]> input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("input_131_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_2_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29266496)))];
+            tensor<fp16, [1, 17, 2048]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16, x = input_131_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_2_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31363712)))];
+            tensor<fp16, [1, 17, 512]> linear_20_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16, x = input_135_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
+            tensor<fp16, []> var_711_to_fp16 = const()[name = tensor<string, []>("op_711_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_712_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_711_to_fp16)[name = tensor<string, []>("op_712_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_712_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
+            tensor<int32, [1]> key_5_axes_0 = const()[name = tensor<string, []>("key_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33460928)))];
+            tensor<fp16, [512]> encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33462016)))];
+            tensor<fp16, [1, 17, 512]> key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")];
+            tensor<bool, []> input_143_interleave_0 = const()[name = tensor<string, []>("input_143_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 87, 512]> input_143_cast_fp16 = concat(axis = var_64, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = tensor<string, []>("input_143_cast_fp16")];
+            tensor<int32, [3]> var_734_begin_0 = const()[name = tensor<string, []>("op_734_begin_0"), val = tensor<int32, [3]>([0, 17, 0])];
+            tensor<int32, [3]> var_734_end_0 = const()[name = tensor<string, []>("op_734_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_734_end_mask_0 = const()[name = tensor<string, []>("op_734_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 53, 512]> var_734_cast_fp16 = slice_by_index(begin = var_734_begin_0, end = var_734_end_0, end_mask = var_734_end_mask_0, x = cache_9_cast_fp16)[name = tensor<string, []>("op_734_cast_fp16")];
+            tensor<bool, []> var_740_interleave_0 = const()[name = tensor<string, []>("op_740_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_740_cast_fp16 = concat(axis = var_64, interleave = var_740_interleave_0, values = (var_734_cast_fp16, key_5_cast_fp16))[name = tensor<string, []>("op_740_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_2_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33463104)))];
+            tensor<fp16, [1, 17, 512]> linear_21_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16, x = key_5_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
+            tensor<int32, [4]> var_744 = const()[name = tensor<string, []>("op_744"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 17, 8, 64]> q_13_cast_fp16 = reshape(shape = var_744, x = linear_21_cast_fp16)[name = tensor<string, []>("q_13_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_2_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33987456)))];
+            tensor<fp16, [1, 87, 512]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
+            tensor<int32, [4]> var_748 = const()[name = tensor<string, []>("op_748"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> k_9_cast_fp16 = reshape(shape = var_748, x = linear_22_cast_fp16)[name = tensor<string, []>("k_9_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_2_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34511808)))];
+            tensor<fp16, [1, 87, 512]> linear_23_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
+            tensor<int32, [4]> var_752 = const()[name = tensor<string, []>("op_752"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> v_5_cast_fp16 = reshape(shape = var_752, x = linear_23_cast_fp16)[name = tensor<string, []>("v_5_cast_fp16")];
+            tensor<int32, [4]> value_7_perm_0 = const()[name = tensor<string, []>("value_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35036160)))];
+            tensor<fp16, [1, 17, 8, 64]> var_764_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_764_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35037248)))];
+            tensor<fp16, [1, 17, 8, 64]> var_766_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_766_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_5_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_59_transpose_x_0 = const()[name = tensor<string, []>("x_59_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_59_transpose_y_0 = const()[name = tensor<string, []>("x_59_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 173]> var_768_to_fp16 = const()[name = tensor<string, []>("op_768_to_fp16"), val = tensor<fp16, [1, 8, 64, 173]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35038336)))];
+            tensor<fp16, [1, 8, 17, 64]> q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_766_cast_fp16)[name = tensor<string, []>("transpose_219")];
+            tensor<fp16, [1, 8, 17, 173]> x_59_cast_fp16 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = var_768_to_fp16)[name = tensor<string, []>("x_59_cast_fp16")];
+            tensor<int32, [8]> x_61_pad_0 = const()[name = tensor<string, []>("x_61_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_61_mode_0 = const()[name = tensor<string, []>("x_61_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_49_to_fp16 = const()[name = tensor<string, []>("const_49_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 17, 174]> x_61_cast_fp16 = pad(constant_val = const_49_to_fp16, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59_cast_fp16)[name = tensor<string, []>("x_61_cast_fp16")];
+            tensor<int32, [4]> var_776 = const()[name = tensor<string, []>("op_776"), val = tensor<int32, [4]>([1, 8, -1, 17])];
+            tensor<fp16, [1, 8, 174, 17]> x_63_cast_fp16 = reshape(shape = var_776, x = x_61_cast_fp16)[name = tensor<string, []>("x_63_cast_fp16")];
+            tensor<int32, [4]> var_780_begin_0 = const()[name = tensor<string, []>("op_780_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_780_end_0 = const()[name = tensor<string, []>("op_780_end_0"), val = tensor<int32, [4]>([1, 8, 174, 17])];
+            tensor<bool, [4]> var_780_end_mask_0 = const()[name = tensor<string, []>("op_780_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 173, 17]> var_780_cast_fp16 = slice_by_index(begin = var_780_begin_0, end = var_780_end_0, end_mask = var_780_end_mask_0, x = x_63_cast_fp16)[name = tensor<string, []>("op_780_cast_fp16")];
+            tensor<int32, [4]> var_781 = const()[name = tensor<string, []>("op_781"), val = tensor<int32, [4]>([1, 8, 17, 173])];
+            tensor<fp16, [1, 8, 17, 173]> matrix_bd_9_cast_fp16 = reshape(shape = var_781, x = var_780_cast_fp16)[name = tensor<string, []>("matrix_bd_9_cast_fp16")];
+            tensor<bool, []> matrix_ac_5_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_5_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_5_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_5_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_55_perm_0 = const()[name = tensor<string, []>("transpose_55_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_56_perm_0 = const()[name = tensor<string, []>("transpose_56_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 87]> transpose_56 = transpose(perm = transpose_56_perm_0, x = k_9_cast_fp16)[name = tensor<string, []>("transpose_217")];
+            tensor<fp16, [1, 8, 17, 64]> transpose_55 = transpose(perm = transpose_55_perm_0, x = var_764_cast_fp16)[name = tensor<string, []>("transpose_218")];
+            tensor<fp16, [1, 8, 17, 87]> matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_55, y = transpose_56)[name = tensor<string, []>("matrix_ac_5_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_11_begin_0 = const()[name = tensor<string, []>("matrix_bd_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_11_end_0 = const()[name = tensor<string, []>("matrix_bd_11_end_0"), val = tensor<int32, [4]>([1, 8, 17, 87])];
+            tensor<bool, [4]> matrix_bd_11_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 17, 87]> matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9_cast_fp16)[name = tensor<string, []>("matrix_bd_11_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_790_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = tensor<string, []>("op_790_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_9_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_9_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 17, 87]> _inversed_scores_9_cast_fp16 = mul(x = var_790_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_9_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> scores_11_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_11_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_796_cast_fp16 = softmax(axis = var_62, x = scores_11_cast_fp16)[name = tensor<string, []>("op_796_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> input_145_cast_fp16 = select(a = var_40_to_fp16, b = var_796_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_145_cast_fp16")];
+            tensor<bool, []> x_65_transpose_x_0 = const()[name = tensor<string, []>("x_65_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_65_transpose_y_0 = const()[name = tensor<string, []>("x_65_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 87, 64]> value_7_cast_fp16 = transpose(perm = value_7_perm_0, x = v_5_cast_fp16)[name = tensor<string, []>("transpose_220")];
+            tensor<fp16, [1, 8, 17, 64]> x_65_cast_fp16 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145_cast_fp16, y = value_7_cast_fp16)[name = tensor<string, []>("x_65_cast_fp16")];
+            tensor<int32, [4]> var_800_perm_0 = const()[name = tensor<string, []>("op_800_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_801 = const()[name = tensor<string, []>("op_801"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 17, 8, 64]> var_800_cast_fp16 = transpose(perm = var_800_perm_0, x = x_65_cast_fp16)[name = tensor<string, []>("transpose_216")];
+            tensor<fp16, [1, 17, 512]> input_147_cast_fp16 = reshape(shape = var_801, x = var_800_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_2_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35215552)))];
+            tensor<fp16, [1, 17, 512]> linear_25_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")];
+            tensor<int32, [1]> x_69_axes_0 = const()[name = tensor<string, []>("x_69_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35739904)))];
+            tensor<fp16, [512]> encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35740992)))];
+            tensor<fp16, [1, 17, 512]> x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = tensor<string, []>("x_69_cast_fp16")];
+            tensor<int32, [3]> input_153_perm_0 = const()[name = tensor<string, []>("input_153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_155_pad_type_0 = const()[name = tensor<string, []>("input_155_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_155_strides_0 = const()[name = tensor<string, []>("input_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_155_pad_0 = const()[name = tensor<string, []>("input_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_155_dilations_0 = const()[name = tensor<string, []>("input_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_155_groups_0 = const()[name = tensor<string, []>("input_155_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_2_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35742080)))];
+            tensor<fp16, [1, 512, 17]> input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = tensor<string, []>("transpose_215")];
+            tensor<fp16, [1, 1024, 17]> input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor<string, []>("input_155_cast_fp16")];
+            tensor<int32, []> x_71_split_num_splits_0 = const()[name = tensor<string, []>("x_71_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_71_split_axis_0 = const()[name = tensor<string, []>("x_71_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 17]> x_71_split_cast_fp16_0, tensor<fp16, [1, 512, 17]> x_71_split_cast_fp16_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155_cast_fp16)[name = tensor<string, []>("x_71_split_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = tensor<string, []>("x_71_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_71_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> input_157_cast_fp16 = select(a = var_40_to_fp16, b = x_71_cast_fp16, cond = var_418)[name = tensor<string, []>("input_157_cast_fp16")];
+            tensor<bool, []> new_x_11_interleave_0 = const()[name = tensor<string, []>("new_x_11_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 25]> new_x_11_cast_fp16 = concat(axis = var_62, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = tensor<string, []>("new_x_11_cast_fp16")];
+            tensor<int32, [3]> var_839_begin_0 = const()[name = tensor<string, []>("op_839_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
+            tensor<int32, [3]> var_839_end_0 = const()[name = tensor<string, []>("op_839_end_0"), val = tensor<int32, [3]>([1, 512, 25])];
+            tensor<bool, [3]> var_839_end_mask_0 = const()[name = tensor<string, []>("op_839_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_839_cast_fp16 = slice_by_index(begin = var_839_begin_0, end = var_839_end_0, end_mask = var_839_end_mask_0, x = new_x_11_cast_fp16)[name = tensor<string, []>("op_839_cast_fp16")];
+            tensor<string, []> x_73_pad_type_0 = const()[name = tensor<string, []>("x_73_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_73_groups_0 = const()[name = tensor<string, []>("x_73_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_73_strides_0 = const()[name = tensor<string, []>("x_73_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_73_pad_0 = const()[name = tensor<string, []>("x_73_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_73_dilations_0 = const()[name = tensor<string, []>("x_73_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_2_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36790720)))];
+            tensor<fp16, [1, 512, 17]> x_73_cast_fp16 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16, x = new_x_11_cast_fp16)[name = tensor<string, []>("x_73_cast_fp16")];
+            tensor<int32, [3]> input_159_perm_0 = const()[name = tensor<string, []>("input_159_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_75_axes_0 = const()[name = tensor<string, []>("x_75_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36800000)))];
+            tensor<fp16, [512]> encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36801088)))];
+            tensor<fp16, [1, 17, 512]> input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = tensor<string, []>("transpose_214")];
+            tensor<fp16, [1, 17, 512]> x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = tensor<string, []>("x_75_cast_fp16")];
+            tensor<int32, [3]> input_161_perm_0 = const()[name = tensor<string, []>("input_161_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 17]> input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = tensor<string, []>("transpose_213")];
+            tensor<fp16, [1, 512, 17]> input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
+            tensor<string, []> x_77_pad_type_0 = const()[name = tensor<string, []>("x_77_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_77_strides_0 = const()[name = tensor<string, []>("x_77_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_77_pad_0 = const()[name = tensor<string, []>("x_77_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_77_dilations_0 = const()[name = tensor<string, []>("x_77_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_77_groups_0 = const()[name = tensor<string, []>("x_77_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_2_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36802176)))];
+            tensor<fp16, [1, 512, 17]> x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16, x = input_163_cast_fp16)[name = tensor<string, []>("x_77_cast_fp16")];
+            tensor<int32, [3]> input_165_perm_0 = const()[name = tensor<string, []>("input_165_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 17, 512]> input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = tensor<string, []>("transpose_212")];
+            tensor<fp16, [1, 17, 512]> input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")];
+            tensor<int32, [1]> input_169_axes_0 = const()[name = tensor<string, []>("input_169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37326528)))];
+            tensor<fp16, [512]> encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37327616)))];
+            tensor<fp16, [1, 17, 512]> input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor<string, []>("input_169_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_2_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37328704)))];
+            tensor<fp16, [1, 17, 2048]> linear_26_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16, x = input_169_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_2_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39425920)))];
+            tensor<fp16, [1, 17, 512]> linear_27_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")];
+            tensor<fp16, []> var_880_to_fp16 = const()[name = tensor<string, []>("op_880_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_881_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_880_to_fp16)[name = tensor<string, []>("op_881_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_881_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
+            tensor<int32, [1]> input_181_axes_0 = const()[name = tensor<string, []>("input_181_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_2_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41523136)))];
+            tensor<fp16, [512]> encoder_layers_2_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41524224)))];
+            tensor<fp16, [1, 17, 512]> input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")];
+            tensor<int32, [4]> cache_13_begin_0 = const()[name = tensor<string, []>("cache_13_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_13_end_0 = const()[name = tensor<string, []>("cache_13_end_0"), val = tensor<int32, [4]>([4, 1, 70, 512])];
+            tensor<bool, [4]> cache_13_end_mask_0 = const()[name = tensor<string, []>("cache_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_13_squeeze_mask_0 = const()[name = tensor<string, []>("cache_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_13_cast_fp16 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_13_cast_fp16")];
+            tensor<int32, [4]> cache_15_begin_0 = const()[name = tensor<string, []>("cache_15_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_15_end_0 = const()[name = tensor<string, []>("cache_15_end_0"), val = tensor<int32, [4]>([4, 1, 512, 8])];
+            tensor<bool, [4]> cache_15_end_mask_0 = const()[name = tensor<string, []>("cache_15_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_15_squeeze_mask_0 = const()[name = tensor<string, []>("cache_15_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_15_cast_fp16 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_15_cast_fp16")];
+            tensor<int32, [1]> input_183_axes_0 = const()[name = tensor<string, []>("input_183_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41525312)))];
+            tensor<fp16, [512]> encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41526400)))];
+            tensor<fp16, [1, 17, 512]> input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor<string, []>("input_183_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_3_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41527488)))];
+            tensor<fp16, [1, 17, 2048]> linear_28_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = tensor<string, []>("input_187_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_3_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43624704)))];
+            tensor<fp16, [1, 17, 512]> linear_29_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")];
+            tensor<fp16, []> var_915_to_fp16 = const()[name = tensor<string, []>("op_915_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_916_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_915_to_fp16)[name = tensor<string, []>("op_916_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_916_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
+            tensor<int32, [1]> key_7_axes_0 = const()[name = tensor<string, []>("key_7_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45721920)))];
+            tensor<fp16, [512]> encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45723008)))];
+            tensor<fp16, [1, 17, 512]> key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = tensor<string, []>("key_7_cast_fp16")];
+            tensor<bool, []> input_195_interleave_0 = const()[name = tensor<string, []>("input_195_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 87, 512]> input_195_cast_fp16 = concat(axis = var_64, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = tensor<string, []>("input_195_cast_fp16")];
+            tensor<int32, [3]> var_938_begin_0 = const()[name = tensor<string, []>("op_938_begin_0"), val = tensor<int32, [3]>([0, 17, 0])];
+            tensor<int32, [3]> var_938_end_0 = const()[name = tensor<string, []>("op_938_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_938_end_mask_0 = const()[name = tensor<string, []>("op_938_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 53, 512]> var_938_cast_fp16 = slice_by_index(begin = var_938_begin_0, end = var_938_end_0, end_mask = var_938_end_mask_0, x = cache_13_cast_fp16)[name = tensor<string, []>("op_938_cast_fp16")];
+            tensor<bool, []> var_944_interleave_0 = const()[name = tensor<string, []>("op_944_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_944_cast_fp16 = concat(axis = var_64, interleave = var_944_interleave_0, values = (var_938_cast_fp16, key_7_cast_fp16))[name = tensor<string, []>("op_944_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_3_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45724096)))];
+            tensor<fp16, [1, 17, 512]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16, x = key_7_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")];
+            tensor<int32, [4]> var_948 = const()[name = tensor<string, []>("op_948"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 17, 8, 64]> q_19_cast_fp16 = reshape(shape = var_948, x = linear_30_cast_fp16)[name = tensor<string, []>("q_19_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_3_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46248448)))];
+            tensor<fp16, [1, 87, 512]> linear_31_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")];
+            tensor<int32, [4]> var_952 = const()[name = tensor<string, []>("op_952"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> k_13_cast_fp16 = reshape(shape = var_952, x = linear_31_cast_fp16)[name = tensor<string, []>("k_13_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_3_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46772800)))];
+            tensor<fp16, [1, 87, 512]> linear_32_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")];
+            tensor<int32, [4]> var_956 = const()[name = tensor<string, []>("op_956"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> v_7_cast_fp16 = reshape(shape = var_956, x = linear_32_cast_fp16)[name = tensor<string, []>("v_7_cast_fp16")];
+            tensor<int32, [4]> value_9_perm_0 = const()[name = tensor<string, []>("value_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47297152)))];
+            tensor<fp16, [1, 17, 8, 64]> var_968_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_968_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47298240)))];
+            tensor<fp16, [1, 17, 8, 64]> var_970_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_970_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_7_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_85_transpose_x_0 = const()[name = tensor<string, []>("x_85_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_85_transpose_y_0 = const()[name = tensor<string, []>("x_85_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 173]> var_972_to_fp16 = const()[name = tensor<string, []>("op_972_to_fp16"), val = tensor<fp16, [1, 8, 64, 173]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47299328)))];
+            tensor<fp16, [1, 8, 17, 64]> q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_970_cast_fp16)[name = tensor<string, []>("transpose_210")];
+            tensor<fp16, [1, 8, 17, 173]> x_85_cast_fp16 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = var_972_to_fp16)[name = tensor<string, []>("x_85_cast_fp16")];
+            tensor<int32, [8]> x_87_pad_0 = const()[name = tensor<string, []>("x_87_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_87_mode_0 = const()[name = tensor<string, []>("x_87_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_62_to_fp16 = const()[name = tensor<string, []>("const_62_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 17, 174]> x_87_cast_fp16 = pad(constant_val = const_62_to_fp16, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85_cast_fp16)[name = tensor<string, []>("x_87_cast_fp16")];
+            tensor<int32, [4]> var_980 = const()[name = tensor<string, []>("op_980"), val = tensor<int32, [4]>([1, 8, -1, 17])];
+            tensor<fp16, [1, 8, 174, 17]> x_89_cast_fp16 = reshape(shape = var_980, x = x_87_cast_fp16)[name = tensor<string, []>("x_89_cast_fp16")];
+            tensor<int32, [4]> var_984_begin_0 = const()[name = tensor<string, []>("op_984_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_984_end_0 = const()[name = tensor<string, []>("op_984_end_0"), val = tensor<int32, [4]>([1, 8, 174, 17])];
+            tensor<bool, [4]> var_984_end_mask_0 = const()[name = tensor<string, []>("op_984_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 173, 17]> var_984_cast_fp16 = slice_by_index(begin = var_984_begin_0, end = var_984_end_0, end_mask = var_984_end_mask_0, x = x_89_cast_fp16)[name = tensor<string, []>("op_984_cast_fp16")];
+            tensor<int32, [4]> var_985 = const()[name = tensor<string, []>("op_985"), val = tensor<int32, [4]>([1, 8, 17, 173])];
+            tensor<fp16, [1, 8, 17, 173]> matrix_bd_13_cast_fp16 = reshape(shape = var_985, x = var_984_cast_fp16)[name = tensor<string, []>("matrix_bd_13_cast_fp16")];
+            tensor<bool, []> matrix_ac_7_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_7_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_57_perm_0 = const()[name = tensor<string, []>("transpose_57_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_58_perm_0 = const()[name = tensor<string, []>("transpose_58_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 87]> transpose_58 = transpose(perm = transpose_58_perm_0, x = k_13_cast_fp16)[name = tensor<string, []>("transpose_208")];
+            tensor<fp16, [1, 8, 17, 64]> transpose_57 = transpose(perm = transpose_57_perm_0, x = var_968_cast_fp16)[name = tensor<string, []>("transpose_209")];
+            tensor<fp16, [1, 8, 17, 87]> matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_57, y = transpose_58)[name = tensor<string, []>("matrix_ac_7_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_15_begin_0 = const()[name = tensor<string, []>("matrix_bd_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_15_end_0 = const()[name = tensor<string, []>("matrix_bd_15_end_0"), val = tensor<int32, [4]>([1, 8, 17, 87])];
+            tensor<bool, [4]> matrix_bd_15_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 17, 87]> matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13_cast_fp16)[name = tensor<string, []>("matrix_bd_15_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_994_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = tensor<string, []>("op_994_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_13_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_13_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 17, 87]> _inversed_scores_13_cast_fp16 = mul(x = var_994_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_13_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> scores_15_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_15_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_1000_cast_fp16 = softmax(axis = var_62, x = scores_15_cast_fp16)[name = tensor<string, []>("op_1000_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> input_197_cast_fp16 = select(a = var_40_to_fp16, b = var_1000_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_197_cast_fp16")];
+            tensor<bool, []> x_91_transpose_x_0 = const()[name = tensor<string, []>("x_91_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_91_transpose_y_0 = const()[name = tensor<string, []>("x_91_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 87, 64]> value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_7_cast_fp16)[name = tensor<string, []>("transpose_211")];
+            tensor<fp16, [1, 8, 17, 64]> x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197_cast_fp16, y = value_9_cast_fp16)[name = tensor<string, []>("x_91_cast_fp16")];
+            tensor<int32, [4]> var_1004_perm_0 = const()[name = tensor<string, []>("op_1004_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1005 = const()[name = tensor<string, []>("op_1005"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 17, 8, 64]> var_1004_cast_fp16 = transpose(perm = var_1004_perm_0, x = x_91_cast_fp16)[name = tensor<string, []>("transpose_207")];
+            tensor<fp16, [1, 17, 512]> input_199_cast_fp16 = reshape(shape = var_1005, x = var_1004_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_3_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47476544)))];
+            tensor<fp16, [1, 17, 512]> linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16, x = input_199_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")];
+            tensor<int32, [1]> x_95_axes_0 = const()[name = tensor<string, []>("x_95_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48000896)))];
+            tensor<fp16, [512]> encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48001984)))];
+            tensor<fp16, [1, 17, 512]> x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = tensor<string, []>("x_95_cast_fp16")];
+            tensor<int32, [3]> input_205_perm_0 = const()[name = tensor<string, []>("input_205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_207_pad_type_0 = const()[name = tensor<string, []>("input_207_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_207_strides_0 = const()[name = tensor<string, []>("input_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_207_pad_0 = const()[name = tensor<string, []>("input_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_207_dilations_0 = const()[name = tensor<string, []>("input_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_207_groups_0 = const()[name = tensor<string, []>("input_207_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_3_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48003072)))];
+            tensor<fp16, [1, 512, 17]> input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = tensor<string, []>("transpose_206")];
+            tensor<fp16, [1, 1024, 17]> input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")];
+            tensor<int32, []> x_97_split_num_splits_0 = const()[name = tensor<string, []>("x_97_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_97_split_axis_0 = const()[name = tensor<string, []>("x_97_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 17]> x_97_split_cast_fp16_0, tensor<fp16, [1, 512, 17]> x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207_cast_fp16)[name = tensor<string, []>("x_97_split_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = tensor<string, []>("x_97_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_97_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> input_209_cast_fp16 = select(a = var_40_to_fp16, b = x_97_cast_fp16, cond = var_418)[name = tensor<string, []>("input_209_cast_fp16")];
+            tensor<bool, []> new_x_15_interleave_0 = const()[name = tensor<string, []>("new_x_15_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 25]> new_x_15_cast_fp16 = concat(axis = var_62, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = tensor<string, []>("new_x_15_cast_fp16")];
+            tensor<int32, [3]> var_1043_begin_0 = const()[name = tensor<string, []>("op_1043_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
+            tensor<int32, [3]> var_1043_end_0 = const()[name = tensor<string, []>("op_1043_end_0"), val = tensor<int32, [3]>([1, 512, 25])];
+            tensor<bool, [3]> var_1043_end_mask_0 = const()[name = tensor<string, []>("op_1043_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_1043_cast_fp16 = slice_by_index(begin = var_1043_begin_0, end = var_1043_end_0, end_mask = var_1043_end_mask_0, x = new_x_15_cast_fp16)[name = tensor<string, []>("op_1043_cast_fp16")];
+            tensor<string, []> x_99_pad_type_0 = const()[name = tensor<string, []>("x_99_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_99_groups_0 = const()[name = tensor<string, []>("x_99_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_99_strides_0 = const()[name = tensor<string, []>("x_99_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_99_pad_0 = const()[name = tensor<string, []>("x_99_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_99_dilations_0 = const()[name = tensor<string, []>("x_99_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_3_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49051712)))];
+            tensor<fp16, [1, 512, 17]> x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16, x = new_x_15_cast_fp16)[name = tensor<string, []>("x_99_cast_fp16")];
+            tensor<int32, [3]> input_211_perm_0 = const()[name = tensor<string, []>("input_211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_101_axes_0 = const()[name = tensor<string, []>("x_101_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49060992)))];
+            tensor<fp16, [512]> encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49062080)))];
+            tensor<fp16, [1, 17, 512]> input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = tensor<string, []>("transpose_205")];
+            tensor<fp16, [1, 17, 512]> x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = tensor<string, []>("x_101_cast_fp16")];
+            tensor<int32, [3]> input_213_perm_0 = const()[name = tensor<string, []>("input_213_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 17]> input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = tensor<string, []>("transpose_204")];
+            tensor<fp16, [1, 512, 17]> input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = tensor<string, []>("input_215_cast_fp16")];
+            tensor<string, []> x_103_pad_type_0 = const()[name = tensor<string, []>("x_103_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_103_strides_0 = const()[name = tensor<string, []>("x_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_103_pad_0 = const()[name = tensor<string, []>("x_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_103_dilations_0 = const()[name = tensor<string, []>("x_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_103_groups_0 = const()[name = tensor<string, []>("x_103_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_3_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49063168)))];
+            tensor<fp16, [1, 512, 17]> x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor<string, []>("x_103_cast_fp16")];
+            tensor<int32, [3]> input_217_perm_0 = const()[name = tensor<string, []>("input_217_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 17, 512]> input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = tensor<string, []>("transpose_203")];
+            tensor<fp16, [1, 17, 512]> input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
+            tensor<int32, [1]> input_221_axes_0 = const()[name = tensor<string, []>("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49587520)))];
+            tensor<fp16, [512]> encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49588608)))];
+            tensor<fp16, [1, 17, 512]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = tensor<string, []>("input_221_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_3_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49589696)))];
+            tensor<fp16, [1, 17, 2048]> linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16, x = input_221_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = tensor<string, []>("input_225_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_3_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51686912)))];
+            tensor<fp16, [1, 17, 512]> linear_36_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor<string, []>("linear_36_cast_fp16")];
+            tensor<fp16, []> var_1084_to_fp16 = const()[name = tensor<string, []>("op_1084_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_1085_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1084_to_fp16)[name = tensor<string, []>("op_1085_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1085_cast_fp16)[name = tensor<string, []>("input_231_cast_fp16")];
+            tensor<int32, [1]> input_233_axes_0 = const()[name = tensor<string, []>("input_233_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_3_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53784128)))];
+            tensor<fp16, [512]> encoder_layers_3_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53785216)))];
+            tensor<fp16, [1, 17, 512]> input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = tensor<string, []>("input_233_cast_fp16")];
+            tensor<int32, [4]> cache_17_begin_0 = const()[name = tensor<string, []>("cache_17_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_17_end_0 = const()[name = tensor<string, []>("cache_17_end_0"), val = tensor<int32, [4]>([5, 1, 70, 512])];
+            tensor<bool, [4]> cache_17_end_mask_0 = const()[name = tensor<string, []>("cache_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_17_squeeze_mask_0 = const()[name = tensor<string, []>("cache_17_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_17_cast_fp16 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_17_cast_fp16")];
+            tensor<int32, [4]> cache_19_begin_0 = const()[name = tensor<string, []>("cache_19_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_19_end_0 = const()[name = tensor<string, []>("cache_19_end_0"), val = tensor<int32, [4]>([5, 1, 512, 8])];
+            tensor<bool, [4]> cache_19_end_mask_0 = const()[name = tensor<string, []>("cache_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_19_squeeze_mask_0 = const()[name = tensor<string, []>("cache_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_19_cast_fp16 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_19_cast_fp16")];
+            tensor<int32, [1]> input_235_axes_0 = const()[name = tensor<string, []>("input_235_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53786304)))];
+            tensor<fp16, [512]> encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53787392)))];
+            tensor<fp16, [1, 17, 512]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = tensor<string, []>("input_235_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_4_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53788480)))];
+            tensor<fp16, [1, 17, 2048]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16, x = input_235_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = tensor<string, []>("input_239_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_4_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55885696)))];
+            tensor<fp16, [1, 17, 512]> linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")];
+            tensor<fp16, []> var_1119_to_fp16 = const()[name = tensor<string, []>("op_1119_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_1120_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1119_to_fp16)[name = tensor<string, []>("op_1120_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1120_cast_fp16)[name = tensor<string, []>("input_245_cast_fp16")];
+            tensor<int32, [1]> key_9_axes_0 = const()[name = tensor<string, []>("key_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57982912)))];
+            tensor<fp16, [512]> encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57984000)))];
+            tensor<fp16, [1, 17, 512]> key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")];
+            tensor<bool, []> input_247_interleave_0 = const()[name = tensor<string, []>("input_247_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 87, 512]> input_247_cast_fp16 = concat(axis = var_64, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = tensor<string, []>("input_247_cast_fp16")];
+            tensor<int32, [3]> var_1142_begin_0 = const()[name = tensor<string, []>("op_1142_begin_0"), val = tensor<int32, [3]>([0, 17, 0])];
+            tensor<int32, [3]> var_1142_end_0 = const()[name = tensor<string, []>("op_1142_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_1142_end_mask_0 = const()[name = tensor<string, []>("op_1142_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 53, 512]> var_1142_cast_fp16 = slice_by_index(begin = var_1142_begin_0, end = var_1142_end_0, end_mask = var_1142_end_mask_0, x = cache_17_cast_fp16)[name = tensor<string, []>("op_1142_cast_fp16")];
+            tensor<bool, []> var_1148_interleave_0 = const()[name = tensor<string, []>("op_1148_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_1148_cast_fp16 = concat(axis = var_64, interleave = var_1148_interleave_0, values = (var_1142_cast_fp16, key_9_cast_fp16))[name = tensor<string, []>("op_1148_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_4_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57985088)))];
+            tensor<fp16, [1, 17, 512]> linear_39_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16, x = key_9_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")];
+            tensor<int32, [4]> var_1152 = const()[name = tensor<string, []>("op_1152"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 17, 8, 64]> q_25_cast_fp16 = reshape(shape = var_1152, x = linear_39_cast_fp16)[name = tensor<string, []>("q_25_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_4_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58509440)))];
+            tensor<fp16, [1, 87, 512]> linear_40_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16, x = input_247_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")];
+            tensor<int32, [4]> var_1156 = const()[name = tensor<string, []>("op_1156"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> k_17_cast_fp16 = reshape(shape = var_1156, x = linear_40_cast_fp16)[name = tensor<string, []>("k_17_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_4_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59033792)))];
+            tensor<fp16, [1, 87, 512]> linear_41_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16, x = input_247_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")];
+            tensor<int32, [4]> var_1160 = const()[name = tensor<string, []>("op_1160"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> v_9_cast_fp16 = reshape(shape = var_1160, x = linear_41_cast_fp16)[name = tensor<string, []>("v_9_cast_fp16")];
+            tensor<int32, [4]> value_11_perm_0 = const()[name = tensor<string, []>("value_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59558144)))];
+            tensor<fp16, [1, 17, 8, 64]> var_1172_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1172_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59559232)))];
+            tensor<fp16, [1, 17, 8, 64]> var_1174_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1174_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_9_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_111_transpose_x_0 = const()[name = tensor<string, []>("x_111_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_111_transpose_y_0 = const()[name = tensor<string, []>("x_111_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 173]> var_1176_to_fp16 = const()[name = tensor<string, []>("op_1176_to_fp16"), val = tensor<fp16, [1, 8, 64, 173]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59560320)))];
+            tensor<fp16, [1, 8, 17, 64]> q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1174_cast_fp16)[name = tensor<string, []>("transpose_201")];
+            tensor<fp16, [1, 8, 17, 173]> x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = var_1176_to_fp16)[name = tensor<string, []>("x_111_cast_fp16")];
+            tensor<int32, [8]> x_113_pad_0 = const()[name = tensor<string, []>("x_113_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_113_mode_0 = const()[name = tensor<string, []>("x_113_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_75_to_fp16 = const()[name = tensor<string, []>("const_75_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 17, 174]> x_113_cast_fp16 = pad(constant_val = const_75_to_fp16, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111_cast_fp16)[name = tensor<string, []>("x_113_cast_fp16")];
+            tensor<int32, [4]> var_1184 = const()[name = tensor<string, []>("op_1184"), val = tensor<int32, [4]>([1, 8, -1, 17])];
+            tensor<fp16, [1, 8, 174, 17]> x_115_cast_fp16 = reshape(shape = var_1184, x = x_113_cast_fp16)[name = tensor<string, []>("x_115_cast_fp16")];
+            tensor<int32, [4]> var_1188_begin_0 = const()[name = tensor<string, []>("op_1188_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1188_end_0 = const()[name = tensor<string, []>("op_1188_end_0"), val = tensor<int32, [4]>([1, 8, 174, 17])];
+            tensor<bool, [4]> var_1188_end_mask_0 = const()[name = tensor<string, []>("op_1188_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 173, 17]> var_1188_cast_fp16 = slice_by_index(begin = var_1188_begin_0, end = var_1188_end_0, end_mask = var_1188_end_mask_0, x = x_115_cast_fp16)[name = tensor<string, []>("op_1188_cast_fp16")];
+            tensor<int32, [4]> var_1189 = const()[name = tensor<string, []>("op_1189"), val = tensor<int32, [4]>([1, 8, 17, 173])];
+            tensor<fp16, [1, 8, 17, 173]> matrix_bd_17_cast_fp16 = reshape(shape = var_1189, x = var_1188_cast_fp16)[name = tensor<string, []>("matrix_bd_17_cast_fp16")];
+            tensor<bool, []> matrix_ac_9_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_9_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_9_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_9_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_59_perm_0 = const()[name = tensor<string, []>("transpose_59_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_60_perm_0 = const()[name = tensor<string, []>("transpose_60_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 87]> transpose_60 = transpose(perm = transpose_60_perm_0, x = k_17_cast_fp16)[name = tensor<string, []>("transpose_199")];
+            tensor<fp16, [1, 8, 17, 64]> transpose_59 = transpose(perm = transpose_59_perm_0, x = var_1172_cast_fp16)[name = tensor<string, []>("transpose_200")];
+            tensor<fp16, [1, 8, 17, 87]> matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_59, y = transpose_60)[name = tensor<string, []>("matrix_ac_9_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_19_begin_0 = const()[name = tensor<string, []>("matrix_bd_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_19_end_0 = const()[name = tensor<string, []>("matrix_bd_19_end_0"), val = tensor<int32, [4]>([1, 8, 17, 87])];
+            tensor<bool, [4]> matrix_bd_19_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 17, 87]> matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17_cast_fp16)[name = tensor<string, []>("matrix_bd_19_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_1198_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = tensor<string, []>("op_1198_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_17_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_17_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 17, 87]> _inversed_scores_17_cast_fp16 = mul(x = var_1198_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_17_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> scores_19_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_19_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_1204_cast_fp16 = softmax(axis = var_62, x = scores_19_cast_fp16)[name = tensor<string, []>("op_1204_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> input_249_cast_fp16 = select(a = var_40_to_fp16, b = var_1204_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_249_cast_fp16")];
+            tensor<bool, []> x_117_transpose_x_0 = const()[name = tensor<string, []>("x_117_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_117_transpose_y_0 = const()[name = tensor<string, []>("x_117_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 87, 64]> value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_9_cast_fp16)[name = tensor<string, []>("transpose_202")];
+            tensor<fp16, [1, 8, 17, 64]> x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249_cast_fp16, y = value_11_cast_fp16)[name = tensor<string, []>("x_117_cast_fp16")];
+            tensor<int32, [4]> var_1208_perm_0 = const()[name = tensor<string, []>("op_1208_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1209 = const()[name = tensor<string, []>("op_1209"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 17, 8, 64]> var_1208_cast_fp16 = transpose(perm = var_1208_perm_0, x = x_117_cast_fp16)[name = tensor<string, []>("transpose_198")];
+            tensor<fp16, [1, 17, 512]> input_251_cast_fp16 = reshape(shape = var_1209, x = var_1208_cast_fp16)[name = tensor<string, []>("input_251_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_4_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59737536)))];
+            tensor<fp16, [1, 17, 512]> linear_43_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16, x = input_251_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = tensor<string, []>("input_255_cast_fp16")];
+            tensor<int32, [1]> x_121_axes_0 = const()[name = tensor<string, []>("x_121_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60261888)))];
+            tensor<fp16, [512]> encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60262976)))];
+            tensor<fp16, [1, 17, 512]> x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = tensor<string, []>("x_121_cast_fp16")];
+            tensor<int32, [3]> input_257_perm_0 = const()[name = tensor<string, []>("input_257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_259_pad_type_0 = const()[name = tensor<string, []>("input_259_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_259_strides_0 = const()[name = tensor<string, []>("input_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_259_pad_0 = const()[name = tensor<string, []>("input_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_259_dilations_0 = const()[name = tensor<string, []>("input_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_259_groups_0 = const()[name = tensor<string, []>("input_259_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_4_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60264064)))];
+            tensor<fp16, [1, 512, 17]> input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = tensor<string, []>("transpose_197")];
+            tensor<fp16, [1, 1024, 17]> input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16, x = input_257_cast_fp16)[name = tensor<string, []>("input_259_cast_fp16")];
+            tensor<int32, []> x_123_split_num_splits_0 = const()[name = tensor<string, []>("x_123_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_123_split_axis_0 = const()[name = tensor<string, []>("x_123_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 17]> x_123_split_cast_fp16_0, tensor<fp16, [1, 512, 17]> x_123_split_cast_fp16_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259_cast_fp16)[name = tensor<string, []>("x_123_split_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = tensor<string, []>("x_123_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_123_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> input_261_cast_fp16 = select(a = var_40_to_fp16, b = x_123_cast_fp16, cond = var_418)[name = tensor<string, []>("input_261_cast_fp16")];
+            tensor<bool, []> new_x_19_interleave_0 = const()[name = tensor<string, []>("new_x_19_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 25]> new_x_19_cast_fp16 = concat(axis = var_62, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = tensor<string, []>("new_x_19_cast_fp16")];
+            tensor<int32, [3]> var_1247_begin_0 = const()[name = tensor<string, []>("op_1247_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
+            tensor<int32, [3]> var_1247_end_0 = const()[name = tensor<string, []>("op_1247_end_0"), val = tensor<int32, [3]>([1, 512, 25])];
+            tensor<bool, [3]> var_1247_end_mask_0 = const()[name = tensor<string, []>("op_1247_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_1247_cast_fp16 = slice_by_index(begin = var_1247_begin_0, end = var_1247_end_0, end_mask = var_1247_end_mask_0, x = new_x_19_cast_fp16)[name = tensor<string, []>("op_1247_cast_fp16")];
+            tensor<string, []> x_125_pad_type_0 = const()[name = tensor<string, []>("x_125_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_125_groups_0 = const()[name = tensor<string, []>("x_125_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_125_strides_0 = const()[name = tensor<string, []>("x_125_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_125_pad_0 = const()[name = tensor<string, []>("x_125_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_125_dilations_0 = const()[name = tensor<string, []>("x_125_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_4_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61312704)))];
+            tensor<fp16, [1, 512, 17]> x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16, x = new_x_19_cast_fp16)[name = tensor<string, []>("x_125_cast_fp16")];
+            tensor<int32, [3]> input_263_perm_0 = const()[name = tensor<string, []>("input_263_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_127_axes_0 = const()[name = tensor<string, []>("x_127_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61321984)))];
+            tensor<fp16, [512]> encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61323072)))];
+            tensor<fp16, [1, 17, 512]> input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = tensor<string, []>("transpose_196")];
+            tensor<fp16, [1, 17, 512]> x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = tensor<string, []>("x_127_cast_fp16")];
+            tensor<int32, [3]> input_265_perm_0 = const()[name = tensor<string, []>("input_265_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 17]> input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = tensor<string, []>("transpose_195")];
+            tensor<fp16, [1, 512, 17]> input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = tensor<string, []>("input_267_cast_fp16")];
+            tensor<string, []> x_129_pad_type_0 = const()[name = tensor<string, []>("x_129_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_129_strides_0 = const()[name = tensor<string, []>("x_129_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_129_pad_0 = const()[name = tensor<string, []>("x_129_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_129_dilations_0 = const()[name = tensor<string, []>("x_129_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_129_groups_0 = const()[name = tensor<string, []>("x_129_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_4_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61324160)))];
+            tensor<fp16, [1, 512, 17]> x_129_cast_fp16 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16, x = input_267_cast_fp16)[name = tensor<string, []>("x_129_cast_fp16")];
+            tensor<int32, [3]> input_269_perm_0 = const()[name = tensor<string, []>("input_269_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 17, 512]> input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = tensor<string, []>("transpose_194")];
+            tensor<fp16, [1, 17, 512]> input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = tensor<string, []>("input_271_cast_fp16")];
+            tensor<int32, [1]> input_273_axes_0 = const()[name = tensor<string, []>("input_273_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61848512)))];
+            tensor<fp16, [512]> encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61849600)))];
+            tensor<fp16, [1, 17, 512]> input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor<string, []>("input_273_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_4_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61850688)))];
+            tensor<fp16, [1, 17, 2048]> linear_44_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16, x = input_273_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = tensor<string, []>("input_277_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_4_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63947904)))];
+            tensor<fp16, [1, 17, 512]> linear_45_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor<string, []>("linear_45_cast_fp16")];
+            tensor<fp16, []> var_1288_to_fp16 = const()[name = tensor<string, []>("op_1288_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_1289_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1288_to_fp16)[name = tensor<string, []>("op_1289_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1289_cast_fp16)[name = tensor<string, []>("input_283_cast_fp16")];
+            tensor<int32, [1]> input_285_axes_0 = const()[name = tensor<string, []>("input_285_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_4_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66045120)))];
+            tensor<fp16, [512]> encoder_layers_4_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66046208)))];
+            tensor<fp16, [1, 17, 512]> input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = tensor<string, []>("input_285_cast_fp16")];
+            tensor<int32, [4]> cache_21_begin_0 = const()[name = tensor<string, []>("cache_21_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_21_end_0 = const()[name = tensor<string, []>("cache_21_end_0"), val = tensor<int32, [4]>([6, 1, 70, 512])];
+            tensor<bool, [4]> cache_21_end_mask_0 = const()[name = tensor<string, []>("cache_21_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_21_squeeze_mask_0 = const()[name = tensor<string, []>("cache_21_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_21_cast_fp16 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_21_cast_fp16")];
+            tensor<int32, [4]> cache_23_begin_0 = const()[name = tensor<string, []>("cache_23_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_23_end_0 = const()[name = tensor<string, []>("cache_23_end_0"), val = tensor<int32, [4]>([6, 1, 512, 8])];
+            tensor<bool, [4]> cache_23_end_mask_0 = const()[name = tensor<string, []>("cache_23_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_23_squeeze_mask_0 = const()[name = tensor<string, []>("cache_23_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_23_cast_fp16 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_23_cast_fp16")];
+            tensor<int32, [1]> input_287_axes_0 = const()[name = tensor<string, []>("input_287_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66047296)))];
+            tensor<fp16, [512]> encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66048384)))];
+            tensor<fp16, [1, 17, 512]> input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = tensor<string, []>("input_287_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_5_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66049472)))];
+            tensor<fp16, [1, 17, 2048]> linear_46_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor<string, []>("linear_46_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = tensor<string, []>("input_291_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_5_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68146688)))];
+            tensor<fp16, [1, 17, 512]> linear_47_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16, x = input_291_cast_fp16)[name = tensor<string, []>("linear_47_cast_fp16")];
+            tensor<fp16, []> var_1323_to_fp16 = const()[name = tensor<string, []>("op_1323_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_1324_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1323_to_fp16)[name = tensor<string, []>("op_1324_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1324_cast_fp16)[name = tensor<string, []>("input_297_cast_fp16")];
+            tensor<int32, [1]> key_11_axes_0 = const()[name = tensor<string, []>("key_11_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70243904)))];
+            tensor<fp16, [512]> encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70244992)))];
+            tensor<fp16, [1, 17, 512]> key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = tensor<string, []>("key_11_cast_fp16")];
+            tensor<bool, []> input_299_interleave_0 = const()[name = tensor<string, []>("input_299_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 87, 512]> input_299_cast_fp16 = concat(axis = var_64, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = tensor<string, []>("input_299_cast_fp16")];
+            tensor<int32, [3]> var_1346_begin_0 = const()[name = tensor<string, []>("op_1346_begin_0"), val = tensor<int32, [3]>([0, 17, 0])];
+            tensor<int32, [3]> var_1346_end_0 = const()[name = tensor<string, []>("op_1346_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_1346_end_mask_0 = const()[name = tensor<string, []>("op_1346_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 53, 512]> var_1346_cast_fp16 = slice_by_index(begin = var_1346_begin_0, end = var_1346_end_0, end_mask = var_1346_end_mask_0, x = cache_21_cast_fp16)[name = tensor<string, []>("op_1346_cast_fp16")];
+            tensor<bool, []> var_1352_interleave_0 = const()[name = tensor<string, []>("op_1352_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_1352_cast_fp16 = concat(axis = var_64, interleave = var_1352_interleave_0, values = (var_1346_cast_fp16, key_11_cast_fp16))[name = tensor<string, []>("op_1352_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_5_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70246080)))];
+            tensor<fp16, [1, 17, 512]> linear_48_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16, x = key_11_cast_fp16)[name = tensor<string, []>("linear_48_cast_fp16")];
+            tensor<int32, [4]> var_1356 = const()[name = tensor<string, []>("op_1356"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 17, 8, 64]> q_31_cast_fp16 = reshape(shape = var_1356, x = linear_48_cast_fp16)[name = tensor<string, []>("q_31_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_5_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70770432)))];
+            tensor<fp16, [1, 87, 512]> linear_49_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16, x = input_299_cast_fp16)[name = tensor<string, []>("linear_49_cast_fp16")];
+            tensor<int32, [4]> var_1360 = const()[name = tensor<string, []>("op_1360"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> k_21_cast_fp16 = reshape(shape = var_1360, x = linear_49_cast_fp16)[name = tensor<string, []>("k_21_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_5_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71294784)))];
+            tensor<fp16, [1, 87, 512]> linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16, x = input_299_cast_fp16)[name = tensor<string, []>("linear_50_cast_fp16")];
+            tensor<int32, [4]> var_1364 = const()[name = tensor<string, []>("op_1364"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> v_11_cast_fp16 = reshape(shape = var_1364, x = linear_50_cast_fp16)[name = tensor<string, []>("v_11_cast_fp16")];
+            tensor<int32, [4]> value_13_perm_0 = const()[name = tensor<string, []>("value_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71819136)))];
+            tensor<fp16, [1, 17, 8, 64]> var_1376_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1376_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71820224)))];
+            tensor<fp16, [1, 17, 8, 64]> var_1378_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1378_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_11_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_137_transpose_x_0 = const()[name = tensor<string, []>("x_137_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_137_transpose_y_0 = const()[name = tensor<string, []>("x_137_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 173]> var_1380_to_fp16 = const()[name = tensor<string, []>("op_1380_to_fp16"), val = tensor<fp16, [1, 8, 64, 173]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71821312)))];
+            tensor<fp16, [1, 8, 17, 64]> q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1378_cast_fp16)[name = tensor<string, []>("transpose_192")];
+            tensor<fp16, [1, 8, 17, 173]> x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = var_1380_to_fp16)[name = tensor<string, []>("x_137_cast_fp16")];
+            tensor<int32, [8]> x_139_pad_0 = const()[name = tensor<string, []>("x_139_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_139_mode_0 = const()[name = tensor<string, []>("x_139_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_88_to_fp16 = const()[name = tensor<string, []>("const_88_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 17, 174]> x_139_cast_fp16 = pad(constant_val = const_88_to_fp16, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137_cast_fp16)[name = tensor<string, []>("x_139_cast_fp16")];
+            tensor<int32, [4]> var_1388 = const()[name = tensor<string, []>("op_1388"), val = tensor<int32, [4]>([1, 8, -1, 17])];
+            tensor<fp16, [1, 8, 174, 17]> x_141_cast_fp16 = reshape(shape = var_1388, x = x_139_cast_fp16)[name = tensor<string, []>("x_141_cast_fp16")];
+            tensor<int32, [4]> var_1392_begin_0 = const()[name = tensor<string, []>("op_1392_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1392_end_0 = const()[name = tensor<string, []>("op_1392_end_0"), val = tensor<int32, [4]>([1, 8, 174, 17])];
+            tensor<bool, [4]> var_1392_end_mask_0 = const()[name = tensor<string, []>("op_1392_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 173, 17]> var_1392_cast_fp16 = slice_by_index(begin = var_1392_begin_0, end = var_1392_end_0, end_mask = var_1392_end_mask_0, x = x_141_cast_fp16)[name = tensor<string, []>("op_1392_cast_fp16")];
+            tensor<int32, [4]> var_1393 = const()[name = tensor<string, []>("op_1393"), val = tensor<int32, [4]>([1, 8, 17, 173])];
+            tensor<fp16, [1, 8, 17, 173]> matrix_bd_21_cast_fp16 = reshape(shape = var_1393, x = var_1392_cast_fp16)[name = tensor<string, []>("matrix_bd_21_cast_fp16")];
+            tensor<bool, []> matrix_ac_11_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_11_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_11_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_11_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_61_perm_0 = const()[name = tensor<string, []>("transpose_61_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_62_perm_0 = const()[name = tensor<string, []>("transpose_62_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 87]> transpose_62 = transpose(perm = transpose_62_perm_0, x = k_21_cast_fp16)[name = tensor<string, []>("transpose_190")];
+            tensor<fp16, [1, 8, 17, 64]> transpose_61 = transpose(perm = transpose_61_perm_0, x = var_1376_cast_fp16)[name = tensor<string, []>("transpose_191")];
+            tensor<fp16, [1, 8, 17, 87]> matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_61, y = transpose_62)[name = tensor<string, []>("matrix_ac_11_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_23_begin_0 = const()[name = tensor<string, []>("matrix_bd_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_23_end_0 = const()[name = tensor<string, []>("matrix_bd_23_end_0"), val = tensor<int32, [4]>([1, 8, 17, 87])];
+            tensor<bool, [4]> matrix_bd_23_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 17, 87]> matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21_cast_fp16)[name = tensor<string, []>("matrix_bd_23_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_1402_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = tensor<string, []>("op_1402_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_21_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_21_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 17, 87]> _inversed_scores_21_cast_fp16 = mul(x = var_1402_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_21_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> scores_23_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_23_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_1408_cast_fp16 = softmax(axis = var_62, x = scores_23_cast_fp16)[name = tensor<string, []>("op_1408_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> input_301_cast_fp16 = select(a = var_40_to_fp16, b = var_1408_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_301_cast_fp16")];
+            tensor<bool, []> x_143_transpose_x_0 = const()[name = tensor<string, []>("x_143_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_143_transpose_y_0 = const()[name = tensor<string, []>("x_143_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 87, 64]> value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_11_cast_fp16)[name = tensor<string, []>("transpose_193")];
+            tensor<fp16, [1, 8, 17, 64]> x_143_cast_fp16 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301_cast_fp16, y = value_13_cast_fp16)[name = tensor<string, []>("x_143_cast_fp16")];
+            tensor<int32, [4]> var_1412_perm_0 = const()[name = tensor<string, []>("op_1412_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1413 = const()[name = tensor<string, []>("op_1413"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 17, 8, 64]> var_1412_cast_fp16 = transpose(perm = var_1412_perm_0, x = x_143_cast_fp16)[name = tensor<string, []>("transpose_189")];
+            tensor<fp16, [1, 17, 512]> input_303_cast_fp16 = reshape(shape = var_1413, x = var_1412_cast_fp16)[name = tensor<string, []>("input_303_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_5_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71998528)))];
+            tensor<fp16, [1, 17, 512]> linear_52_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16, x = input_303_cast_fp16)[name = tensor<string, []>("linear_52_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = tensor<string, []>("input_307_cast_fp16")];
+            tensor<int32, [1]> x_147_axes_0 = const()[name = tensor<string, []>("x_147_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72522880)))];
+            tensor<fp16, [512]> encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72523968)))];
+            tensor<fp16, [1, 17, 512]> x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = tensor<string, []>("x_147_cast_fp16")];
+            tensor<int32, [3]> input_309_perm_0 = const()[name = tensor<string, []>("input_309_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_311_pad_type_0 = const()[name = tensor<string, []>("input_311_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_311_strides_0 = const()[name = tensor<string, []>("input_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_311_pad_0 = const()[name = tensor<string, []>("input_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_311_dilations_0 = const()[name = tensor<string, []>("input_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_311_groups_0 = const()[name = tensor<string, []>("input_311_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_5_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72525056)))];
+            tensor<fp16, [1, 512, 17]> input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = tensor<string, []>("transpose_188")];
+            tensor<fp16, [1, 1024, 17]> input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16, x = input_309_cast_fp16)[name = tensor<string, []>("input_311_cast_fp16")];
+            tensor<int32, []> x_149_split_num_splits_0 = const()[name = tensor<string, []>("x_149_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_149_split_axis_0 = const()[name = tensor<string, []>("x_149_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 17]> x_149_split_cast_fp16_0, tensor<fp16, [1, 512, 17]> x_149_split_cast_fp16_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311_cast_fp16)[name = tensor<string, []>("x_149_split_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = tensor<string, []>("x_149_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_149_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> input_313_cast_fp16 = select(a = var_40_to_fp16, b = x_149_cast_fp16, cond = var_418)[name = tensor<string, []>("input_313_cast_fp16")];
+            tensor<bool, []> new_x_23_interleave_0 = const()[name = tensor<string, []>("new_x_23_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 25]> new_x_23_cast_fp16 = concat(axis = var_62, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = tensor<string, []>("new_x_23_cast_fp16")];
+            tensor<int32, [3]> var_1451_begin_0 = const()[name = tensor<string, []>("op_1451_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
+            tensor<int32, [3]> var_1451_end_0 = const()[name = tensor<string, []>("op_1451_end_0"), val = tensor<int32, [3]>([1, 512, 25])];
+            tensor<bool, [3]> var_1451_end_mask_0 = const()[name = tensor<string, []>("op_1451_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_1451_cast_fp16 = slice_by_index(begin = var_1451_begin_0, end = var_1451_end_0, end_mask = var_1451_end_mask_0, x = new_x_23_cast_fp16)[name = tensor<string, []>("op_1451_cast_fp16")];
+            tensor<string, []> x_151_pad_type_0 = const()[name = tensor<string, []>("x_151_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_151_groups_0 = const()[name = tensor<string, []>("x_151_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_151_strides_0 = const()[name = tensor<string, []>("x_151_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_151_pad_0 = const()[name = tensor<string, []>("x_151_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_151_dilations_0 = const()[name = tensor<string, []>("x_151_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_5_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73573696)))];
+            tensor<fp16, [1, 512, 17]> x_151_cast_fp16 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16, x = new_x_23_cast_fp16)[name = tensor<string, []>("x_151_cast_fp16")];
+            tensor<int32, [3]> input_315_perm_0 = const()[name = tensor<string, []>("input_315_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_153_axes_0 = const()[name = tensor<string, []>("x_153_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73582976)))];
+            tensor<fp16, [512]> encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73584064)))];
+            tensor<fp16, [1, 17, 512]> input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = tensor<string, []>("transpose_187")];
+            tensor<fp16, [1, 17, 512]> x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = tensor<string, []>("x_153_cast_fp16")];
+            tensor<int32, [3]> input_317_perm_0 = const()[name = tensor<string, []>("input_317_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 17]> input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = tensor<string, []>("transpose_186")];
+            tensor<fp16, [1, 512, 17]> input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = tensor<string, []>("input_319_cast_fp16")];
+            tensor<string, []> x_155_pad_type_0 = const()[name = tensor<string, []>("x_155_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_155_strides_0 = const()[name = tensor<string, []>("x_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_155_pad_0 = const()[name = tensor<string, []>("x_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_155_dilations_0 = const()[name = tensor<string, []>("x_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_155_groups_0 = const()[name = tensor<string, []>("x_155_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_5_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73585152)))];
+            tensor<fp16, [1, 512, 17]> x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16, x = input_319_cast_fp16)[name = tensor<string, []>("x_155_cast_fp16")];
+            tensor<int32, [3]> input_321_perm_0 = const()[name = tensor<string, []>("input_321_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 17, 512]> input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = tensor<string, []>("transpose_185")];
+            tensor<fp16, [1, 17, 512]> input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = tensor<string, []>("input_323_cast_fp16")];
+            tensor<int32, [1]> input_325_axes_0 = const()[name = tensor<string, []>("input_325_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74109504)))];
+            tensor<fp16, [512]> encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74110592)))];
+            tensor<fp16, [1, 17, 512]> input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = tensor<string, []>("input_325_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_5_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74111680)))];
+            tensor<fp16, [1, 17, 2048]> linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16, x = input_325_cast_fp16)[name = tensor<string, []>("linear_53_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = tensor<string, []>("input_329_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_5_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76208896)))];
+            tensor<fp16, [1, 17, 512]> linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16, x = input_329_cast_fp16)[name = tensor<string, []>("linear_54_cast_fp16")];
+            tensor<fp16, []> var_1492_to_fp16 = const()[name = tensor<string, []>("op_1492_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_1493_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1492_to_fp16)[name = tensor<string, []>("op_1493_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1493_cast_fp16)[name = tensor<string, []>("input_335_cast_fp16")];
+            tensor<int32, [1]> input_337_axes_0 = const()[name = tensor<string, []>("input_337_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_5_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78306112)))];
+            tensor<fp16, [512]> encoder_layers_5_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78307200)))];
+            tensor<fp16, [1, 17, 512]> input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = tensor<string, []>("input_337_cast_fp16")];
+            tensor<int32, [4]> cache_25_begin_0 = const()[name = tensor<string, []>("cache_25_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_25_end_0 = const()[name = tensor<string, []>("cache_25_end_0"), val = tensor<int32, [4]>([7, 1, 70, 512])];
+            tensor<bool, [4]> cache_25_end_mask_0 = const()[name = tensor<string, []>("cache_25_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_25_squeeze_mask_0 = const()[name = tensor<string, []>("cache_25_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_25_cast_fp16 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_25_cast_fp16")];
+            tensor<int32, [4]> cache_27_begin_0 = const()[name = tensor<string, []>("cache_27_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_27_end_0 = const()[name = tensor<string, []>("cache_27_end_0"), val = tensor<int32, [4]>([7, 1, 512, 8])];
+            tensor<bool, [4]> cache_27_end_mask_0 = const()[name = tensor<string, []>("cache_27_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_27_squeeze_mask_0 = const()[name = tensor<string, []>("cache_27_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_27_cast_fp16 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_27_cast_fp16")];
+            tensor<int32, [1]> input_339_axes_0 = const()[name = tensor<string, []>("input_339_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78308288)))];
+            tensor<fp16, [512]> encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78309376)))];
+            tensor<fp16, [1, 17, 512]> input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = tensor<string, []>("input_339_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_6_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78310464)))];
+            tensor<fp16, [1, 17, 2048]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16, x = input_339_cast_fp16)[name = tensor<string, []>("linear_55_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = tensor<string, []>("input_343_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_6_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80407680)))];
+            tensor<fp16, [1, 17, 512]> linear_56_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16, x = input_343_cast_fp16)[name = tensor<string, []>("linear_56_cast_fp16")];
+            tensor<fp16, []> var_1527_to_fp16 = const()[name = tensor<string, []>("op_1527_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_1528_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1527_to_fp16)[name = tensor<string, []>("op_1528_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1528_cast_fp16)[name = tensor<string, []>("input_349_cast_fp16")];
+            tensor<int32, [1]> key_13_axes_0 = const()[name = tensor<string, []>("key_13_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82504896)))];
+            tensor<fp16, [512]> encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82505984)))];
+            tensor<fp16, [1, 17, 512]> key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")];
+            tensor<bool, []> input_351_interleave_0 = const()[name = tensor<string, []>("input_351_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 87, 512]> input_351_cast_fp16 = concat(axis = var_64, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = tensor<string, []>("input_351_cast_fp16")];
+            tensor<int32, [3]> var_1550_begin_0 = const()[name = tensor<string, []>("op_1550_begin_0"), val = tensor<int32, [3]>([0, 17, 0])];
+            tensor<int32, [3]> var_1550_end_0 = const()[name = tensor<string, []>("op_1550_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_1550_end_mask_0 = const()[name = tensor<string, []>("op_1550_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 53, 512]> var_1550_cast_fp16 = slice_by_index(begin = var_1550_begin_0, end = var_1550_end_0, end_mask = var_1550_end_mask_0, x = cache_25_cast_fp16)[name = tensor<string, []>("op_1550_cast_fp16")];
+            tensor<bool, []> var_1556_interleave_0 = const()[name = tensor<string, []>("op_1556_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_1556_cast_fp16 = concat(axis = var_64, interleave = var_1556_interleave_0, values = (var_1550_cast_fp16, key_13_cast_fp16))[name = tensor<string, []>("op_1556_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_6_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82507072)))];
+            tensor<fp16, [1, 17, 512]> linear_57_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16, x = key_13_cast_fp16)[name = tensor<string, []>("linear_57_cast_fp16")];
+            tensor<int32, [4]> var_1560 = const()[name = tensor<string, []>("op_1560"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 17, 8, 64]> q_37_cast_fp16 = reshape(shape = var_1560, x = linear_57_cast_fp16)[name = tensor<string, []>("q_37_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_6_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83031424)))];
+            tensor<fp16, [1, 87, 512]> linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16, x = input_351_cast_fp16)[name = tensor<string, []>("linear_58_cast_fp16")];
+            tensor<int32, [4]> var_1564 = const()[name = tensor<string, []>("op_1564"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> k_25_cast_fp16 = reshape(shape = var_1564, x = linear_58_cast_fp16)[name = tensor<string, []>("k_25_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_6_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83555776)))];
+            tensor<fp16, [1, 87, 512]> linear_59_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16, x = input_351_cast_fp16)[name = tensor<string, []>("linear_59_cast_fp16")];
+            tensor<int32, [4]> var_1568 = const()[name = tensor<string, []>("op_1568"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> v_13_cast_fp16 = reshape(shape = var_1568, x = linear_59_cast_fp16)[name = tensor<string, []>("v_13_cast_fp16")];
+            tensor<int32, [4]> value_15_perm_0 = const()[name = tensor<string, []>("value_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84080128)))];
+            tensor<fp16, [1, 17, 8, 64]> var_1580_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1580_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84081216)))];
+            tensor<fp16, [1, 17, 8, 64]> var_1582_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1582_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_13_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_163_transpose_x_0 = const()[name = tensor<string, []>("x_163_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_163_transpose_y_0 = const()[name = tensor<string, []>("x_163_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 173]> var_1584_to_fp16 = const()[name = tensor<string, []>("op_1584_to_fp16"), val = tensor<fp16, [1, 8, 64, 173]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84082304)))];
+            tensor<fp16, [1, 8, 17, 64]> q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1582_cast_fp16)[name = tensor<string, []>("transpose_183")];
+            tensor<fp16, [1, 8, 17, 173]> x_163_cast_fp16 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = var_1584_to_fp16)[name = tensor<string, []>("x_163_cast_fp16")];
+            tensor<int32, [8]> x_165_pad_0 = const()[name = tensor<string, []>("x_165_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_165_mode_0 = const()[name = tensor<string, []>("x_165_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_101_to_fp16 = const()[name = tensor<string, []>("const_101_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 17, 174]> x_165_cast_fp16 = pad(constant_val = const_101_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = tensor<string, []>("x_165_cast_fp16")];
+            tensor<int32, [4]> var_1592 = const()[name = tensor<string, []>("op_1592"), val = tensor<int32, [4]>([1, 8, -1, 17])];
+            tensor<fp16, [1, 8, 174, 17]> x_167_cast_fp16 = reshape(shape = var_1592, x = x_165_cast_fp16)[name = tensor<string, []>("x_167_cast_fp16")];
+            tensor<int32, [4]> var_1596_begin_0 = const()[name = tensor<string, []>("op_1596_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1596_end_0 = const()[name = tensor<string, []>("op_1596_end_0"), val = tensor<int32, [4]>([1, 8, 174, 17])];
+            tensor<bool, [4]> var_1596_end_mask_0 = const()[name = tensor<string, []>("op_1596_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 173, 17]> var_1596_cast_fp16 = slice_by_index(begin = var_1596_begin_0, end = var_1596_end_0, end_mask = var_1596_end_mask_0, x = x_167_cast_fp16)[name = tensor<string, []>("op_1596_cast_fp16")];
+            tensor<int32, [4]> var_1597 = const()[name = tensor<string, []>("op_1597"), val = tensor<int32, [4]>([1, 8, 17, 173])];
+            tensor<fp16, [1, 8, 17, 173]> matrix_bd_25_cast_fp16 = reshape(shape = var_1597, x = var_1596_cast_fp16)[name = tensor<string, []>("matrix_bd_25_cast_fp16")];
+            tensor<bool, []> matrix_ac_13_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_13_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_63_perm_0 = const()[name = tensor<string, []>("transpose_63_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_64_perm_0 = const()[name = tensor<string, []>("transpose_64_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 87]> transpose_64 = transpose(perm = transpose_64_perm_0, x = k_25_cast_fp16)[name = tensor<string, []>("transpose_181")];
+            tensor<fp16, [1, 8, 17, 64]> transpose_63 = transpose(perm = transpose_63_perm_0, x = var_1580_cast_fp16)[name = tensor<string, []>("transpose_182")];
+            tensor<fp16, [1, 8, 17, 87]> matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_63, y = transpose_64)[name = tensor<string, []>("matrix_ac_13_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_27_begin_0 = const()[name = tensor<string, []>("matrix_bd_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_27_end_0 = const()[name = tensor<string, []>("matrix_bd_27_end_0"), val = tensor<int32, [4]>([1, 8, 17, 87])];
+            tensor<bool, [4]> matrix_bd_27_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 17, 87]> matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25_cast_fp16)[name = tensor<string, []>("matrix_bd_27_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_1606_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = tensor<string, []>("op_1606_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_25_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_25_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 17, 87]> _inversed_scores_25_cast_fp16 = mul(x = var_1606_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_25_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> scores_27_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_27_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_1612_cast_fp16 = softmax(axis = var_62, x = scores_27_cast_fp16)[name = tensor<string, []>("op_1612_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> input_353_cast_fp16 = select(a = var_40_to_fp16, b = var_1612_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_353_cast_fp16")];
+            tensor<bool, []> x_169_transpose_x_0 = const()[name = tensor<string, []>("x_169_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_169_transpose_y_0 = const()[name = tensor<string, []>("x_169_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 87, 64]> value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_13_cast_fp16)[name = tensor<string, []>("transpose_184")];
+            tensor<fp16, [1, 8, 17, 64]> x_169_cast_fp16 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353_cast_fp16, y = value_15_cast_fp16)[name = tensor<string, []>("x_169_cast_fp16")];
+            tensor<int32, [4]> var_1616_perm_0 = const()[name = tensor<string, []>("op_1616_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1617 = const()[name = tensor<string, []>("op_1617"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 17, 8, 64]> var_1616_cast_fp16 = transpose(perm = var_1616_perm_0, x = x_169_cast_fp16)[name = tensor<string, []>("transpose_180")];
+            tensor<fp16, [1, 17, 512]> input_355_cast_fp16 = reshape(shape = var_1617, x = var_1616_cast_fp16)[name = tensor<string, []>("input_355_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_6_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84259520)))];
+            tensor<fp16, [1, 17, 512]> linear_61_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16, x = input_355_cast_fp16)[name = tensor<string, []>("linear_61_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = tensor<string, []>("input_359_cast_fp16")];
+            tensor<int32, [1]> x_173_axes_0 = const()[name = tensor<string, []>("x_173_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84783872)))];
+            tensor<fp16, [512]> encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84784960)))];
+            tensor<fp16, [1, 17, 512]> x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = tensor<string, []>("x_173_cast_fp16")];
+            tensor<int32, [3]> input_361_perm_0 = const()[name = tensor<string, []>("input_361_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_363_pad_type_0 = const()[name = tensor<string, []>("input_363_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_363_strides_0 = const()[name = tensor<string, []>("input_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_363_pad_0 = const()[name = tensor<string, []>("input_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_363_dilations_0 = const()[name = tensor<string, []>("input_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_363_groups_0 = const()[name = tensor<string, []>("input_363_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_6_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84786048)))];
+            tensor<fp16, [1, 512, 17]> input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = tensor<string, []>("transpose_179")];
+            tensor<fp16, [1, 1024, 17]> input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16, x = input_361_cast_fp16)[name = tensor<string, []>("input_363_cast_fp16")];
+            tensor<int32, []> x_175_split_num_splits_0 = const()[name = tensor<string, []>("x_175_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_175_split_axis_0 = const()[name = tensor<string, []>("x_175_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 17]> x_175_split_cast_fp16_0, tensor<fp16, [1, 512, 17]> x_175_split_cast_fp16_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363_cast_fp16)[name = tensor<string, []>("x_175_split_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = tensor<string, []>("x_175_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_175_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> input_365_cast_fp16 = select(a = var_40_to_fp16, b = x_175_cast_fp16, cond = var_418)[name = tensor<string, []>("input_365_cast_fp16")];
+            tensor<bool, []> new_x_27_interleave_0 = const()[name = tensor<string, []>("new_x_27_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 25]> new_x_27_cast_fp16 = concat(axis = var_62, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = tensor<string, []>("new_x_27_cast_fp16")];
+            tensor<int32, [3]> var_1655_begin_0 = const()[name = tensor<string, []>("op_1655_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
+            tensor<int32, [3]> var_1655_end_0 = const()[name = tensor<string, []>("op_1655_end_0"), val = tensor<int32, [3]>([1, 512, 25])];
+            tensor<bool, [3]> var_1655_end_mask_0 = const()[name = tensor<string, []>("op_1655_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_1655_cast_fp16 = slice_by_index(begin = var_1655_begin_0, end = var_1655_end_0, end_mask = var_1655_end_mask_0, x = new_x_27_cast_fp16)[name = tensor<string, []>("op_1655_cast_fp16")];
+            tensor<string, []> x_177_pad_type_0 = const()[name = tensor<string, []>("x_177_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_177_groups_0 = const()[name = tensor<string, []>("x_177_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_177_strides_0 = const()[name = tensor<string, []>("x_177_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_177_pad_0 = const()[name = tensor<string, []>("x_177_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_177_dilations_0 = const()[name = tensor<string, []>("x_177_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_6_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85834688)))];
+            tensor<fp16, [1, 512, 17]> x_177_cast_fp16 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16, x = new_x_27_cast_fp16)[name = tensor<string, []>("x_177_cast_fp16")];
+            tensor<int32, [3]> input_367_perm_0 = const()[name = tensor<string, []>("input_367_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_179_axes_0 = const()[name = tensor<string, []>("x_179_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85843968)))];
+            tensor<fp16, [512]> encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85845056)))];
+            tensor<fp16, [1, 17, 512]> input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = tensor<string, []>("transpose_178")];
+            tensor<fp16, [1, 17, 512]> x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = tensor<string, []>("x_179_cast_fp16")];
+            tensor<int32, [3]> input_369_perm_0 = const()[name = tensor<string, []>("input_369_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 17]> input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = tensor<string, []>("transpose_177")];
+            tensor<fp16, [1, 512, 17]> input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = tensor<string, []>("input_371_cast_fp16")];
+            tensor<string, []> x_181_pad_type_0 = const()[name = tensor<string, []>("x_181_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_181_strides_0 = const()[name = tensor<string, []>("x_181_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_181_pad_0 = const()[name = tensor<string, []>("x_181_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_181_dilations_0 = const()[name = tensor<string, []>("x_181_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_181_groups_0 = const()[name = tensor<string, []>("x_181_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_6_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85846144)))];
+            tensor<fp16, [1, 512, 17]> x_181_cast_fp16 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16, x = input_371_cast_fp16)[name = tensor<string, []>("x_181_cast_fp16")];
+            tensor<int32, [3]> input_373_perm_0 = const()[name = tensor<string, []>("input_373_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 17, 512]> input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = tensor<string, []>("transpose_176")];
+            tensor<fp16, [1, 17, 512]> input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = tensor<string, []>("input_375_cast_fp16")];
+            tensor<int32, [1]> input_377_axes_0 = const()[name = tensor<string, []>("input_377_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86370496)))];
+            tensor<fp16, [512]> encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86371584)))];
+            tensor<fp16, [1, 17, 512]> input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = tensor<string, []>("input_377_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_6_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86372672)))];
+            tensor<fp16, [1, 17, 2048]> linear_62_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16, x = input_377_cast_fp16)[name = tensor<string, []>("linear_62_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = tensor<string, []>("input_381_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_6_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88469888)))];
+            tensor<fp16, [1, 17, 512]> linear_63_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16, x = input_381_cast_fp16)[name = tensor<string, []>("linear_63_cast_fp16")];
+            tensor<fp16, []> var_1696_to_fp16 = const()[name = tensor<string, []>("op_1696_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_1697_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1696_to_fp16)[name = tensor<string, []>("op_1697_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1697_cast_fp16)[name = tensor<string, []>("input_387_cast_fp16")];
+            tensor<int32, [1]> input_389_axes_0 = const()[name = tensor<string, []>("input_389_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_6_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90567104)))];
+            tensor<fp16, [512]> encoder_layers_6_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90568192)))];
+            tensor<fp16, [1, 17, 512]> input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = tensor<string, []>("input_389_cast_fp16")];
+            tensor<int32, [4]> cache_29_begin_0 = const()[name = tensor<string, []>("cache_29_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_29_end_0 = const()[name = tensor<string, []>("cache_29_end_0"), val = tensor<int32, [4]>([8, 1, 70, 512])];
+            tensor<bool, [4]> cache_29_end_mask_0 = const()[name = tensor<string, []>("cache_29_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_29_squeeze_mask_0 = const()[name = tensor<string, []>("cache_29_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_29_cast_fp16 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_29_cast_fp16")];
+            tensor<int32, [4]> cache_31_begin_0 = const()[name = tensor<string, []>("cache_31_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_31_end_0 = const()[name = tensor<string, []>("cache_31_end_0"), val = tensor<int32, [4]>([8, 1, 512, 8])];
+            tensor<bool, [4]> cache_31_end_mask_0 = const()[name = tensor<string, []>("cache_31_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_31_squeeze_mask_0 = const()[name = tensor<string, []>("cache_31_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_31_cast_fp16 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_31_cast_fp16")];
+            tensor<int32, [1]> input_391_axes_0 = const()[name = tensor<string, []>("input_391_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90569280)))];
+            tensor<fp16, [512]> encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90570368)))];
+            tensor<fp16, [1, 17, 512]> input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = tensor<string, []>("input_391_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_7_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90571456)))];
+            tensor<fp16, [1, 17, 2048]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16, x = input_391_cast_fp16)[name = tensor<string, []>("linear_64_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = tensor<string, []>("input_395_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_7_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92668672)))];
+            tensor<fp16, [1, 17, 512]> linear_65_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16, x = input_395_cast_fp16)[name = tensor<string, []>("linear_65_cast_fp16")];
+            tensor<fp16, []> var_1731_to_fp16 = const()[name = tensor<string, []>("op_1731_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_1732_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1731_to_fp16)[name = tensor<string, []>("op_1732_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1732_cast_fp16)[name = tensor<string, []>("input_401_cast_fp16")];
+            tensor<int32, [1]> key_15_axes_0 = const()[name = tensor<string, []>("key_15_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94765888)))];
+            tensor<fp16, [512]> encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94766976)))];
+            tensor<fp16, [1, 17, 512]> key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = tensor<string, []>("key_15_cast_fp16")];
+            tensor<bool, []> input_403_interleave_0 = const()[name = tensor<string, []>("input_403_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 87, 512]> input_403_cast_fp16 = concat(axis = var_64, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = tensor<string, []>("input_403_cast_fp16")];
+            tensor<int32, [3]> var_1754_begin_0 = const()[name = tensor<string, []>("op_1754_begin_0"), val = tensor<int32, [3]>([0, 17, 0])];
+            tensor<int32, [3]> var_1754_end_0 = const()[name = tensor<string, []>("op_1754_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_1754_end_mask_0 = const()[name = tensor<string, []>("op_1754_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 53, 512]> var_1754_cast_fp16 = slice_by_index(begin = var_1754_begin_0, end = var_1754_end_0, end_mask = var_1754_end_mask_0, x = cache_29_cast_fp16)[name = tensor<string, []>("op_1754_cast_fp16")];
+            tensor<bool, []> var_1760_interleave_0 = const()[name = tensor<string, []>("op_1760_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_1760_cast_fp16 = concat(axis = var_64, interleave = var_1760_interleave_0, values = (var_1754_cast_fp16, key_15_cast_fp16))[name = tensor<string, []>("op_1760_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_7_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94768064)))];
+            tensor<fp16, [1, 17, 512]> linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16, x = key_15_cast_fp16)[name = tensor<string, []>("linear_66_cast_fp16")];
+            tensor<int32, [4]> var_1764 = const()[name = tensor<string, []>("op_1764"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 17, 8, 64]> q_43_cast_fp16 = reshape(shape = var_1764, x = linear_66_cast_fp16)[name = tensor<string, []>("q_43_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_7_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(95292416)))];
+            tensor<fp16, [1, 87, 512]> linear_67_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16, x = input_403_cast_fp16)[name = tensor<string, []>("linear_67_cast_fp16")];
+            tensor<int32, [4]> var_1768 = const()[name = tensor<string, []>("op_1768"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> k_29_cast_fp16 = reshape(shape = var_1768, x = linear_67_cast_fp16)[name = tensor<string, []>("k_29_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_7_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(95816768)))];
+            tensor<fp16, [1, 87, 512]> linear_68_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16, x = input_403_cast_fp16)[name = tensor<string, []>("linear_68_cast_fp16")];
+            tensor<int32, [4]> var_1772 = const()[name = tensor<string, []>("op_1772"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> v_15_cast_fp16 = reshape(shape = var_1772, x = linear_68_cast_fp16)[name = tensor<string, []>("v_15_cast_fp16")];
+            tensor<int32, [4]> value_17_perm_0 = const()[name = tensor<string, []>("value_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96341120)))];
+            tensor<fp16, [1, 17, 8, 64]> var_1784_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1784_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96342208)))];
+            tensor<fp16, [1, 17, 8, 64]> var_1786_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1786_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_15_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_189_transpose_x_0 = const()[name = tensor<string, []>("x_189_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_189_transpose_y_0 = const()[name = tensor<string, []>("x_189_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 173]> var_1788_to_fp16 = const()[name = tensor<string, []>("op_1788_to_fp16"), val = tensor<fp16, [1, 8, 64, 173]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96343296)))];
+            tensor<fp16, [1, 8, 17, 64]> q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_1786_cast_fp16)[name = tensor<string, []>("transpose_174")];
+            tensor<fp16, [1, 8, 17, 173]> x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = var_1788_to_fp16)[name = tensor<string, []>("x_189_cast_fp16")];
+            tensor<int32, [8]> x_191_pad_0 = const()[name = tensor<string, []>("x_191_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_191_mode_0 = const()[name = tensor<string, []>("x_191_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_114_to_fp16 = const()[name = tensor<string, []>("const_114_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 17, 174]> x_191_cast_fp16 = pad(constant_val = const_114_to_fp16, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189_cast_fp16)[name = tensor<string, []>("x_191_cast_fp16")];
+            tensor<int32, [4]> var_1796 = const()[name = tensor<string, []>("op_1796"), val = tensor<int32, [4]>([1, 8, -1, 17])];
+            tensor<fp16, [1, 8, 174, 17]> x_193_cast_fp16 = reshape(shape = var_1796, x = x_191_cast_fp16)[name = tensor<string, []>("x_193_cast_fp16")];
+            tensor<int32, [4]> var_1800_begin_0 = const()[name = tensor<string, []>("op_1800_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1800_end_0 = const()[name = tensor<string, []>("op_1800_end_0"), val = tensor<int32, [4]>([1, 8, 174, 17])];
+            tensor<bool, [4]> var_1800_end_mask_0 = const()[name = tensor<string, []>("op_1800_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 173, 17]> var_1800_cast_fp16 = slice_by_index(begin = var_1800_begin_0, end = var_1800_end_0, end_mask = var_1800_end_mask_0, x = x_193_cast_fp16)[name = tensor<string, []>("op_1800_cast_fp16")];
+            tensor<int32, [4]> var_1801 = const()[name = tensor<string, []>("op_1801"), val = tensor<int32, [4]>([1, 8, 17, 173])];
+            tensor<fp16, [1, 8, 17, 173]> matrix_bd_29_cast_fp16 = reshape(shape = var_1801, x = var_1800_cast_fp16)[name = tensor<string, []>("matrix_bd_29_cast_fp16")];
+            tensor<bool, []> matrix_ac_15_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_15_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_15_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_15_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_65_perm_0 = const()[name = tensor<string, []>("transpose_65_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_66_perm_0 = const()[name = tensor<string, []>("transpose_66_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 87]> transpose_66 = transpose(perm = transpose_66_perm_0, x = k_29_cast_fp16)[name = tensor<string, []>("transpose_172")];
+            tensor<fp16, [1, 8, 17, 64]> transpose_65 = transpose(perm = transpose_65_perm_0, x = var_1784_cast_fp16)[name = tensor<string, []>("transpose_173")];
+            tensor<fp16, [1, 8, 17, 87]> matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_65, y = transpose_66)[name = tensor<string, []>("matrix_ac_15_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_31_begin_0 = const()[name = tensor<string, []>("matrix_bd_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_31_end_0 = const()[name = tensor<string, []>("matrix_bd_31_end_0"), val = tensor<int32, [4]>([1, 8, 17, 87])];
+            tensor<bool, [4]> matrix_bd_31_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 17, 87]> matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29_cast_fp16)[name = tensor<string, []>("matrix_bd_31_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_1810_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = tensor<string, []>("op_1810_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_29_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_29_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 17, 87]> _inversed_scores_29_cast_fp16 = mul(x = var_1810_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_29_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> scores_31_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_31_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_1816_cast_fp16 = softmax(axis = var_62, x = scores_31_cast_fp16)[name = tensor<string, []>("op_1816_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> input_405_cast_fp16 = select(a = var_40_to_fp16, b = var_1816_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_405_cast_fp16")];
+            tensor<bool, []> x_195_transpose_x_0 = const()[name = tensor<string, []>("x_195_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_195_transpose_y_0 = const()[name = tensor<string, []>("x_195_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 87, 64]> value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_15_cast_fp16)[name = tensor<string, []>("transpose_175")];
+            tensor<fp16, [1, 8, 17, 64]> x_195_cast_fp16 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405_cast_fp16, y = value_17_cast_fp16)[name = tensor<string, []>("x_195_cast_fp16")];
+            tensor<int32, [4]> var_1820_perm_0 = const()[name = tensor<string, []>("op_1820_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1821 = const()[name = tensor<string, []>("op_1821"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 17, 8, 64]> var_1820_cast_fp16 = transpose(perm = var_1820_perm_0, x = x_195_cast_fp16)[name = tensor<string, []>("transpose_171")];
+            tensor<fp16, [1, 17, 512]> input_407_cast_fp16 = reshape(shape = var_1821, x = var_1820_cast_fp16)[name = tensor<string, []>("input_407_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_7_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96520512)))];
+            tensor<fp16, [1, 17, 512]> linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16, x = input_407_cast_fp16)[name = tensor<string, []>("linear_70_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = tensor<string, []>("input_411_cast_fp16")];
+            tensor<int32, [1]> x_199_axes_0 = const()[name = tensor<string, []>("x_199_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97044864)))];
+            tensor<fp16, [512]> encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97045952)))];
+            tensor<fp16, [1, 17, 512]> x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = tensor<string, []>("x_199_cast_fp16")];
+            tensor<int32, [3]> input_413_perm_0 = const()[name = tensor<string, []>("input_413_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_415_pad_type_0 = const()[name = tensor<string, []>("input_415_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_415_strides_0 = const()[name = tensor<string, []>("input_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_415_pad_0 = const()[name = tensor<string, []>("input_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_415_dilations_0 = const()[name = tensor<string, []>("input_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_415_groups_0 = const()[name = tensor<string, []>("input_415_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_7_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97047040)))];
+            tensor<fp16, [1, 512, 17]> input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = tensor<string, []>("transpose_170")];
+            tensor<fp16, [1, 1024, 17]> input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16, x = input_413_cast_fp16)[name = tensor<string, []>("input_415_cast_fp16")];
+            tensor<int32, []> x_201_split_num_splits_0 = const()[name = tensor<string, []>("x_201_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_201_split_axis_0 = const()[name = tensor<string, []>("x_201_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 17]> x_201_split_cast_fp16_0, tensor<fp16, [1, 512, 17]> x_201_split_cast_fp16_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415_cast_fp16)[name = tensor<string, []>("x_201_split_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = tensor<string, []>("x_201_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_201_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> input_417_cast_fp16 = select(a = var_40_to_fp16, b = x_201_cast_fp16, cond = var_418)[name = tensor<string, []>("input_417_cast_fp16")];
+            tensor<bool, []> new_x_31_interleave_0 = const()[name = tensor<string, []>("new_x_31_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 25]> new_x_31_cast_fp16 = concat(axis = var_62, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = tensor<string, []>("new_x_31_cast_fp16")];
+            tensor<int32, [3]> var_1859_begin_0 = const()[name = tensor<string, []>("op_1859_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
+            tensor<int32, [3]> var_1859_end_0 = const()[name = tensor<string, []>("op_1859_end_0"), val = tensor<int32, [3]>([1, 512, 25])];
+            tensor<bool, [3]> var_1859_end_mask_0 = const()[name = tensor<string, []>("op_1859_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_1859_cast_fp16 = slice_by_index(begin = var_1859_begin_0, end = var_1859_end_0, end_mask = var_1859_end_mask_0, x = new_x_31_cast_fp16)[name = tensor<string, []>("op_1859_cast_fp16")];
+            tensor<string, []> x_203_pad_type_0 = const()[name = tensor<string, []>("x_203_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_203_groups_0 = const()[name = tensor<string, []>("x_203_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_203_strides_0 = const()[name = tensor<string, []>("x_203_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_203_pad_0 = const()[name = tensor<string, []>("x_203_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_203_dilations_0 = const()[name = tensor<string, []>("x_203_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_7_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98095680)))];
+            tensor<fp16, [1, 512, 17]> x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16, x = new_x_31_cast_fp16)[name = tensor<string, []>("x_203_cast_fp16")];
+            tensor<int32, [3]> input_419_perm_0 = const()[name = tensor<string, []>("input_419_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_205_axes_0 = const()[name = tensor<string, []>("x_205_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98104960)))];
+            tensor<fp16, [512]> encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98106048)))];
+            tensor<fp16, [1, 17, 512]> input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = tensor<string, []>("transpose_169")];
+            tensor<fp16, [1, 17, 512]> x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = tensor<string, []>("x_205_cast_fp16")];
+            tensor<int32, [3]> input_421_perm_0 = const()[name = tensor<string, []>("input_421_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 17]> input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = tensor<string, []>("transpose_168")];
+            tensor<fp16, [1, 512, 17]> input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor<string, []>("input_423_cast_fp16")];
+            tensor<string, []> x_207_pad_type_0 = const()[name = tensor<string, []>("x_207_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_207_strides_0 = const()[name = tensor<string, []>("x_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_207_pad_0 = const()[name = tensor<string, []>("x_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_207_dilations_0 = const()[name = tensor<string, []>("x_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_207_groups_0 = const()[name = tensor<string, []>("x_207_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_7_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98107136)))];
+            tensor<fp16, [1, 512, 17]> x_207_cast_fp16 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16, x = input_423_cast_fp16)[name = tensor<string, []>("x_207_cast_fp16")];
+            tensor<int32, [3]> input_425_perm_0 = const()[name = tensor<string, []>("input_425_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 17, 512]> input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = tensor<string, []>("transpose_167")];
+            tensor<fp16, [1, 17, 512]> input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = tensor<string, []>("input_427_cast_fp16")];
+            tensor<int32, [1]> input_429_axes_0 = const()[name = tensor<string, []>("input_429_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98631488)))];
+            tensor<fp16, [512]> encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98632576)))];
+            tensor<fp16, [1, 17, 512]> input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = tensor<string, []>("input_429_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_7_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98633664)))];
+            tensor<fp16, [1, 17, 2048]> linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16, x = input_429_cast_fp16)[name = tensor<string, []>("linear_71_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = tensor<string, []>("input_433_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_7_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100730880)))];
+            tensor<fp16, [1, 17, 512]> linear_72_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16, x = input_433_cast_fp16)[name = tensor<string, []>("linear_72_cast_fp16")];
+            tensor<fp16, []> var_1900_to_fp16 = const()[name = tensor<string, []>("op_1900_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_1901_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_1900_to_fp16)[name = tensor<string, []>("op_1901_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_1901_cast_fp16)[name = tensor<string, []>("input_439_cast_fp16")];
+            tensor<int32, [1]> input_441_axes_0 = const()[name = tensor<string, []>("input_441_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_7_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102828096)))];
+            tensor<fp16, [512]> encoder_layers_7_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102829184)))];
+            tensor<fp16, [1, 17, 512]> input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = tensor<string, []>("input_441_cast_fp16")];
+            tensor<int32, [4]> cache_33_begin_0 = const()[name = tensor<string, []>("cache_33_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_33_end_0 = const()[name = tensor<string, []>("cache_33_end_0"), val = tensor<int32, [4]>([9, 1, 70, 512])];
+            tensor<bool, [4]> cache_33_end_mask_0 = const()[name = tensor<string, []>("cache_33_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_33_squeeze_mask_0 = const()[name = tensor<string, []>("cache_33_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_33_cast_fp16 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_33_cast_fp16")];
+            tensor<int32, [4]> cache_35_begin_0 = const()[name = tensor<string, []>("cache_35_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_35_end_0 = const()[name = tensor<string, []>("cache_35_end_0"), val = tensor<int32, [4]>([9, 1, 512, 8])];
+            tensor<bool, [4]> cache_35_end_mask_0 = const()[name = tensor<string, []>("cache_35_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_35_squeeze_mask_0 = const()[name = tensor<string, []>("cache_35_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_35_cast_fp16 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_35_cast_fp16")];
+            tensor<int32, [1]> input_443_axes_0 = const()[name = tensor<string, []>("input_443_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102830272)))];
+            tensor<fp16, [512]> encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102831360)))];
+            tensor<fp16, [1, 17, 512]> input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = tensor<string, []>("input_443_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_8_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102832448)))];
+            tensor<fp16, [1, 17, 2048]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16, x = input_443_cast_fp16)[name = tensor<string, []>("linear_73_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = tensor<string, []>("input_447_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_8_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104929664)))];
+            tensor<fp16, [1, 17, 512]> linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16, x = input_447_cast_fp16)[name = tensor<string, []>("linear_74_cast_fp16")];
+            tensor<fp16, []> var_1935_to_fp16 = const()[name = tensor<string, []>("op_1935_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_1936_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_1935_to_fp16)[name = tensor<string, []>("op_1936_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_1936_cast_fp16)[name = tensor<string, []>("input_453_cast_fp16")];
+            tensor<int32, [1]> key_17_axes_0 = const()[name = tensor<string, []>("key_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107026880)))];
+            tensor<fp16, [512]> encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107027968)))];
+            tensor<fp16, [1, 17, 512]> key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = tensor<string, []>("key_17_cast_fp16")];
+            tensor<bool, []> input_455_interleave_0 = const()[name = tensor<string, []>("input_455_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 87, 512]> input_455_cast_fp16 = concat(axis = var_64, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = tensor<string, []>("input_455_cast_fp16")];
+            tensor<int32, [3]> var_1958_begin_0 = const()[name = tensor<string, []>("op_1958_begin_0"), val = tensor<int32, [3]>([0, 17, 0])];
+            tensor<int32, [3]> var_1958_end_0 = const()[name = tensor<string, []>("op_1958_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_1958_end_mask_0 = const()[name = tensor<string, []>("op_1958_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 53, 512]> var_1958_cast_fp16 = slice_by_index(begin = var_1958_begin_0, end = var_1958_end_0, end_mask = var_1958_end_mask_0, x = cache_33_cast_fp16)[name = tensor<string, []>("op_1958_cast_fp16")];
+            tensor<bool, []> var_1964_interleave_0 = const()[name = tensor<string, []>("op_1964_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_1964_cast_fp16 = concat(axis = var_64, interleave = var_1964_interleave_0, values = (var_1958_cast_fp16, key_17_cast_fp16))[name = tensor<string, []>("op_1964_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_8_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107029056)))];
+            tensor<fp16, [1, 17, 512]> linear_75_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16, x = key_17_cast_fp16)[name = tensor<string, []>("linear_75_cast_fp16")];
+            tensor<int32, [4]> var_1968 = const()[name = tensor<string, []>("op_1968"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 17, 8, 64]> q_49_cast_fp16 = reshape(shape = var_1968, x = linear_75_cast_fp16)[name = tensor<string, []>("q_49_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_8_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107553408)))];
+            tensor<fp16, [1, 87, 512]> linear_76_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16, x = input_455_cast_fp16)[name = tensor<string, []>("linear_76_cast_fp16")];
+            tensor<int32, [4]> var_1972 = const()[name = tensor<string, []>("op_1972"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> k_33_cast_fp16 = reshape(shape = var_1972, x = linear_76_cast_fp16)[name = tensor<string, []>("k_33_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_8_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108077760)))];
+            tensor<fp16, [1, 87, 512]> linear_77_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16, x = input_455_cast_fp16)[name = tensor<string, []>("linear_77_cast_fp16")];
+            tensor<int32, [4]> var_1976 = const()[name = tensor<string, []>("op_1976"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> v_17_cast_fp16 = reshape(shape = var_1976, x = linear_77_cast_fp16)[name = tensor<string, []>("v_17_cast_fp16")];
+            tensor<int32, [4]> value_19_perm_0 = const()[name = tensor<string, []>("value_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108602112)))];
+            tensor<fp16, [1, 17, 8, 64]> var_1988_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1988_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108603200)))];
+            tensor<fp16, [1, 17, 8, 64]> var_1990_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1990_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_17_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_215_transpose_x_0 = const()[name = tensor<string, []>("x_215_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_215_transpose_y_0 = const()[name = tensor<string, []>("x_215_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 173]> var_1992_to_fp16 = const()[name = tensor<string, []>("op_1992_to_fp16"), val = tensor<fp16, [1, 8, 64, 173]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108604288)))];
+            tensor<fp16, [1, 8, 17, 64]> q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_1990_cast_fp16)[name = tensor<string, []>("transpose_165")];
+            tensor<fp16, [1, 8, 17, 173]> x_215_cast_fp16 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = var_1992_to_fp16)[name = tensor<string, []>("x_215_cast_fp16")];
+            tensor<int32, [8]> x_217_pad_0 = const()[name = tensor<string, []>("x_217_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_217_mode_0 = const()[name = tensor<string, []>("x_217_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_127_to_fp16 = const()[name = tensor<string, []>("const_127_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 17, 174]> x_217_cast_fp16 = pad(constant_val = const_127_to_fp16, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215_cast_fp16)[name = tensor<string, []>("x_217_cast_fp16")];
+            tensor<int32, [4]> var_2000 = const()[name = tensor<string, []>("op_2000"), val = tensor<int32, [4]>([1, 8, -1, 17])];
+            tensor<fp16, [1, 8, 174, 17]> x_219_cast_fp16 = reshape(shape = var_2000, x = x_217_cast_fp16)[name = tensor<string, []>("x_219_cast_fp16")];
+            tensor<int32, [4]> var_2004_begin_0 = const()[name = tensor<string, []>("op_2004_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2004_end_0 = const()[name = tensor<string, []>("op_2004_end_0"), val = tensor<int32, [4]>([1, 8, 174, 17])];
+            tensor<bool, [4]> var_2004_end_mask_0 = const()[name = tensor<string, []>("op_2004_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 173, 17]> var_2004_cast_fp16 = slice_by_index(begin = var_2004_begin_0, end = var_2004_end_0, end_mask = var_2004_end_mask_0, x = x_219_cast_fp16)[name = tensor<string, []>("op_2004_cast_fp16")];
+            tensor<int32, [4]> var_2005 = const()[name = tensor<string, []>("op_2005"), val = tensor<int32, [4]>([1, 8, 17, 173])];
+            tensor<fp16, [1, 8, 17, 173]> matrix_bd_33_cast_fp16 = reshape(shape = var_2005, x = var_2004_cast_fp16)[name = tensor<string, []>("matrix_bd_33_cast_fp16")];
+            tensor<bool, []> matrix_ac_17_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_17_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_17_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_17_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_67_perm_0 = const()[name = tensor<string, []>("transpose_67_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_68_perm_0 = const()[name = tensor<string, []>("transpose_68_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 87]> transpose_68 = transpose(perm = transpose_68_perm_0, x = k_33_cast_fp16)[name = tensor<string, []>("transpose_163")];
+            tensor<fp16, [1, 8, 17, 64]> transpose_67 = transpose(perm = transpose_67_perm_0, x = var_1988_cast_fp16)[name = tensor<string, []>("transpose_164")];
+            tensor<fp16, [1, 8, 17, 87]> matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_67, y = transpose_68)[name = tensor<string, []>("matrix_ac_17_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_35_begin_0 = const()[name = tensor<string, []>("matrix_bd_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_35_end_0 = const()[name = tensor<string, []>("matrix_bd_35_end_0"), val = tensor<int32, [4]>([1, 8, 17, 87])];
+            tensor<bool, [4]> matrix_bd_35_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 17, 87]> matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33_cast_fp16)[name = tensor<string, []>("matrix_bd_35_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_2014_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = tensor<string, []>("op_2014_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_33_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_33_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 17, 87]> _inversed_scores_33_cast_fp16 = mul(x = var_2014_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_33_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> scores_35_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_35_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_2020_cast_fp16 = softmax(axis = var_62, x = scores_35_cast_fp16)[name = tensor<string, []>("op_2020_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> input_457_cast_fp16 = select(a = var_40_to_fp16, b = var_2020_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_457_cast_fp16")];
+            tensor<bool, []> x_221_transpose_x_0 = const()[name = tensor<string, []>("x_221_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_221_transpose_y_0 = const()[name = tensor<string, []>("x_221_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 87, 64]> value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_17_cast_fp16)[name = tensor<string, []>("transpose_166")];
+            tensor<fp16, [1, 8, 17, 64]> x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457_cast_fp16, y = value_19_cast_fp16)[name = tensor<string, []>("x_221_cast_fp16")];
+            tensor<int32, [4]> var_2024_perm_0 = const()[name = tensor<string, []>("op_2024_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2025 = const()[name = tensor<string, []>("op_2025"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 17, 8, 64]> var_2024_cast_fp16 = transpose(perm = var_2024_perm_0, x = x_221_cast_fp16)[name = tensor<string, []>("transpose_162")];
+            tensor<fp16, [1, 17, 512]> input_459_cast_fp16 = reshape(shape = var_2025, x = var_2024_cast_fp16)[name = tensor<string, []>("input_459_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_8_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108781504)))];
+            tensor<fp16, [1, 17, 512]> linear_79_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16, x = input_459_cast_fp16)[name = tensor<string, []>("linear_79_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = tensor<string, []>("input_463_cast_fp16")];
+            tensor<int32, [1]> x_225_axes_0 = const()[name = tensor<string, []>("x_225_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109305856)))];
+            tensor<fp16, [512]> encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109306944)))];
+            tensor<fp16, [1, 17, 512]> x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = tensor<string, []>("x_225_cast_fp16")];
+            tensor<int32, [3]> input_465_perm_0 = const()[name = tensor<string, []>("input_465_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_467_pad_type_0 = const()[name = tensor<string, []>("input_467_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_467_strides_0 = const()[name = tensor<string, []>("input_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_467_pad_0 = const()[name = tensor<string, []>("input_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_467_dilations_0 = const()[name = tensor<string, []>("input_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_467_groups_0 = const()[name = tensor<string, []>("input_467_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_8_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109308032)))];
+            tensor<fp16, [1, 512, 17]> input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = tensor<string, []>("transpose_161")];
+            tensor<fp16, [1, 1024, 17]> input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16, x = input_465_cast_fp16)[name = tensor<string, []>("input_467_cast_fp16")];
+            tensor<int32, []> x_227_split_num_splits_0 = const()[name = tensor<string, []>("x_227_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_227_split_axis_0 = const()[name = tensor<string, []>("x_227_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 17]> x_227_split_cast_fp16_0, tensor<fp16, [1, 512, 17]> x_227_split_cast_fp16_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467_cast_fp16)[name = tensor<string, []>("x_227_split_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = tensor<string, []>("x_227_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_227_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> input_469_cast_fp16 = select(a = var_40_to_fp16, b = x_227_cast_fp16, cond = var_418)[name = tensor<string, []>("input_469_cast_fp16")];
+            tensor<bool, []> new_x_35_interleave_0 = const()[name = tensor<string, []>("new_x_35_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 25]> new_x_35_cast_fp16 = concat(axis = var_62, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = tensor<string, []>("new_x_35_cast_fp16")];
+            tensor<int32, [3]> var_2063_begin_0 = const()[name = tensor<string, []>("op_2063_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
+            tensor<int32, [3]> var_2063_end_0 = const()[name = tensor<string, []>("op_2063_end_0"), val = tensor<int32, [3]>([1, 512, 25])];
+            tensor<bool, [3]> var_2063_end_mask_0 = const()[name = tensor<string, []>("op_2063_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_2063_cast_fp16 = slice_by_index(begin = var_2063_begin_0, end = var_2063_end_0, end_mask = var_2063_end_mask_0, x = new_x_35_cast_fp16)[name = tensor<string, []>("op_2063_cast_fp16")];
+            tensor<string, []> x_229_pad_type_0 = const()[name = tensor<string, []>("x_229_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_229_groups_0 = const()[name = tensor<string, []>("x_229_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_229_strides_0 = const()[name = tensor<string, []>("x_229_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_229_pad_0 = const()[name = tensor<string, []>("x_229_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_229_dilations_0 = const()[name = tensor<string, []>("x_229_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_8_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110356672)))];
+            tensor<fp16, [1, 512, 17]> x_229_cast_fp16 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16, x = new_x_35_cast_fp16)[name = tensor<string, []>("x_229_cast_fp16")];
+            tensor<int32, [3]> input_471_perm_0 = const()[name = tensor<string, []>("input_471_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_231_axes_0 = const()[name = tensor<string, []>("x_231_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110365952)))];
+            tensor<fp16, [512]> encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110367040)))];
+            tensor<fp16, [1, 17, 512]> input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = tensor<string, []>("transpose_160")];
+            tensor<fp16, [1, 17, 512]> x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = tensor<string, []>("x_231_cast_fp16")];
+            tensor<int32, [3]> input_473_perm_0 = const()[name = tensor<string, []>("input_473_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 17]> input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = tensor<string, []>("transpose_159")];
+            tensor<fp16, [1, 512, 17]> input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = tensor<string, []>("input_475_cast_fp16")];
+            tensor<string, []> x_233_pad_type_0 = const()[name = tensor<string, []>("x_233_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_233_strides_0 = const()[name = tensor<string, []>("x_233_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_233_pad_0 = const()[name = tensor<string, []>("x_233_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_233_dilations_0 = const()[name = tensor<string, []>("x_233_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_233_groups_0 = const()[name = tensor<string, []>("x_233_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_8_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110368128)))];
+            tensor<fp16, [1, 512, 17]> x_233_cast_fp16 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16, x = input_475_cast_fp16)[name = tensor<string, []>("x_233_cast_fp16")];
+            tensor<int32, [3]> input_477_perm_0 = const()[name = tensor<string, []>("input_477_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 17, 512]> input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = tensor<string, []>("transpose_158")];
+            tensor<fp16, [1, 17, 512]> input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = tensor<string, []>("input_479_cast_fp16")];
+            tensor<int32, [1]> input_481_axes_0 = const()[name = tensor<string, []>("input_481_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110892480)))];
+            tensor<fp16, [512]> encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110893568)))];
+            tensor<fp16, [1, 17, 512]> input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = tensor<string, []>("input_481_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_8_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110894656)))];
+            tensor<fp16, [1, 17, 2048]> linear_80_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16, x = input_481_cast_fp16)[name = tensor<string, []>("linear_80_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = tensor<string, []>("input_485_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_8_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112991872)))];
+            tensor<fp16, [1, 17, 512]> linear_81_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16, x = input_485_cast_fp16)[name = tensor<string, []>("linear_81_cast_fp16")];
+            tensor<fp16, []> var_2104_to_fp16 = const()[name = tensor<string, []>("op_2104_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_2105_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2104_to_fp16)[name = tensor<string, []>("op_2105_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2105_cast_fp16)[name = tensor<string, []>("input_491_cast_fp16")];
+            tensor<int32, [1]> input_493_axes_0 = const()[name = tensor<string, []>("input_493_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_8_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115089088)))];
+            tensor<fp16, [512]> encoder_layers_8_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115090176)))];
+            tensor<fp16, [1, 17, 512]> input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = tensor<string, []>("input_493_cast_fp16")];
+            tensor<int32, [4]> cache_37_begin_0 = const()[name = tensor<string, []>("cache_37_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_37_end_0 = const()[name = tensor<string, []>("cache_37_end_0"), val = tensor<int32, [4]>([10, 1, 70, 512])];
+            tensor<bool, [4]> cache_37_end_mask_0 = const()[name = tensor<string, []>("cache_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_37_squeeze_mask_0 = const()[name = tensor<string, []>("cache_37_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_37_cast_fp16 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_37_cast_fp16")];
+            tensor<int32, [4]> cache_39_begin_0 = const()[name = tensor<string, []>("cache_39_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_39_end_0 = const()[name = tensor<string, []>("cache_39_end_0"), val = tensor<int32, [4]>([10, 1, 512, 8])];
+            tensor<bool, [4]> cache_39_end_mask_0 = const()[name = tensor<string, []>("cache_39_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_39_squeeze_mask_0 = const()[name = tensor<string, []>("cache_39_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_39_cast_fp16 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_39_cast_fp16")];
+            tensor<int32, [1]> input_495_axes_0 = const()[name = tensor<string, []>("input_495_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115091264)))];
+            tensor<fp16, [512]> encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115092352)))];
+            tensor<fp16, [1, 17, 512]> input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = tensor<string, []>("input_495_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_9_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115093440)))];
+            tensor<fp16, [1, 17, 2048]> linear_82_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16, x = input_495_cast_fp16)[name = tensor<string, []>("linear_82_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = tensor<string, []>("input_499_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_9_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117190656)))];
+            tensor<fp16, [1, 17, 512]> linear_83_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16, x = input_499_cast_fp16)[name = tensor<string, []>("linear_83_cast_fp16")];
+            tensor<fp16, []> var_2139_to_fp16 = const()[name = tensor<string, []>("op_2139_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_2140_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2139_to_fp16)[name = tensor<string, []>("op_2140_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2140_cast_fp16)[name = tensor<string, []>("input_505_cast_fp16")];
+            tensor<int32, [1]> key_19_axes_0 = const()[name = tensor<string, []>("key_19_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119287872)))];
+            tensor<fp16, [512]> encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119288960)))];
+            tensor<fp16, [1, 17, 512]> key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = tensor<string, []>("key_19_cast_fp16")];
+            tensor<bool, []> input_507_interleave_0 = const()[name = tensor<string, []>("input_507_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 87, 512]> input_507_cast_fp16 = concat(axis = var_64, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = tensor<string, []>("input_507_cast_fp16")];
+            tensor<int32, [3]> var_2162_begin_0 = const()[name = tensor<string, []>("op_2162_begin_0"), val = tensor<int32, [3]>([0, 17, 0])];
+            tensor<int32, [3]> var_2162_end_0 = const()[name = tensor<string, []>("op_2162_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_2162_end_mask_0 = const()[name = tensor<string, []>("op_2162_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 53, 512]> var_2162_cast_fp16 = slice_by_index(begin = var_2162_begin_0, end = var_2162_end_0, end_mask = var_2162_end_mask_0, x = cache_37_cast_fp16)[name = tensor<string, []>("op_2162_cast_fp16")];
+            tensor<bool, []> var_2168_interleave_0 = const()[name = tensor<string, []>("op_2168_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_2168_cast_fp16 = concat(axis = var_64, interleave = var_2168_interleave_0, values = (var_2162_cast_fp16, key_19_cast_fp16))[name = tensor<string, []>("op_2168_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_9_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119290048)))];
+            tensor<fp16, [1, 17, 512]> linear_84_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16, x = key_19_cast_fp16)[name = tensor<string, []>("linear_84_cast_fp16")];
+            tensor<int32, [4]> var_2172 = const()[name = tensor<string, []>("op_2172"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 17, 8, 64]> q_55_cast_fp16 = reshape(shape = var_2172, x = linear_84_cast_fp16)[name = tensor<string, []>("q_55_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_9_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119814400)))];
+            tensor<fp16, [1, 87, 512]> linear_85_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16, x = input_507_cast_fp16)[name = tensor<string, []>("linear_85_cast_fp16")];
+            tensor<int32, [4]> var_2176 = const()[name = tensor<string, []>("op_2176"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> k_37_cast_fp16 = reshape(shape = var_2176, x = linear_85_cast_fp16)[name = tensor<string, []>("k_37_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_9_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120338752)))];
+            tensor<fp16, [1, 87, 512]> linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16, x = input_507_cast_fp16)[name = tensor<string, []>("linear_86_cast_fp16")];
+            tensor<int32, [4]> var_2180 = const()[name = tensor<string, []>("op_2180"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> v_19_cast_fp16 = reshape(shape = var_2180, x = linear_86_cast_fp16)[name = tensor<string, []>("v_19_cast_fp16")];
+            tensor<int32, [4]> value_21_perm_0 = const()[name = tensor<string, []>("value_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120863104)))];
+            tensor<fp16, [1, 17, 8, 64]> var_2192_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2192_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120864192)))];
+            tensor<fp16, [1, 17, 8, 64]> var_2194_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2194_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_19_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_241_transpose_x_0 = const()[name = tensor<string, []>("x_241_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_241_transpose_y_0 = const()[name = tensor<string, []>("x_241_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 173]> var_2196_to_fp16 = const()[name = tensor<string, []>("op_2196_to_fp16"), val = tensor<fp16, [1, 8, 64, 173]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120865280)))];
+            tensor<fp16, [1, 8, 17, 64]> q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2194_cast_fp16)[name = tensor<string, []>("transpose_156")];
+            tensor<fp16, [1, 8, 17, 173]> x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = var_2196_to_fp16)[name = tensor<string, []>("x_241_cast_fp16")];
+            tensor<int32, [8]> x_243_pad_0 = const()[name = tensor<string, []>("x_243_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_243_mode_0 = const()[name = tensor<string, []>("x_243_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_140_to_fp16 = const()[name = tensor<string, []>("const_140_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 17, 174]> x_243_cast_fp16 = pad(constant_val = const_140_to_fp16, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241_cast_fp16)[name = tensor<string, []>("x_243_cast_fp16")];
+            tensor<int32, [4]> var_2204 = const()[name = tensor<string, []>("op_2204"), val = tensor<int32, [4]>([1, 8, -1, 17])];
+            tensor<fp16, [1, 8, 174, 17]> x_245_cast_fp16 = reshape(shape = var_2204, x = x_243_cast_fp16)[name = tensor<string, []>("x_245_cast_fp16")];
+            tensor<int32, [4]> var_2208_begin_0 = const()[name = tensor<string, []>("op_2208_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2208_end_0 = const()[name = tensor<string, []>("op_2208_end_0"), val = tensor<int32, [4]>([1, 8, 174, 17])];
+            tensor<bool, [4]> var_2208_end_mask_0 = const()[name = tensor<string, []>("op_2208_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 173, 17]> var_2208_cast_fp16 = slice_by_index(begin = var_2208_begin_0, end = var_2208_end_0, end_mask = var_2208_end_mask_0, x = x_245_cast_fp16)[name = tensor<string, []>("op_2208_cast_fp16")];
+            tensor<int32, [4]> var_2209 = const()[name = tensor<string, []>("op_2209"), val = tensor<int32, [4]>([1, 8, 17, 173])];
+            tensor<fp16, [1, 8, 17, 173]> matrix_bd_37_cast_fp16 = reshape(shape = var_2209, x = var_2208_cast_fp16)[name = tensor<string, []>("matrix_bd_37_cast_fp16")];
+            tensor<bool, []> matrix_ac_19_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_19_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_19_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_19_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_69_perm_0 = const()[name = tensor<string, []>("transpose_69_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_70_perm_0 = const()[name = tensor<string, []>("transpose_70_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 87]> transpose_70 = transpose(perm = transpose_70_perm_0, x = k_37_cast_fp16)[name = tensor<string, []>("transpose_154")];
+            tensor<fp16, [1, 8, 17, 64]> transpose_69 = transpose(perm = transpose_69_perm_0, x = var_2192_cast_fp16)[name = tensor<string, []>("transpose_155")];
+            tensor<fp16, [1, 8, 17, 87]> matrix_ac_19_cast_fp16 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_69, y = transpose_70)[name = tensor<string, []>("matrix_ac_19_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_39_begin_0 = const()[name = tensor<string, []>("matrix_bd_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_39_end_0 = const()[name = tensor<string, []>("matrix_bd_39_end_0"), val = tensor<int32, [4]>([1, 8, 17, 87])];
+            tensor<bool, [4]> matrix_bd_39_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 17, 87]> matrix_bd_39_cast_fp16 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37_cast_fp16)[name = tensor<string, []>("matrix_bd_39_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_2218_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = tensor<string, []>("op_2218_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_37_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_37_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 17, 87]> _inversed_scores_37_cast_fp16 = mul(x = var_2218_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_37_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> scores_39_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_39_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_2224_cast_fp16 = softmax(axis = var_62, x = scores_39_cast_fp16)[name = tensor<string, []>("op_2224_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> input_509_cast_fp16 = select(a = var_40_to_fp16, b = var_2224_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_509_cast_fp16")];
+            tensor<bool, []> x_247_transpose_x_0 = const()[name = tensor<string, []>("x_247_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_247_transpose_y_0 = const()[name = tensor<string, []>("x_247_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 87, 64]> value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_19_cast_fp16)[name = tensor<string, []>("transpose_157")];
+            tensor<fp16, [1, 8, 17, 64]> x_247_cast_fp16 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509_cast_fp16, y = value_21_cast_fp16)[name = tensor<string, []>("x_247_cast_fp16")];
+            tensor<int32, [4]> var_2228_perm_0 = const()[name = tensor<string, []>("op_2228_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2229 = const()[name = tensor<string, []>("op_2229"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 17, 8, 64]> var_2228_cast_fp16 = transpose(perm = var_2228_perm_0, x = x_247_cast_fp16)[name = tensor<string, []>("transpose_153")];
+            tensor<fp16, [1, 17, 512]> input_511_cast_fp16 = reshape(shape = var_2229, x = var_2228_cast_fp16)[name = tensor<string, []>("input_511_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_9_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121042496)))];
+            tensor<fp16, [1, 17, 512]> linear_88_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16, x = input_511_cast_fp16)[name = tensor<string, []>("linear_88_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = tensor<string, []>("input_515_cast_fp16")];
+            tensor<int32, [1]> x_251_axes_0 = const()[name = tensor<string, []>("x_251_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121566848)))];
+            tensor<fp16, [512]> encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121567936)))];
+            tensor<fp16, [1, 17, 512]> x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = tensor<string, []>("x_251_cast_fp16")];
+            tensor<int32, [3]> input_517_perm_0 = const()[name = tensor<string, []>("input_517_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_519_pad_type_0 = const()[name = tensor<string, []>("input_519_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_519_strides_0 = const()[name = tensor<string, []>("input_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_519_pad_0 = const()[name = tensor<string, []>("input_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_519_dilations_0 = const()[name = tensor<string, []>("input_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_519_groups_0 = const()[name = tensor<string, []>("input_519_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_9_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121569024)))];
+            tensor<fp16, [1, 512, 17]> input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = tensor<string, []>("transpose_152")];
+            tensor<fp16, [1, 1024, 17]> input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16, x = input_517_cast_fp16)[name = tensor<string, []>("input_519_cast_fp16")];
+            tensor<int32, []> x_253_split_num_splits_0 = const()[name = tensor<string, []>("x_253_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_253_split_axis_0 = const()[name = tensor<string, []>("x_253_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 17]> x_253_split_cast_fp16_0, tensor<fp16, [1, 512, 17]> x_253_split_cast_fp16_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519_cast_fp16)[name = tensor<string, []>("x_253_split_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = tensor<string, []>("x_253_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_253_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> input_521_cast_fp16 = select(a = var_40_to_fp16, b = x_253_cast_fp16, cond = var_418)[name = tensor<string, []>("input_521_cast_fp16")];
+            tensor<bool, []> new_x_39_interleave_0 = const()[name = tensor<string, []>("new_x_39_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 25]> new_x_39_cast_fp16 = concat(axis = var_62, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = tensor<string, []>("new_x_39_cast_fp16")];
+            tensor<int32, [3]> var_2267_begin_0 = const()[name = tensor<string, []>("op_2267_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
+            tensor<int32, [3]> var_2267_end_0 = const()[name = tensor<string, []>("op_2267_end_0"), val = tensor<int32, [3]>([1, 512, 25])];
+            tensor<bool, [3]> var_2267_end_mask_0 = const()[name = tensor<string, []>("op_2267_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_2267_cast_fp16 = slice_by_index(begin = var_2267_begin_0, end = var_2267_end_0, end_mask = var_2267_end_mask_0, x = new_x_39_cast_fp16)[name = tensor<string, []>("op_2267_cast_fp16")];
+            tensor<string, []> x_255_pad_type_0 = const()[name = tensor<string, []>("x_255_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_255_groups_0 = const()[name = tensor<string, []>("x_255_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_255_strides_0 = const()[name = tensor<string, []>("x_255_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_255_pad_0 = const()[name = tensor<string, []>("x_255_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_255_dilations_0 = const()[name = tensor<string, []>("x_255_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_9_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122617664)))];
+            tensor<fp16, [1, 512, 17]> x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16, x = new_x_39_cast_fp16)[name = tensor<string, []>("x_255_cast_fp16")];
+            tensor<int32, [3]> input_523_perm_0 = const()[name = tensor<string, []>("input_523_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_257_axes_0 = const()[name = tensor<string, []>("x_257_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122626944)))];
+            tensor<fp16, [512]> encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122628032)))];
+            tensor<fp16, [1, 17, 512]> input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = tensor<string, []>("transpose_151")];
+            tensor<fp16, [1, 17, 512]> x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = tensor<string, []>("x_257_cast_fp16")];
+            tensor<int32, [3]> input_525_perm_0 = const()[name = tensor<string, []>("input_525_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 17]> input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = tensor<string, []>("transpose_150")];
+            tensor<fp16, [1, 512, 17]> input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = tensor<string, []>("input_527_cast_fp16")];
+            tensor<string, []> x_259_pad_type_0 = const()[name = tensor<string, []>("x_259_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_259_strides_0 = const()[name = tensor<string, []>("x_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_259_pad_0 = const()[name = tensor<string, []>("x_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_259_dilations_0 = const()[name = tensor<string, []>("x_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_259_groups_0 = const()[name = tensor<string, []>("x_259_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_9_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122629120)))];
+            tensor<fp16, [1, 512, 17]> x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16, x = input_527_cast_fp16)[name = tensor<string, []>("x_259_cast_fp16")];
+            tensor<int32, [3]> input_529_perm_0 = const()[name = tensor<string, []>("input_529_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 17, 512]> input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = tensor<string, []>("transpose_149")];
+            tensor<fp16, [1, 17, 512]> input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = tensor<string, []>("input_531_cast_fp16")];
+            tensor<int32, [1]> input_533_axes_0 = const()[name = tensor<string, []>("input_533_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123153472)))];
+            tensor<fp16, [512]> encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123154560)))];
+            tensor<fp16, [1, 17, 512]> input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = tensor<string, []>("input_533_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_9_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123155648)))];
+            tensor<fp16, [1, 17, 2048]> linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16, x = input_533_cast_fp16)[name = tensor<string, []>("linear_89_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = tensor<string, []>("input_537_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_9_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125252864)))];
+            tensor<fp16, [1, 17, 512]> linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16, x = input_537_cast_fp16)[name = tensor<string, []>("linear_90_cast_fp16")];
+            tensor<fp16, []> var_2308_to_fp16 = const()[name = tensor<string, []>("op_2308_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_2309_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2308_to_fp16)[name = tensor<string, []>("op_2309_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2309_cast_fp16)[name = tensor<string, []>("input_543_cast_fp16")];
+            tensor<int32, [1]> input_545_axes_0 = const()[name = tensor<string, []>("input_545_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_9_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127350080)))];
+            tensor<fp16, [512]> encoder_layers_9_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127351168)))];
+            tensor<fp16, [1, 17, 512]> input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = tensor<string, []>("input_545_cast_fp16")];
+            tensor<int32, [4]> cache_41_begin_0 = const()[name = tensor<string, []>("cache_41_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_41_end_0 = const()[name = tensor<string, []>("cache_41_end_0"), val = tensor<int32, [4]>([11, 1, 70, 512])];
+            tensor<bool, [4]> cache_41_end_mask_0 = const()[name = tensor<string, []>("cache_41_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_41_squeeze_mask_0 = const()[name = tensor<string, []>("cache_41_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_41_cast_fp16 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_41_cast_fp16")];
+            tensor<int32, [4]> cache_43_begin_0 = const()[name = tensor<string, []>("cache_43_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_43_end_0 = const()[name = tensor<string, []>("cache_43_end_0"), val = tensor<int32, [4]>([11, 1, 512, 8])];
+            tensor<bool, [4]> cache_43_end_mask_0 = const()[name = tensor<string, []>("cache_43_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_43_squeeze_mask_0 = const()[name = tensor<string, []>("cache_43_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_43_cast_fp16 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_43_cast_fp16")];
+            tensor<int32, [1]> input_547_axes_0 = const()[name = tensor<string, []>("input_547_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127352256)))];
+            tensor<fp16, [512]> encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127353344)))];
+            tensor<fp16, [1, 17, 512]> input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = tensor<string, []>("input_547_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_10_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127354432)))];
+            tensor<fp16, [1, 17, 2048]> linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16, x = input_547_cast_fp16)[name = tensor<string, []>("linear_91_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = tensor<string, []>("input_551_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_10_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129451648)))];
+            tensor<fp16, [1, 17, 512]> linear_92_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16, x = input_551_cast_fp16)[name = tensor<string, []>("linear_92_cast_fp16")];
+            tensor<fp16, []> var_2343_to_fp16 = const()[name = tensor<string, []>("op_2343_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_2344_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2343_to_fp16)[name = tensor<string, []>("op_2344_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2344_cast_fp16)[name = tensor<string, []>("input_557_cast_fp16")];
+            tensor<int32, [1]> key_21_axes_0 = const()[name = tensor<string, []>("key_21_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131548864)))];
+            tensor<fp16, [512]> encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131549952)))];
+            tensor<fp16, [1, 17, 512]> key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = tensor<string, []>("key_21_cast_fp16")];
+            tensor<bool, []> input_559_interleave_0 = const()[name = tensor<string, []>("input_559_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 87, 512]> input_559_cast_fp16 = concat(axis = var_64, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = tensor<string, []>("input_559_cast_fp16")];
+            tensor<int32, [3]> var_2366_begin_0 = const()[name = tensor<string, []>("op_2366_begin_0"), val = tensor<int32, [3]>([0, 17, 0])];
+            tensor<int32, [3]> var_2366_end_0 = const()[name = tensor<string, []>("op_2366_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_2366_end_mask_0 = const()[name = tensor<string, []>("op_2366_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 53, 512]> var_2366_cast_fp16 = slice_by_index(begin = var_2366_begin_0, end = var_2366_end_0, end_mask = var_2366_end_mask_0, x = cache_41_cast_fp16)[name = tensor<string, []>("op_2366_cast_fp16")];
+            tensor<bool, []> var_2372_interleave_0 = const()[name = tensor<string, []>("op_2372_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_2372_cast_fp16 = concat(axis = var_64, interleave = var_2372_interleave_0, values = (var_2366_cast_fp16, key_21_cast_fp16))[name = tensor<string, []>("op_2372_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_10_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131551040)))];
+            tensor<fp16, [1, 17, 512]> linear_93_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16, x = key_21_cast_fp16)[name = tensor<string, []>("linear_93_cast_fp16")];
+            tensor<int32, [4]> var_2376 = const()[name = tensor<string, []>("op_2376"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 17, 8, 64]> q_61_cast_fp16 = reshape(shape = var_2376, x = linear_93_cast_fp16)[name = tensor<string, []>("q_61_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_10_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132075392)))];
+            tensor<fp16, [1, 87, 512]> linear_94_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16, x = input_559_cast_fp16)[name = tensor<string, []>("linear_94_cast_fp16")];
+            tensor<int32, [4]> var_2380 = const()[name = tensor<string, []>("op_2380"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> k_41_cast_fp16 = reshape(shape = var_2380, x = linear_94_cast_fp16)[name = tensor<string, []>("k_41_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_10_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132599744)))];
+            tensor<fp16, [1, 87, 512]> linear_95_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16, x = input_559_cast_fp16)[name = tensor<string, []>("linear_95_cast_fp16")];
+            tensor<int32, [4]> var_2384 = const()[name = tensor<string, []>("op_2384"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> v_21_cast_fp16 = reshape(shape = var_2384, x = linear_95_cast_fp16)[name = tensor<string, []>("v_21_cast_fp16")];
+            tensor<int32, [4]> value_23_perm_0 = const()[name = tensor<string, []>("value_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133124096)))];
+            tensor<fp16, [1, 17, 8, 64]> var_2396_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2396_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133125184)))];
+            tensor<fp16, [1, 17, 8, 64]> var_2398_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2398_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_21_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_267_transpose_x_0 = const()[name = tensor<string, []>("x_267_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_267_transpose_y_0 = const()[name = tensor<string, []>("x_267_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 173]> var_2400_to_fp16 = const()[name = tensor<string, []>("op_2400_to_fp16"), val = tensor<fp16, [1, 8, 64, 173]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133126272)))];
+            tensor<fp16, [1, 8, 17, 64]> q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2398_cast_fp16)[name = tensor<string, []>("transpose_147")];
+            tensor<fp16, [1, 8, 17, 173]> x_267_cast_fp16 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = var_2400_to_fp16)[name = tensor<string, []>("x_267_cast_fp16")];
+            tensor<int32, [8]> x_269_pad_0 = const()[name = tensor<string, []>("x_269_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_269_mode_0 = const()[name = tensor<string, []>("x_269_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_153_to_fp16 = const()[name = tensor<string, []>("const_153_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 17, 174]> x_269_cast_fp16 = pad(constant_val = const_153_to_fp16, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267_cast_fp16)[name = tensor<string, []>("x_269_cast_fp16")];
+            tensor<int32, [4]> var_2408 = const()[name = tensor<string, []>("op_2408"), val = tensor<int32, [4]>([1, 8, -1, 17])];
+            tensor<fp16, [1, 8, 174, 17]> x_271_cast_fp16 = reshape(shape = var_2408, x = x_269_cast_fp16)[name = tensor<string, []>("x_271_cast_fp16")];
+            tensor<int32, [4]> var_2412_begin_0 = const()[name = tensor<string, []>("op_2412_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2412_end_0 = const()[name = tensor<string, []>("op_2412_end_0"), val = tensor<int32, [4]>([1, 8, 174, 17])];
+            tensor<bool, [4]> var_2412_end_mask_0 = const()[name = tensor<string, []>("op_2412_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 173, 17]> var_2412_cast_fp16 = slice_by_index(begin = var_2412_begin_0, end = var_2412_end_0, end_mask = var_2412_end_mask_0, x = x_271_cast_fp16)[name = tensor<string, []>("op_2412_cast_fp16")];
+            tensor<int32, [4]> var_2413 = const()[name = tensor<string, []>("op_2413"), val = tensor<int32, [4]>([1, 8, 17, 173])];
+            tensor<fp16, [1, 8, 17, 173]> matrix_bd_41_cast_fp16 = reshape(shape = var_2413, x = var_2412_cast_fp16)[name = tensor<string, []>("matrix_bd_41_cast_fp16")];
+            tensor<bool, []> matrix_ac_21_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_21_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_21_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_21_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_71_perm_0 = const()[name = tensor<string, []>("transpose_71_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_72_perm_0 = const()[name = tensor<string, []>("transpose_72_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 87]> transpose_72 = transpose(perm = transpose_72_perm_0, x = k_41_cast_fp16)[name = tensor<string, []>("transpose_145")];
+            tensor<fp16, [1, 8, 17, 64]> transpose_71 = transpose(perm = transpose_71_perm_0, x = var_2396_cast_fp16)[name = tensor<string, []>("transpose_146")];
+            tensor<fp16, [1, 8, 17, 87]> matrix_ac_21_cast_fp16 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_71, y = transpose_72)[name = tensor<string, []>("matrix_ac_21_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_43_begin_0 = const()[name = tensor<string, []>("matrix_bd_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_43_end_0 = const()[name = tensor<string, []>("matrix_bd_43_end_0"), val = tensor<int32, [4]>([1, 8, 17, 87])];
+            tensor<bool, [4]> matrix_bd_43_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 17, 87]> matrix_bd_43_cast_fp16 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41_cast_fp16)[name = tensor<string, []>("matrix_bd_43_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_2422_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = tensor<string, []>("op_2422_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_41_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_41_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 17, 87]> _inversed_scores_41_cast_fp16 = mul(x = var_2422_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_41_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> scores_43_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_43_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_2428_cast_fp16 = softmax(axis = var_62, x = scores_43_cast_fp16)[name = tensor<string, []>("op_2428_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> input_561_cast_fp16 = select(a = var_40_to_fp16, b = var_2428_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_561_cast_fp16")];
+            tensor<bool, []> x_273_transpose_x_0 = const()[name = tensor<string, []>("x_273_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_273_transpose_y_0 = const()[name = tensor<string, []>("x_273_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 87, 64]> value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_21_cast_fp16)[name = tensor<string, []>("transpose_148")];
+            tensor<fp16, [1, 8, 17, 64]> x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561_cast_fp16, y = value_23_cast_fp16)[name = tensor<string, []>("x_273_cast_fp16")];
+            tensor<int32, [4]> var_2432_perm_0 = const()[name = tensor<string, []>("op_2432_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2433 = const()[name = tensor<string, []>("op_2433"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 17, 8, 64]> var_2432_cast_fp16 = transpose(perm = var_2432_perm_0, x = x_273_cast_fp16)[name = tensor<string, []>("transpose_144")];
+            tensor<fp16, [1, 17, 512]> input_563_cast_fp16 = reshape(shape = var_2433, x = var_2432_cast_fp16)[name = tensor<string, []>("input_563_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_10_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133303488)))];
+            tensor<fp16, [1, 17, 512]> linear_97_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16, x = input_563_cast_fp16)[name = tensor<string, []>("linear_97_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = tensor<string, []>("input_567_cast_fp16")];
+            tensor<int32, [1]> x_277_axes_0 = const()[name = tensor<string, []>("x_277_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133827840)))];
+            tensor<fp16, [512]> encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133828928)))];
+            tensor<fp16, [1, 17, 512]> x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = tensor<string, []>("x_277_cast_fp16")];
+            tensor<int32, [3]> input_569_perm_0 = const()[name = tensor<string, []>("input_569_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_571_pad_type_0 = const()[name = tensor<string, []>("input_571_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_571_strides_0 = const()[name = tensor<string, []>("input_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_571_pad_0 = const()[name = tensor<string, []>("input_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_571_dilations_0 = const()[name = tensor<string, []>("input_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_571_groups_0 = const()[name = tensor<string, []>("input_571_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_10_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133830016)))];
+            tensor<fp16, [1, 512, 17]> input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = tensor<string, []>("transpose_143")];
+            tensor<fp16, [1, 1024, 17]> input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16, x = input_569_cast_fp16)[name = tensor<string, []>("input_571_cast_fp16")];
+            tensor<int32, []> x_279_split_num_splits_0 = const()[name = tensor<string, []>("x_279_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_279_split_axis_0 = const()[name = tensor<string, []>("x_279_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 17]> x_279_split_cast_fp16_0, tensor<fp16, [1, 512, 17]> x_279_split_cast_fp16_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571_cast_fp16)[name = tensor<string, []>("x_279_split_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = tensor<string, []>("x_279_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_279_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> input_573_cast_fp16 = select(a = var_40_to_fp16, b = x_279_cast_fp16, cond = var_418)[name = tensor<string, []>("input_573_cast_fp16")];
+            tensor<bool, []> new_x_43_interleave_0 = const()[name = tensor<string, []>("new_x_43_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 25]> new_x_43_cast_fp16 = concat(axis = var_62, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = tensor<string, []>("new_x_43_cast_fp16")];
+            tensor<int32, [3]> var_2471_begin_0 = const()[name = tensor<string, []>("op_2471_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
+            tensor<int32, [3]> var_2471_end_0 = const()[name = tensor<string, []>("op_2471_end_0"), val = tensor<int32, [3]>([1, 512, 25])];
+            tensor<bool, [3]> var_2471_end_mask_0 = const()[name = tensor<string, []>("op_2471_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_2471_cast_fp16 = slice_by_index(begin = var_2471_begin_0, end = var_2471_end_0, end_mask = var_2471_end_mask_0, x = new_x_43_cast_fp16)[name = tensor<string, []>("op_2471_cast_fp16")];
+            tensor<string, []> x_281_pad_type_0 = const()[name = tensor<string, []>("x_281_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_281_groups_0 = const()[name = tensor<string, []>("x_281_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_281_strides_0 = const()[name = tensor<string, []>("x_281_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_281_pad_0 = const()[name = tensor<string, []>("x_281_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_281_dilations_0 = const()[name = tensor<string, []>("x_281_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_10_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134878656)))];
+            tensor<fp16, [1, 512, 17]> x_281_cast_fp16 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16, x = new_x_43_cast_fp16)[name = tensor<string, []>("x_281_cast_fp16")];
+            tensor<int32, [3]> input_575_perm_0 = const()[name = tensor<string, []>("input_575_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_283_axes_0 = const()[name = tensor<string, []>("x_283_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134887936)))];
+            tensor<fp16, [512]> encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134889024)))];
+            tensor<fp16, [1, 17, 512]> input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = tensor<string, []>("transpose_142")];
+            tensor<fp16, [1, 17, 512]> x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = tensor<string, []>("x_283_cast_fp16")];
+            tensor<int32, [3]> input_577_perm_0 = const()[name = tensor<string, []>("input_577_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 17]> input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = tensor<string, []>("transpose_141")];
+            tensor<fp16, [1, 512, 17]> input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = tensor<string, []>("input_579_cast_fp16")];
+            tensor<string, []> x_285_pad_type_0 = const()[name = tensor<string, []>("x_285_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_285_strides_0 = const()[name = tensor<string, []>("x_285_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_285_pad_0 = const()[name = tensor<string, []>("x_285_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_285_dilations_0 = const()[name = tensor<string, []>("x_285_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_285_groups_0 = const()[name = tensor<string, []>("x_285_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_10_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134890112)))];
+            tensor<fp16, [1, 512, 17]> x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16, x = input_579_cast_fp16)[name = tensor<string, []>("x_285_cast_fp16")];
+            tensor<int32, [3]> input_581_perm_0 = const()[name = tensor<string, []>("input_581_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 17, 512]> input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = tensor<string, []>("transpose_140")];
+            tensor<fp16, [1, 17, 512]> input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = tensor<string, []>("input_583_cast_fp16")];
+            tensor<int32, [1]> input_585_axes_0 = const()[name = tensor<string, []>("input_585_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135414464)))];
+            tensor<fp16, [512]> encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135415552)))];
+            tensor<fp16, [1, 17, 512]> input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = tensor<string, []>("input_585_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_10_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135416640)))];
+            tensor<fp16, [1, 17, 2048]> linear_98_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16, x = input_585_cast_fp16)[name = tensor<string, []>("linear_98_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = tensor<string, []>("input_589_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_10_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137513856)))];
+            tensor<fp16, [1, 17, 512]> linear_99_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16, x = input_589_cast_fp16)[name = tensor<string, []>("linear_99_cast_fp16")];
+            tensor<fp16, []> var_2512_to_fp16 = const()[name = tensor<string, []>("op_2512_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_2513_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2512_to_fp16)[name = tensor<string, []>("op_2513_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2513_cast_fp16)[name = tensor<string, []>("input_595_cast_fp16")];
+            tensor<int32, [1]> input_597_axes_0 = const()[name = tensor<string, []>("input_597_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_10_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139611072)))];
+            tensor<fp16, [512]> encoder_layers_10_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139612160)))];
+            tensor<fp16, [1, 17, 512]> input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = tensor<string, []>("input_597_cast_fp16")];
+            tensor<int32, [4]> cache_45_begin_0 = const()[name = tensor<string, []>("cache_45_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_45_end_0 = const()[name = tensor<string, []>("cache_45_end_0"), val = tensor<int32, [4]>([12, 1, 70, 512])];
+            tensor<bool, [4]> cache_45_end_mask_0 = const()[name = tensor<string, []>("cache_45_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_45_squeeze_mask_0 = const()[name = tensor<string, []>("cache_45_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_45_cast_fp16 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_45_cast_fp16")];
+            tensor<int32, [4]> cache_47_begin_0 = const()[name = tensor<string, []>("cache_47_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_47_end_0 = const()[name = tensor<string, []>("cache_47_end_0"), val = tensor<int32, [4]>([12, 1, 512, 8])];
+            tensor<bool, [4]> cache_47_end_mask_0 = const()[name = tensor<string, []>("cache_47_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_47_squeeze_mask_0 = const()[name = tensor<string, []>("cache_47_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_47_cast_fp16 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_47_cast_fp16")];
+            tensor<int32, [1]> input_599_axes_0 = const()[name = tensor<string, []>("input_599_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139613248)))];
+            tensor<fp16, [512]> encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139614336)))];
+            tensor<fp16, [1, 17, 512]> input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = tensor<string, []>("input_599_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_11_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139615424)))];
+            tensor<fp16, [1, 17, 2048]> linear_100_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16, x = input_599_cast_fp16)[name = tensor<string, []>("linear_100_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = tensor<string, []>("input_603_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_11_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141712640)))];
+            tensor<fp16, [1, 17, 512]> linear_101_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16, x = input_603_cast_fp16)[name = tensor<string, []>("linear_101_cast_fp16")];
+            tensor<fp16, []> var_2547_to_fp16 = const()[name = tensor<string, []>("op_2547_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_2548_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2547_to_fp16)[name = tensor<string, []>("op_2548_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2548_cast_fp16)[name = tensor<string, []>("input_609_cast_fp16")];
+            tensor<int32, [1]> key_23_axes_0 = const()[name = tensor<string, []>("key_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143809856)))];
+            tensor<fp16, [512]> encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143810944)))];
+            tensor<fp16, [1, 17, 512]> key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = tensor<string, []>("key_23_cast_fp16")];
+            tensor<bool, []> input_611_interleave_0 = const()[name = tensor<string, []>("input_611_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 87, 512]> input_611_cast_fp16 = concat(axis = var_64, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = tensor<string, []>("input_611_cast_fp16")];
+            tensor<int32, [3]> var_2570_begin_0 = const()[name = tensor<string, []>("op_2570_begin_0"), val = tensor<int32, [3]>([0, 17, 0])];
+            tensor<int32, [3]> var_2570_end_0 = const()[name = tensor<string, []>("op_2570_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_2570_end_mask_0 = const()[name = tensor<string, []>("op_2570_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 53, 512]> var_2570_cast_fp16 = slice_by_index(begin = var_2570_begin_0, end = var_2570_end_0, end_mask = var_2570_end_mask_0, x = cache_45_cast_fp16)[name = tensor<string, []>("op_2570_cast_fp16")];
+            tensor<bool, []> var_2576_interleave_0 = const()[name = tensor<string, []>("op_2576_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_2576_cast_fp16 = concat(axis = var_64, interleave = var_2576_interleave_0, values = (var_2570_cast_fp16, key_23_cast_fp16))[name = tensor<string, []>("op_2576_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_11_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143812032)))];
+            tensor<fp16, [1, 17, 512]> linear_102_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16, x = key_23_cast_fp16)[name = tensor<string, []>("linear_102_cast_fp16")];
+            tensor<int32, [4]> var_2580 = const()[name = tensor<string, []>("op_2580"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 17, 8, 64]> q_67_cast_fp16 = reshape(shape = var_2580, x = linear_102_cast_fp16)[name = tensor<string, []>("q_67_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_11_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144336384)))];
+            tensor<fp16, [1, 87, 512]> linear_103_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16, x = input_611_cast_fp16)[name = tensor<string, []>("linear_103_cast_fp16")];
+            tensor<int32, [4]> var_2584 = const()[name = tensor<string, []>("op_2584"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> k_45_cast_fp16 = reshape(shape = var_2584, x = linear_103_cast_fp16)[name = tensor<string, []>("k_45_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_11_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144860736)))];
+            tensor<fp16, [1, 87, 512]> linear_104_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16, x = input_611_cast_fp16)[name = tensor<string, []>("linear_104_cast_fp16")];
+            tensor<int32, [4]> var_2588 = const()[name = tensor<string, []>("op_2588"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> v_23_cast_fp16 = reshape(shape = var_2588, x = linear_104_cast_fp16)[name = tensor<string, []>("v_23_cast_fp16")];
+            tensor<int32, [4]> value_25_perm_0 = const()[name = tensor<string, []>("value_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145385088)))];
+            tensor<fp16, [1, 17, 8, 64]> var_2600_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2600_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145386176)))];
+            tensor<fp16, [1, 17, 8, 64]> var_2602_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2602_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_23_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_293_transpose_x_0 = const()[name = tensor<string, []>("x_293_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_293_transpose_y_0 = const()[name = tensor<string, []>("x_293_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 173]> var_2604_to_fp16 = const()[name = tensor<string, []>("op_2604_to_fp16"), val = tensor<fp16, [1, 8, 64, 173]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145387264)))];
+            tensor<fp16, [1, 8, 17, 64]> q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2602_cast_fp16)[name = tensor<string, []>("transpose_138")];
+            tensor<fp16, [1, 8, 17, 173]> x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = var_2604_to_fp16)[name = tensor<string, []>("x_293_cast_fp16")];
+            tensor<int32, [8]> x_295_pad_0 = const()[name = tensor<string, []>("x_295_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_295_mode_0 = const()[name = tensor<string, []>("x_295_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_166_to_fp16 = const()[name = tensor<string, []>("const_166_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 17, 174]> x_295_cast_fp16 = pad(constant_val = const_166_to_fp16, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293_cast_fp16)[name = tensor<string, []>("x_295_cast_fp16")];
+            tensor<int32, [4]> var_2612 = const()[name = tensor<string, []>("op_2612"), val = tensor<int32, [4]>([1, 8, -1, 17])];
+            tensor<fp16, [1, 8, 174, 17]> x_297_cast_fp16 = reshape(shape = var_2612, x = x_295_cast_fp16)[name = tensor<string, []>("x_297_cast_fp16")];
+            tensor<int32, [4]> var_2616_begin_0 = const()[name = tensor<string, []>("op_2616_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2616_end_0 = const()[name = tensor<string, []>("op_2616_end_0"), val = tensor<int32, [4]>([1, 8, 174, 17])];
+            tensor<bool, [4]> var_2616_end_mask_0 = const()[name = tensor<string, []>("op_2616_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 173, 17]> var_2616_cast_fp16 = slice_by_index(begin = var_2616_begin_0, end = var_2616_end_0, end_mask = var_2616_end_mask_0, x = x_297_cast_fp16)[name = tensor<string, []>("op_2616_cast_fp16")];
+            tensor<int32, [4]> var_2617 = const()[name = tensor<string, []>("op_2617"), val = tensor<int32, [4]>([1, 8, 17, 173])];
+            tensor<fp16, [1, 8, 17, 173]> matrix_bd_45_cast_fp16 = reshape(shape = var_2617, x = var_2616_cast_fp16)[name = tensor<string, []>("matrix_bd_45_cast_fp16")];
+            tensor<bool, []> matrix_ac_23_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_23_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_23_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_23_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_73_perm_0 = const()[name = tensor<string, []>("transpose_73_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_74_perm_0 = const()[name = tensor<string, []>("transpose_74_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 87]> transpose_74 = transpose(perm = transpose_74_perm_0, x = k_45_cast_fp16)[name = tensor<string, []>("transpose_136")];
+            tensor<fp16, [1, 8, 17, 64]> transpose_73 = transpose(perm = transpose_73_perm_0, x = var_2600_cast_fp16)[name = tensor<string, []>("transpose_137")];
+            tensor<fp16, [1, 8, 17, 87]> matrix_ac_23_cast_fp16 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_73, y = transpose_74)[name = tensor<string, []>("matrix_ac_23_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_47_begin_0 = const()[name = tensor<string, []>("matrix_bd_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_47_end_0 = const()[name = tensor<string, []>("matrix_bd_47_end_0"), val = tensor<int32, [4]>([1, 8, 17, 87])];
+            tensor<bool, [4]> matrix_bd_47_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 17, 87]> matrix_bd_47_cast_fp16 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45_cast_fp16)[name = tensor<string, []>("matrix_bd_47_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_2626_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = tensor<string, []>("op_2626_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_45_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_45_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 17, 87]> _inversed_scores_45_cast_fp16 = mul(x = var_2626_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_45_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> scores_47_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_47_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_2632_cast_fp16 = softmax(axis = var_62, x = scores_47_cast_fp16)[name = tensor<string, []>("op_2632_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> input_613_cast_fp16 = select(a = var_40_to_fp16, b = var_2632_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_613_cast_fp16")];
+            tensor<bool, []> x_299_transpose_x_0 = const()[name = tensor<string, []>("x_299_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_299_transpose_y_0 = const()[name = tensor<string, []>("x_299_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 87, 64]> value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_23_cast_fp16)[name = tensor<string, []>("transpose_139")];
+            tensor<fp16, [1, 8, 17, 64]> x_299_cast_fp16 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613_cast_fp16, y = value_25_cast_fp16)[name = tensor<string, []>("x_299_cast_fp16")];
+            tensor<int32, [4]> var_2636_perm_0 = const()[name = tensor<string, []>("op_2636_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2637 = const()[name = tensor<string, []>("op_2637"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 17, 8, 64]> var_2636_cast_fp16 = transpose(perm = var_2636_perm_0, x = x_299_cast_fp16)[name = tensor<string, []>("transpose_135")];
+            tensor<fp16, [1, 17, 512]> input_615_cast_fp16 = reshape(shape = var_2637, x = var_2636_cast_fp16)[name = tensor<string, []>("input_615_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_11_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145564480)))];
+            tensor<fp16, [1, 17, 512]> linear_106_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16, x = input_615_cast_fp16)[name = tensor<string, []>("linear_106_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = tensor<string, []>("input_619_cast_fp16")];
+            tensor<int32, [1]> x_303_axes_0 = const()[name = tensor<string, []>("x_303_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146088832)))];
+            tensor<fp16, [512]> encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146089920)))];
+            tensor<fp16, [1, 17, 512]> x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = tensor<string, []>("x_303_cast_fp16")];
+            tensor<int32, [3]> input_621_perm_0 = const()[name = tensor<string, []>("input_621_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_623_pad_type_0 = const()[name = tensor<string, []>("input_623_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_623_strides_0 = const()[name = tensor<string, []>("input_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_623_pad_0 = const()[name = tensor<string, []>("input_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_623_dilations_0 = const()[name = tensor<string, []>("input_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_623_groups_0 = const()[name = tensor<string, []>("input_623_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_11_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146091008)))];
+            tensor<fp16, [1, 512, 17]> input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = tensor<string, []>("transpose_134")];
+            tensor<fp16, [1, 1024, 17]> input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16, x = input_621_cast_fp16)[name = tensor<string, []>("input_623_cast_fp16")];
+            tensor<int32, []> x_305_split_num_splits_0 = const()[name = tensor<string, []>("x_305_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_305_split_axis_0 = const()[name = tensor<string, []>("x_305_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 17]> x_305_split_cast_fp16_0, tensor<fp16, [1, 512, 17]> x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623_cast_fp16)[name = tensor<string, []>("x_305_split_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = tensor<string, []>("x_305_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_305_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> input_625_cast_fp16 = select(a = var_40_to_fp16, b = x_305_cast_fp16, cond = var_418)[name = tensor<string, []>("input_625_cast_fp16")];
+            tensor<bool, []> new_x_47_interleave_0 = const()[name = tensor<string, []>("new_x_47_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 25]> new_x_47_cast_fp16 = concat(axis = var_62, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = tensor<string, []>("new_x_47_cast_fp16")];
+            tensor<int32, [3]> var_2675_begin_0 = const()[name = tensor<string, []>("op_2675_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
+            tensor<int32, [3]> var_2675_end_0 = const()[name = tensor<string, []>("op_2675_end_0"), val = tensor<int32, [3]>([1, 512, 25])];
+            tensor<bool, [3]> var_2675_end_mask_0 = const()[name = tensor<string, []>("op_2675_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_2675_cast_fp16 = slice_by_index(begin = var_2675_begin_0, end = var_2675_end_0, end_mask = var_2675_end_mask_0, x = new_x_47_cast_fp16)[name = tensor<string, []>("op_2675_cast_fp16")];
+            tensor<string, []> x_307_pad_type_0 = const()[name = tensor<string, []>("x_307_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_307_groups_0 = const()[name = tensor<string, []>("x_307_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_307_strides_0 = const()[name = tensor<string, []>("x_307_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_307_pad_0 = const()[name = tensor<string, []>("x_307_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_307_dilations_0 = const()[name = tensor<string, []>("x_307_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_11_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147139648)))];
+            tensor<fp16, [1, 512, 17]> x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16, x = new_x_47_cast_fp16)[name = tensor<string, []>("x_307_cast_fp16")];
+            tensor<int32, [3]> input_627_perm_0 = const()[name = tensor<string, []>("input_627_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_309_axes_0 = const()[name = tensor<string, []>("x_309_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147148928)))];
+            tensor<fp16, [512]> encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147150016)))];
+            tensor<fp16, [1, 17, 512]> input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = tensor<string, []>("transpose_133")];
+            tensor<fp16, [1, 17, 512]> x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = tensor<string, []>("x_309_cast_fp16")];
+            tensor<int32, [3]> input_629_perm_0 = const()[name = tensor<string, []>("input_629_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 17]> input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = tensor<string, []>("transpose_132")];
+            tensor<fp16, [1, 512, 17]> input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = tensor<string, []>("input_631_cast_fp16")];
+            tensor<string, []> x_311_pad_type_0 = const()[name = tensor<string, []>("x_311_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_311_strides_0 = const()[name = tensor<string, []>("x_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_311_pad_0 = const()[name = tensor<string, []>("x_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_311_dilations_0 = const()[name = tensor<string, []>("x_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_311_groups_0 = const()[name = tensor<string, []>("x_311_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_11_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147151104)))];
+            tensor<fp16, [1, 512, 17]> x_311_cast_fp16 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16, x = input_631_cast_fp16)[name = tensor<string, []>("x_311_cast_fp16")];
+            tensor<int32, [3]> input_633_perm_0 = const()[name = tensor<string, []>("input_633_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 17, 512]> input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = tensor<string, []>("transpose_131")];
+            tensor<fp16, [1, 17, 512]> input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = tensor<string, []>("input_635_cast_fp16")];
+            tensor<int32, [1]> input_637_axes_0 = const()[name = tensor<string, []>("input_637_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147675456)))];
+            tensor<fp16, [512]> encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147676544)))];
+            tensor<fp16, [1, 17, 512]> input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = tensor<string, []>("input_637_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_11_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147677632)))];
+            tensor<fp16, [1, 17, 2048]> linear_107_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16, x = input_637_cast_fp16)[name = tensor<string, []>("linear_107_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = tensor<string, []>("input_641_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_11_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149774848)))];
+            tensor<fp16, [1, 17, 512]> linear_108_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16, x = input_641_cast_fp16)[name = tensor<string, []>("linear_108_cast_fp16")];
+            tensor<fp16, []> var_2716_to_fp16 = const()[name = tensor<string, []>("op_2716_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_2717_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2716_to_fp16)[name = tensor<string, []>("op_2717_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2717_cast_fp16)[name = tensor<string, []>("input_647_cast_fp16")];
+            tensor<int32, [1]> input_649_axes_0 = const()[name = tensor<string, []>("input_649_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_11_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151872064)))];
+            tensor<fp16, [512]> encoder_layers_11_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151873152)))];
+            tensor<fp16, [1, 17, 512]> input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = tensor<string, []>("input_649_cast_fp16")];
+            tensor<int32, [4]> cache_49_begin_0 = const()[name = tensor<string, []>("cache_49_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_49_end_0 = const()[name = tensor<string, []>("cache_49_end_0"), val = tensor<int32, [4]>([13, 1, 70, 512])];
+            tensor<bool, [4]> cache_49_end_mask_0 = const()[name = tensor<string, []>("cache_49_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_49_squeeze_mask_0 = const()[name = tensor<string, []>("cache_49_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_49_cast_fp16 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_49_cast_fp16")];
+            tensor<int32, [4]> cache_51_begin_0 = const()[name = tensor<string, []>("cache_51_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_51_end_0 = const()[name = tensor<string, []>("cache_51_end_0"), val = tensor<int32, [4]>([13, 1, 512, 8])];
+            tensor<bool, [4]> cache_51_end_mask_0 = const()[name = tensor<string, []>("cache_51_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_51_squeeze_mask_0 = const()[name = tensor<string, []>("cache_51_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_51_cast_fp16 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_51_cast_fp16")];
+            tensor<int32, [1]> input_651_axes_0 = const()[name = tensor<string, []>("input_651_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151874240)))];
+            tensor<fp16, [512]> encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151875328)))];
+            tensor<fp16, [1, 17, 512]> input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = tensor<string, []>("input_651_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_12_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151876416)))];
+            tensor<fp16, [1, 17, 2048]> linear_109_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16, x = input_651_cast_fp16)[name = tensor<string, []>("linear_109_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = tensor<string, []>("input_655_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_12_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153973632)))];
+            tensor<fp16, [1, 17, 512]> linear_110_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16, x = input_655_cast_fp16)[name = tensor<string, []>("linear_110_cast_fp16")];
+            tensor<fp16, []> var_2751_to_fp16 = const()[name = tensor<string, []>("op_2751_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_2752_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_2751_to_fp16)[name = tensor<string, []>("op_2752_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_2752_cast_fp16)[name = tensor<string, []>("input_661_cast_fp16")];
+            tensor<int32, [1]> key_25_axes_0 = const()[name = tensor<string, []>("key_25_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156070848)))];
+            tensor<fp16, [512]> encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156071936)))];
+            tensor<fp16, [1, 17, 512]> key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = tensor<string, []>("key_25_cast_fp16")];
+            tensor<bool, []> input_663_interleave_0 = const()[name = tensor<string, []>("input_663_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 87, 512]> input_663_cast_fp16 = concat(axis = var_64, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = tensor<string, []>("input_663_cast_fp16")];
+            tensor<int32, [3]> var_2774_begin_0 = const()[name = tensor<string, []>("op_2774_begin_0"), val = tensor<int32, [3]>([0, 17, 0])];
+            tensor<int32, [3]> var_2774_end_0 = const()[name = tensor<string, []>("op_2774_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_2774_end_mask_0 = const()[name = tensor<string, []>("op_2774_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 53, 512]> var_2774_cast_fp16 = slice_by_index(begin = var_2774_begin_0, end = var_2774_end_0, end_mask = var_2774_end_mask_0, x = cache_49_cast_fp16)[name = tensor<string, []>("op_2774_cast_fp16")];
+            tensor<bool, []> var_2780_interleave_0 = const()[name = tensor<string, []>("op_2780_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_2780_cast_fp16 = concat(axis = var_64, interleave = var_2780_interleave_0, values = (var_2774_cast_fp16, key_25_cast_fp16))[name = tensor<string, []>("op_2780_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_12_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156073024)))];
+            tensor<fp16, [1, 17, 512]> linear_111_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16, x = key_25_cast_fp16)[name = tensor<string, []>("linear_111_cast_fp16")];
+            tensor<int32, [4]> var_2784 = const()[name = tensor<string, []>("op_2784"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 17, 8, 64]> q_73_cast_fp16 = reshape(shape = var_2784, x = linear_111_cast_fp16)[name = tensor<string, []>("q_73_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_12_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156597376)))];
+            tensor<fp16, [1, 87, 512]> linear_112_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16, x = input_663_cast_fp16)[name = tensor<string, []>("linear_112_cast_fp16")];
+            tensor<int32, [4]> var_2788 = const()[name = tensor<string, []>("op_2788"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> k_49_cast_fp16 = reshape(shape = var_2788, x = linear_112_cast_fp16)[name = tensor<string, []>("k_49_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_12_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157121728)))];
+            tensor<fp16, [1, 87, 512]> linear_113_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16, x = input_663_cast_fp16)[name = tensor<string, []>("linear_113_cast_fp16")];
+            tensor<int32, [4]> var_2792 = const()[name = tensor<string, []>("op_2792"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> v_25_cast_fp16 = reshape(shape = var_2792, x = linear_113_cast_fp16)[name = tensor<string, []>("v_25_cast_fp16")];
+            tensor<int32, [4]> value_27_perm_0 = const()[name = tensor<string, []>("value_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157646080)))];
+            tensor<fp16, [1, 17, 8, 64]> var_2804_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2804_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157647168)))];
+            tensor<fp16, [1, 17, 8, 64]> var_2806_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2806_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_25_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_319_transpose_x_0 = const()[name = tensor<string, []>("x_319_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_319_transpose_y_0 = const()[name = tensor<string, []>("x_319_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 173]> var_2808_to_fp16 = const()[name = tensor<string, []>("op_2808_to_fp16"), val = tensor<fp16, [1, 8, 64, 173]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157648256)))];
+            tensor<fp16, [1, 8, 17, 64]> q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_2806_cast_fp16)[name = tensor<string, []>("transpose_129")];
+            tensor<fp16, [1, 8, 17, 173]> x_319_cast_fp16 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = var_2808_to_fp16)[name = tensor<string, []>("x_319_cast_fp16")];
+            tensor<int32, [8]> x_321_pad_0 = const()[name = tensor<string, []>("x_321_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_321_mode_0 = const()[name = tensor<string, []>("x_321_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_179_to_fp16 = const()[name = tensor<string, []>("const_179_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 17, 174]> x_321_cast_fp16 = pad(constant_val = const_179_to_fp16, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319_cast_fp16)[name = tensor<string, []>("x_321_cast_fp16")];
+            tensor<int32, [4]> var_2816 = const()[name = tensor<string, []>("op_2816"), val = tensor<int32, [4]>([1, 8, -1, 17])];
+            tensor<fp16, [1, 8, 174, 17]> x_323_cast_fp16 = reshape(shape = var_2816, x = x_321_cast_fp16)[name = tensor<string, []>("x_323_cast_fp16")];
+            tensor<int32, [4]> var_2820_begin_0 = const()[name = tensor<string, []>("op_2820_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2820_end_0 = const()[name = tensor<string, []>("op_2820_end_0"), val = tensor<int32, [4]>([1, 8, 174, 17])];
+            tensor<bool, [4]> var_2820_end_mask_0 = const()[name = tensor<string, []>("op_2820_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 173, 17]> var_2820_cast_fp16 = slice_by_index(begin = var_2820_begin_0, end = var_2820_end_0, end_mask = var_2820_end_mask_0, x = x_323_cast_fp16)[name = tensor<string, []>("op_2820_cast_fp16")];
+            tensor<int32, [4]> var_2821 = const()[name = tensor<string, []>("op_2821"), val = tensor<int32, [4]>([1, 8, 17, 173])];
+            tensor<fp16, [1, 8, 17, 173]> matrix_bd_49_cast_fp16 = reshape(shape = var_2821, x = var_2820_cast_fp16)[name = tensor<string, []>("matrix_bd_49_cast_fp16")];
+            tensor<bool, []> matrix_ac_25_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_25_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_25_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_25_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_75_perm_0 = const()[name = tensor<string, []>("transpose_75_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_76_perm_0 = const()[name = tensor<string, []>("transpose_76_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 87]> transpose_76 = transpose(perm = transpose_76_perm_0, x = k_49_cast_fp16)[name = tensor<string, []>("transpose_127")];
+            tensor<fp16, [1, 8, 17, 64]> transpose_75 = transpose(perm = transpose_75_perm_0, x = var_2804_cast_fp16)[name = tensor<string, []>("transpose_128")];
+            tensor<fp16, [1, 8, 17, 87]> matrix_ac_25_cast_fp16 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_75, y = transpose_76)[name = tensor<string, []>("matrix_ac_25_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_51_begin_0 = const()[name = tensor<string, []>("matrix_bd_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_51_end_0 = const()[name = tensor<string, []>("matrix_bd_51_end_0"), val = tensor<int32, [4]>([1, 8, 17, 87])];
+            tensor<bool, [4]> matrix_bd_51_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 17, 87]> matrix_bd_51_cast_fp16 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49_cast_fp16)[name = tensor<string, []>("matrix_bd_51_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_2830_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = tensor<string, []>("op_2830_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_49_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_49_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 17, 87]> _inversed_scores_49_cast_fp16 = mul(x = var_2830_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_49_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> scores_51_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_51_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_2836_cast_fp16 = softmax(axis = var_62, x = scores_51_cast_fp16)[name = tensor<string, []>("op_2836_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> input_665_cast_fp16 = select(a = var_40_to_fp16, b = var_2836_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_665_cast_fp16")];
+            tensor<bool, []> x_325_transpose_x_0 = const()[name = tensor<string, []>("x_325_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_325_transpose_y_0 = const()[name = tensor<string, []>("x_325_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 87, 64]> value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_25_cast_fp16)[name = tensor<string, []>("transpose_130")];
+            tensor<fp16, [1, 8, 17, 64]> x_325_cast_fp16 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665_cast_fp16, y = value_27_cast_fp16)[name = tensor<string, []>("x_325_cast_fp16")];
+            tensor<int32, [4]> var_2840_perm_0 = const()[name = tensor<string, []>("op_2840_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2841 = const()[name = tensor<string, []>("op_2841"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 17, 8, 64]> var_2840_cast_fp16 = transpose(perm = var_2840_perm_0, x = x_325_cast_fp16)[name = tensor<string, []>("transpose_126")];
+            tensor<fp16, [1, 17, 512]> input_667_cast_fp16 = reshape(shape = var_2841, x = var_2840_cast_fp16)[name = tensor<string, []>("input_667_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_12_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157825472)))];
+            tensor<fp16, [1, 17, 512]> linear_115_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16, x = input_667_cast_fp16)[name = tensor<string, []>("linear_115_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = tensor<string, []>("input_671_cast_fp16")];
+            tensor<int32, [1]> x_329_axes_0 = const()[name = tensor<string, []>("x_329_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158349824)))];
+            tensor<fp16, [512]> encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158350912)))];
+            tensor<fp16, [1, 17, 512]> x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = tensor<string, []>("x_329_cast_fp16")];
+            tensor<int32, [3]> input_673_perm_0 = const()[name = tensor<string, []>("input_673_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_675_pad_type_0 = const()[name = tensor<string, []>("input_675_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_675_strides_0 = const()[name = tensor<string, []>("input_675_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_675_pad_0 = const()[name = tensor<string, []>("input_675_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_675_dilations_0 = const()[name = tensor<string, []>("input_675_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_675_groups_0 = const()[name = tensor<string, []>("input_675_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_12_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158352000)))];
+            tensor<fp16, [1, 512, 17]> input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = tensor<string, []>("transpose_125")];
+            tensor<fp16, [1, 1024, 17]> input_675_cast_fp16 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16, x = input_673_cast_fp16)[name = tensor<string, []>("input_675_cast_fp16")];
+            tensor<int32, []> x_331_split_num_splits_0 = const()[name = tensor<string, []>("x_331_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_331_split_axis_0 = const()[name = tensor<string, []>("x_331_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 17]> x_331_split_cast_fp16_0, tensor<fp16, [1, 512, 17]> x_331_split_cast_fp16_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675_cast_fp16)[name = tensor<string, []>("x_331_split_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = tensor<string, []>("x_331_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_331_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> input_677_cast_fp16 = select(a = var_40_to_fp16, b = x_331_cast_fp16, cond = var_418)[name = tensor<string, []>("input_677_cast_fp16")];
+            tensor<bool, []> new_x_51_interleave_0 = const()[name = tensor<string, []>("new_x_51_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 25]> new_x_51_cast_fp16 = concat(axis = var_62, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = tensor<string, []>("new_x_51_cast_fp16")];
+            tensor<int32, [3]> var_2879_begin_0 = const()[name = tensor<string, []>("op_2879_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
+            tensor<int32, [3]> var_2879_end_0 = const()[name = tensor<string, []>("op_2879_end_0"), val = tensor<int32, [3]>([1, 512, 25])];
+            tensor<bool, [3]> var_2879_end_mask_0 = const()[name = tensor<string, []>("op_2879_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_2879_cast_fp16 = slice_by_index(begin = var_2879_begin_0, end = var_2879_end_0, end_mask = var_2879_end_mask_0, x = new_x_51_cast_fp16)[name = tensor<string, []>("op_2879_cast_fp16")];
+            tensor<string, []> x_333_pad_type_0 = const()[name = tensor<string, []>("x_333_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_333_groups_0 = const()[name = tensor<string, []>("x_333_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_333_strides_0 = const()[name = tensor<string, []>("x_333_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_333_pad_0 = const()[name = tensor<string, []>("x_333_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_333_dilations_0 = const()[name = tensor<string, []>("x_333_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_12_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159400640)))];
+            tensor<fp16, [1, 512, 17]> x_333_cast_fp16 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16, x = new_x_51_cast_fp16)[name = tensor<string, []>("x_333_cast_fp16")];
+            tensor<int32, [3]> input_679_perm_0 = const()[name = tensor<string, []>("input_679_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_335_axes_0 = const()[name = tensor<string, []>("x_335_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159409920)))];
+            tensor<fp16, [512]> encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159411008)))];
+            tensor<fp16, [1, 17, 512]> input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = tensor<string, []>("transpose_124")];
+            tensor<fp16, [1, 17, 512]> x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = tensor<string, []>("x_335_cast_fp16")];
+            tensor<int32, [3]> input_681_perm_0 = const()[name = tensor<string, []>("input_681_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 17]> input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = tensor<string, []>("transpose_123")];
+            tensor<fp16, [1, 512, 17]> input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = tensor<string, []>("input_683_cast_fp16")];
+            tensor<string, []> x_337_pad_type_0 = const()[name = tensor<string, []>("x_337_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_337_strides_0 = const()[name = tensor<string, []>("x_337_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_337_pad_0 = const()[name = tensor<string, []>("x_337_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_337_dilations_0 = const()[name = tensor<string, []>("x_337_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_337_groups_0 = const()[name = tensor<string, []>("x_337_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_12_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159412096)))];
+            tensor<fp16, [1, 512, 17]> x_337_cast_fp16 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16, x = input_683_cast_fp16)[name = tensor<string, []>("x_337_cast_fp16")];
+            tensor<int32, [3]> input_685_perm_0 = const()[name = tensor<string, []>("input_685_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 17, 512]> input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = tensor<string, []>("transpose_122")];
+            tensor<fp16, [1, 17, 512]> input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = tensor<string, []>("input_687_cast_fp16")];
+            tensor<int32, [1]> input_689_axes_0 = const()[name = tensor<string, []>("input_689_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159936448)))];
+            tensor<fp16, [512]> encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159937536)))];
+            tensor<fp16, [1, 17, 512]> input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = tensor<string, []>("input_689_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_12_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159938624)))];
+            tensor<fp16, [1, 17, 2048]> linear_116_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16, x = input_689_cast_fp16)[name = tensor<string, []>("linear_116_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = tensor<string, []>("input_693_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_12_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162035840)))];
+            tensor<fp16, [1, 17, 512]> linear_117_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16, x = input_693_cast_fp16)[name = tensor<string, []>("linear_117_cast_fp16")];
+            tensor<fp16, []> var_2920_to_fp16 = const()[name = tensor<string, []>("op_2920_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_2921_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_2920_to_fp16)[name = tensor<string, []>("op_2921_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_2921_cast_fp16)[name = tensor<string, []>("input_699_cast_fp16")];
+            tensor<int32, [1]> input_701_axes_0 = const()[name = tensor<string, []>("input_701_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_12_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164133056)))];
+            tensor<fp16, [512]> encoder_layers_12_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164134144)))];
+            tensor<fp16, [1, 17, 512]> input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = tensor<string, []>("input_701_cast_fp16")];
+            tensor<int32, [4]> cache_53_begin_0 = const()[name = tensor<string, []>("cache_53_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_53_end_0 = const()[name = tensor<string, []>("cache_53_end_0"), val = tensor<int32, [4]>([14, 1, 70, 512])];
+            tensor<bool, [4]> cache_53_end_mask_0 = const()[name = tensor<string, []>("cache_53_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_53_squeeze_mask_0 = const()[name = tensor<string, []>("cache_53_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_53_cast_fp16 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_53_cast_fp16")];
+            tensor<int32, [4]> cache_55_begin_0 = const()[name = tensor<string, []>("cache_55_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_55_end_0 = const()[name = tensor<string, []>("cache_55_end_0"), val = tensor<int32, [4]>([14, 1, 512, 8])];
+            tensor<bool, [4]> cache_55_end_mask_0 = const()[name = tensor<string, []>("cache_55_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_55_squeeze_mask_0 = const()[name = tensor<string, []>("cache_55_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_55_cast_fp16 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_55_cast_fp16")];
+            tensor<int32, [1]> input_703_axes_0 = const()[name = tensor<string, []>("input_703_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164135232)))];
+            tensor<fp16, [512]> encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164136320)))];
+            tensor<fp16, [1, 17, 512]> input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = tensor<string, []>("input_703_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_13_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164137408)))];
+            tensor<fp16, [1, 17, 2048]> linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16, x = input_703_cast_fp16)[name = tensor<string, []>("linear_118_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = tensor<string, []>("input_707_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_13_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166234624)))];
+            tensor<fp16, [1, 17, 512]> linear_119_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16, x = input_707_cast_fp16)[name = tensor<string, []>("linear_119_cast_fp16")];
+            tensor<fp16, []> var_2955_to_fp16 = const()[name = tensor<string, []>("op_2955_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_2956_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_2955_to_fp16)[name = tensor<string, []>("op_2956_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_2956_cast_fp16)[name = tensor<string, []>("input_713_cast_fp16")];
+            tensor<int32, [1]> key_27_axes_0 = const()[name = tensor<string, []>("key_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168331840)))];
+            tensor<fp16, [512]> encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168332928)))];
+            tensor<fp16, [1, 17, 512]> key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = tensor<string, []>("key_27_cast_fp16")];
+            tensor<bool, []> input_715_interleave_0 = const()[name = tensor<string, []>("input_715_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 87, 512]> input_715_cast_fp16 = concat(axis = var_64, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = tensor<string, []>("input_715_cast_fp16")];
+            tensor<int32, [3]> var_2978_begin_0 = const()[name = tensor<string, []>("op_2978_begin_0"), val = tensor<int32, [3]>([0, 17, 0])];
+            tensor<int32, [3]> var_2978_end_0 = const()[name = tensor<string, []>("op_2978_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_2978_end_mask_0 = const()[name = tensor<string, []>("op_2978_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 53, 512]> var_2978_cast_fp16 = slice_by_index(begin = var_2978_begin_0, end = var_2978_end_0, end_mask = var_2978_end_mask_0, x = cache_53_cast_fp16)[name = tensor<string, []>("op_2978_cast_fp16")];
+            tensor<bool, []> var_2984_interleave_0 = const()[name = tensor<string, []>("op_2984_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_2984_cast_fp16 = concat(axis = var_64, interleave = var_2984_interleave_0, values = (var_2978_cast_fp16, key_27_cast_fp16))[name = tensor<string, []>("op_2984_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_13_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168334016)))];
+            tensor<fp16, [1, 17, 512]> linear_120_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16, x = key_27_cast_fp16)[name = tensor<string, []>("linear_120_cast_fp16")];
+            tensor<int32, [4]> var_2988 = const()[name = tensor<string, []>("op_2988"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 17, 8, 64]> q_79_cast_fp16 = reshape(shape = var_2988, x = linear_120_cast_fp16)[name = tensor<string, []>("q_79_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_13_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168858368)))];
+            tensor<fp16, [1, 87, 512]> linear_121_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16, x = input_715_cast_fp16)[name = tensor<string, []>("linear_121_cast_fp16")];
+            tensor<int32, [4]> var_2992 = const()[name = tensor<string, []>("op_2992"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> k_53_cast_fp16 = reshape(shape = var_2992, x = linear_121_cast_fp16)[name = tensor<string, []>("k_53_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_13_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169382720)))];
+            tensor<fp16, [1, 87, 512]> linear_122_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16, x = input_715_cast_fp16)[name = tensor<string, []>("linear_122_cast_fp16")];
+            tensor<int32, [4]> var_2996 = const()[name = tensor<string, []>("op_2996"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> v_27_cast_fp16 = reshape(shape = var_2996, x = linear_122_cast_fp16)[name = tensor<string, []>("v_27_cast_fp16")];
+            tensor<int32, [4]> value_29_perm_0 = const()[name = tensor<string, []>("value_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169907072)))];
+            tensor<fp16, [1, 17, 8, 64]> var_3008_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3008_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169908160)))];
+            tensor<fp16, [1, 17, 8, 64]> var_3010_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3010_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_27_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_345_transpose_x_0 = const()[name = tensor<string, []>("x_345_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_345_transpose_y_0 = const()[name = tensor<string, []>("x_345_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 173]> var_3012_to_fp16 = const()[name = tensor<string, []>("op_3012_to_fp16"), val = tensor<fp16, [1, 8, 64, 173]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169909248)))];
+            tensor<fp16, [1, 8, 17, 64]> q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3010_cast_fp16)[name = tensor<string, []>("transpose_120")];
+            tensor<fp16, [1, 8, 17, 173]> x_345_cast_fp16 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = var_3012_to_fp16)[name = tensor<string, []>("x_345_cast_fp16")];
+            tensor<int32, [8]> x_347_pad_0 = const()[name = tensor<string, []>("x_347_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_347_mode_0 = const()[name = tensor<string, []>("x_347_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_192_to_fp16 = const()[name = tensor<string, []>("const_192_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 17, 174]> x_347_cast_fp16 = pad(constant_val = const_192_to_fp16, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345_cast_fp16)[name = tensor<string, []>("x_347_cast_fp16")];
+            tensor<int32, [4]> var_3020 = const()[name = tensor<string, []>("op_3020"), val = tensor<int32, [4]>([1, 8, -1, 17])];
+            tensor<fp16, [1, 8, 174, 17]> x_349_cast_fp16 = reshape(shape = var_3020, x = x_347_cast_fp16)[name = tensor<string, []>("x_349_cast_fp16")];
+            tensor<int32, [4]> var_3024_begin_0 = const()[name = tensor<string, []>("op_3024_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3024_end_0 = const()[name = tensor<string, []>("op_3024_end_0"), val = tensor<int32, [4]>([1, 8, 174, 17])];
+            tensor<bool, [4]> var_3024_end_mask_0 = const()[name = tensor<string, []>("op_3024_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 173, 17]> var_3024_cast_fp16 = slice_by_index(begin = var_3024_begin_0, end = var_3024_end_0, end_mask = var_3024_end_mask_0, x = x_349_cast_fp16)[name = tensor<string, []>("op_3024_cast_fp16")];
+            tensor<int32, [4]> var_3025 = const()[name = tensor<string, []>("op_3025"), val = tensor<int32, [4]>([1, 8, 17, 173])];
+            tensor<fp16, [1, 8, 17, 173]> matrix_bd_53_cast_fp16 = reshape(shape = var_3025, x = var_3024_cast_fp16)[name = tensor<string, []>("matrix_bd_53_cast_fp16")];
+            tensor<bool, []> matrix_ac_27_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_27_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_27_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_27_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_77_perm_0 = const()[name = tensor<string, []>("transpose_77_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_78_perm_0 = const()[name = tensor<string, []>("transpose_78_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 87]> transpose_78 = transpose(perm = transpose_78_perm_0, x = k_53_cast_fp16)[name = tensor<string, []>("transpose_118")];
+            tensor<fp16, [1, 8, 17, 64]> transpose_77 = transpose(perm = transpose_77_perm_0, x = var_3008_cast_fp16)[name = tensor<string, []>("transpose_119")];
+            tensor<fp16, [1, 8, 17, 87]> matrix_ac_27_cast_fp16 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_77, y = transpose_78)[name = tensor<string, []>("matrix_ac_27_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_55_begin_0 = const()[name = tensor<string, []>("matrix_bd_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_55_end_0 = const()[name = tensor<string, []>("matrix_bd_55_end_0"), val = tensor<int32, [4]>([1, 8, 17, 87])];
+            tensor<bool, [4]> matrix_bd_55_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 17, 87]> matrix_bd_55_cast_fp16 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53_cast_fp16)[name = tensor<string, []>("matrix_bd_55_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_3034_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = tensor<string, []>("op_3034_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_53_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_53_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 17, 87]> _inversed_scores_53_cast_fp16 = mul(x = var_3034_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_53_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> scores_55_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_55_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_3040_cast_fp16 = softmax(axis = var_62, x = scores_55_cast_fp16)[name = tensor<string, []>("op_3040_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> input_717_cast_fp16 = select(a = var_40_to_fp16, b = var_3040_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_717_cast_fp16")];
+            tensor<bool, []> x_351_transpose_x_0 = const()[name = tensor<string, []>("x_351_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_351_transpose_y_0 = const()[name = tensor<string, []>("x_351_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 87, 64]> value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_27_cast_fp16)[name = tensor<string, []>("transpose_121")];
+            tensor<fp16, [1, 8, 17, 64]> x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717_cast_fp16, y = value_29_cast_fp16)[name = tensor<string, []>("x_351_cast_fp16")];
+            tensor<int32, [4]> var_3044_perm_0 = const()[name = tensor<string, []>("op_3044_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3045 = const()[name = tensor<string, []>("op_3045"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 17, 8, 64]> var_3044_cast_fp16 = transpose(perm = var_3044_perm_0, x = x_351_cast_fp16)[name = tensor<string, []>("transpose_117")];
+            tensor<fp16, [1, 17, 512]> input_719_cast_fp16 = reshape(shape = var_3045, x = var_3044_cast_fp16)[name = tensor<string, []>("input_719_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_13_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170086464)))];
+            tensor<fp16, [1, 17, 512]> linear_124_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16, x = input_719_cast_fp16)[name = tensor<string, []>("linear_124_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = tensor<string, []>("input_723_cast_fp16")];
+            tensor<int32, [1]> x_355_axes_0 = const()[name = tensor<string, []>("x_355_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170610816)))];
+            tensor<fp16, [512]> encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170611904)))];
+            tensor<fp16, [1, 17, 512]> x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = tensor<string, []>("x_355_cast_fp16")];
+            tensor<int32, [3]> input_725_perm_0 = const()[name = tensor<string, []>("input_725_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_727_pad_type_0 = const()[name = tensor<string, []>("input_727_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_727_strides_0 = const()[name = tensor<string, []>("input_727_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_727_pad_0 = const()[name = tensor<string, []>("input_727_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_727_dilations_0 = const()[name = tensor<string, []>("input_727_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_727_groups_0 = const()[name = tensor<string, []>("input_727_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_13_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170612992)))];
+            tensor<fp16, [1, 512, 17]> input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = tensor<string, []>("transpose_116")];
+            tensor<fp16, [1, 1024, 17]> input_727_cast_fp16 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16, x = input_725_cast_fp16)[name = tensor<string, []>("input_727_cast_fp16")];
+            tensor<int32, []> x_357_split_num_splits_0 = const()[name = tensor<string, []>("x_357_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_357_split_axis_0 = const()[name = tensor<string, []>("x_357_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 17]> x_357_split_cast_fp16_0, tensor<fp16, [1, 512, 17]> x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727_cast_fp16)[name = tensor<string, []>("x_357_split_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = tensor<string, []>("x_357_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_357_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> input_729_cast_fp16 = select(a = var_40_to_fp16, b = x_357_cast_fp16, cond = var_418)[name = tensor<string, []>("input_729_cast_fp16")];
+            tensor<bool, []> new_x_55_interleave_0 = const()[name = tensor<string, []>("new_x_55_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 25]> new_x_55_cast_fp16 = concat(axis = var_62, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = tensor<string, []>("new_x_55_cast_fp16")];
+            tensor<int32, [3]> var_3083_begin_0 = const()[name = tensor<string, []>("op_3083_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
+            tensor<int32, [3]> var_3083_end_0 = const()[name = tensor<string, []>("op_3083_end_0"), val = tensor<int32, [3]>([1, 512, 25])];
+            tensor<bool, [3]> var_3083_end_mask_0 = const()[name = tensor<string, []>("op_3083_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_3083_cast_fp16 = slice_by_index(begin = var_3083_begin_0, end = var_3083_end_0, end_mask = var_3083_end_mask_0, x = new_x_55_cast_fp16)[name = tensor<string, []>("op_3083_cast_fp16")];
+            tensor<string, []> x_359_pad_type_0 = const()[name = tensor<string, []>("x_359_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_359_groups_0 = const()[name = tensor<string, []>("x_359_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_359_strides_0 = const()[name = tensor<string, []>("x_359_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_359_pad_0 = const()[name = tensor<string, []>("x_359_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_359_dilations_0 = const()[name = tensor<string, []>("x_359_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_13_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171661632)))];
+            tensor<fp16, [1, 512, 17]> x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16, x = new_x_55_cast_fp16)[name = tensor<string, []>("x_359_cast_fp16")];
+            tensor<int32, [3]> input_731_perm_0 = const()[name = tensor<string, []>("input_731_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_361_axes_0 = const()[name = tensor<string, []>("x_361_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171670912)))];
+            tensor<fp16, [512]> encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171672000)))];
+            tensor<fp16, [1, 17, 512]> input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = tensor<string, []>("transpose_115")];
+            tensor<fp16, [1, 17, 512]> x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = tensor<string, []>("x_361_cast_fp16")];
+            tensor<int32, [3]> input_733_perm_0 = const()[name = tensor<string, []>("input_733_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 17]> input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = tensor<string, []>("transpose_114")];
+            tensor<fp16, [1, 512, 17]> input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = tensor<string, []>("input_735_cast_fp16")];
+            tensor<string, []> x_363_pad_type_0 = const()[name = tensor<string, []>("x_363_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_363_strides_0 = const()[name = tensor<string, []>("x_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_363_pad_0 = const()[name = tensor<string, []>("x_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_363_dilations_0 = const()[name = tensor<string, []>("x_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_363_groups_0 = const()[name = tensor<string, []>("x_363_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_13_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171673088)))];
+            tensor<fp16, [1, 512, 17]> x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16, x = input_735_cast_fp16)[name = tensor<string, []>("x_363_cast_fp16")];
+            tensor<int32, [3]> input_737_perm_0 = const()[name = tensor<string, []>("input_737_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 17, 512]> input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = tensor<string, []>("transpose_113")];
+            tensor<fp16, [1, 17, 512]> input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = tensor<string, []>("input_739_cast_fp16")];
+            tensor<int32, [1]> input_741_axes_0 = const()[name = tensor<string, []>("input_741_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172197440)))];
+            tensor<fp16, [512]> encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172198528)))];
+            tensor<fp16, [1, 17, 512]> input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = tensor<string, []>("input_741_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_13_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172199616)))];
+            tensor<fp16, [1, 17, 2048]> linear_125_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16, x = input_741_cast_fp16)[name = tensor<string, []>("linear_125_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = tensor<string, []>("input_745_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_13_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174296832)))];
+            tensor<fp16, [1, 17, 512]> linear_126_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16, x = input_745_cast_fp16)[name = tensor<string, []>("linear_126_cast_fp16")];
+            tensor<fp16, []> var_3124_to_fp16 = const()[name = tensor<string, []>("op_3124_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_3125_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3124_to_fp16)[name = tensor<string, []>("op_3125_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3125_cast_fp16)[name = tensor<string, []>("input_751_cast_fp16")];
+            tensor<int32, [1]> input_753_axes_0 = const()[name = tensor<string, []>("input_753_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_13_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176394048)))];
+            tensor<fp16, [512]> encoder_layers_13_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176395136)))];
+            tensor<fp16, [1, 17, 512]> input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = tensor<string, []>("input_753_cast_fp16")];
+            tensor<int32, [4]> cache_57_begin_0 = const()[name = tensor<string, []>("cache_57_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_57_end_0 = const()[name = tensor<string, []>("cache_57_end_0"), val = tensor<int32, [4]>([15, 1, 70, 512])];
+            tensor<bool, [4]> cache_57_end_mask_0 = const()[name = tensor<string, []>("cache_57_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_57_squeeze_mask_0 = const()[name = tensor<string, []>("cache_57_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_57_cast_fp16 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_57_cast_fp16")];
+            tensor<int32, [4]> cache_59_begin_0 = const()[name = tensor<string, []>("cache_59_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_59_end_0 = const()[name = tensor<string, []>("cache_59_end_0"), val = tensor<int32, [4]>([15, 1, 512, 8])];
+            tensor<bool, [4]> cache_59_end_mask_0 = const()[name = tensor<string, []>("cache_59_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_59_squeeze_mask_0 = const()[name = tensor<string, []>("cache_59_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_59_cast_fp16 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_59_cast_fp16")];
+            tensor<int32, [1]> input_755_axes_0 = const()[name = tensor<string, []>("input_755_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176396224)))];
+            tensor<fp16, [512]> encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176397312)))];
+            tensor<fp16, [1, 17, 512]> input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = tensor<string, []>("input_755_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_14_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176398400)))];
+            tensor<fp16, [1, 17, 2048]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16, x = input_755_cast_fp16)[name = tensor<string, []>("linear_127_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = tensor<string, []>("input_759_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_14_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178495616)))];
+            tensor<fp16, [1, 17, 512]> linear_128_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16, x = input_759_cast_fp16)[name = tensor<string, []>("linear_128_cast_fp16")];
+            tensor<fp16, []> var_3159_to_fp16 = const()[name = tensor<string, []>("op_3159_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_3160_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3159_to_fp16)[name = tensor<string, []>("op_3160_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3160_cast_fp16)[name = tensor<string, []>("input_765_cast_fp16")];
+            tensor<int32, [1]> key_29_axes_0 = const()[name = tensor<string, []>("key_29_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180592832)))];
+            tensor<fp16, [512]> encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180593920)))];
+            tensor<fp16, [1, 17, 512]> key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = tensor<string, []>("key_29_cast_fp16")];
+            tensor<bool, []> input_767_interleave_0 = const()[name = tensor<string, []>("input_767_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 87, 512]> input_767_cast_fp16 = concat(axis = var_64, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = tensor<string, []>("input_767_cast_fp16")];
+            tensor<int32, [3]> var_3182_begin_0 = const()[name = tensor<string, []>("op_3182_begin_0"), val = tensor<int32, [3]>([0, 17, 0])];
+            tensor<int32, [3]> var_3182_end_0 = const()[name = tensor<string, []>("op_3182_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_3182_end_mask_0 = const()[name = tensor<string, []>("op_3182_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 53, 512]> var_3182_cast_fp16 = slice_by_index(begin = var_3182_begin_0, end = var_3182_end_0, end_mask = var_3182_end_mask_0, x = cache_57_cast_fp16)[name = tensor<string, []>("op_3182_cast_fp16")];
+            tensor<bool, []> var_3188_interleave_0 = const()[name = tensor<string, []>("op_3188_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_3188_cast_fp16 = concat(axis = var_64, interleave = var_3188_interleave_0, values = (var_3182_cast_fp16, key_29_cast_fp16))[name = tensor<string, []>("op_3188_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_14_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180595008)))];
+            tensor<fp16, [1, 17, 512]> linear_129_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16, x = key_29_cast_fp16)[name = tensor<string, []>("linear_129_cast_fp16")];
+            tensor<int32, [4]> var_3192 = const()[name = tensor<string, []>("op_3192"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 17, 8, 64]> q_85_cast_fp16 = reshape(shape = var_3192, x = linear_129_cast_fp16)[name = tensor<string, []>("q_85_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_14_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181119360)))];
+            tensor<fp16, [1, 87, 512]> linear_130_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16, x = input_767_cast_fp16)[name = tensor<string, []>("linear_130_cast_fp16")];
+            tensor<int32, [4]> var_3196 = const()[name = tensor<string, []>("op_3196"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> k_57_cast_fp16 = reshape(shape = var_3196, x = linear_130_cast_fp16)[name = tensor<string, []>("k_57_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_14_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181643712)))];
+            tensor<fp16, [1, 87, 512]> linear_131_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16, x = input_767_cast_fp16)[name = tensor<string, []>("linear_131_cast_fp16")];
+            tensor<int32, [4]> var_3200 = const()[name = tensor<string, []>("op_3200"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> v_29_cast_fp16 = reshape(shape = var_3200, x = linear_131_cast_fp16)[name = tensor<string, []>("v_29_cast_fp16")];
+            tensor<int32, [4]> value_31_perm_0 = const()[name = tensor<string, []>("value_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182168064)))];
+            tensor<fp16, [1, 17, 8, 64]> var_3212_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3212_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182169152)))];
+            tensor<fp16, [1, 17, 8, 64]> var_3214_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3214_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_29_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_371_transpose_x_0 = const()[name = tensor<string, []>("x_371_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_371_transpose_y_0 = const()[name = tensor<string, []>("x_371_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 173]> var_3216_to_fp16 = const()[name = tensor<string, []>("op_3216_to_fp16"), val = tensor<fp16, [1, 8, 64, 173]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182170240)))];
+            tensor<fp16, [1, 8, 17, 64]> q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3214_cast_fp16)[name = tensor<string, []>("transpose_111")];
+            tensor<fp16, [1, 8, 17, 173]> x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = var_3216_to_fp16)[name = tensor<string, []>("x_371_cast_fp16")];
+            tensor<int32, [8]> x_373_pad_0 = const()[name = tensor<string, []>("x_373_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_373_mode_0 = const()[name = tensor<string, []>("x_373_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_205_to_fp16 = const()[name = tensor<string, []>("const_205_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 17, 174]> x_373_cast_fp16 = pad(constant_val = const_205_to_fp16, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371_cast_fp16)[name = tensor<string, []>("x_373_cast_fp16")];
+            tensor<int32, [4]> var_3224 = const()[name = tensor<string, []>("op_3224"), val = tensor<int32, [4]>([1, 8, -1, 17])];
+            tensor<fp16, [1, 8, 174, 17]> x_375_cast_fp16 = reshape(shape = var_3224, x = x_373_cast_fp16)[name = tensor<string, []>("x_375_cast_fp16")];
+            tensor<int32, [4]> var_3228_begin_0 = const()[name = tensor<string, []>("op_3228_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3228_end_0 = const()[name = tensor<string, []>("op_3228_end_0"), val = tensor<int32, [4]>([1, 8, 174, 17])];
+            tensor<bool, [4]> var_3228_end_mask_0 = const()[name = tensor<string, []>("op_3228_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 173, 17]> var_3228_cast_fp16 = slice_by_index(begin = var_3228_begin_0, end = var_3228_end_0, end_mask = var_3228_end_mask_0, x = x_375_cast_fp16)[name = tensor<string, []>("op_3228_cast_fp16")];
+            tensor<int32, [4]> var_3229 = const()[name = tensor<string, []>("op_3229"), val = tensor<int32, [4]>([1, 8, 17, 173])];
+            tensor<fp16, [1, 8, 17, 173]> matrix_bd_57_cast_fp16 = reshape(shape = var_3229, x = var_3228_cast_fp16)[name = tensor<string, []>("matrix_bd_57_cast_fp16")];
+            tensor<bool, []> matrix_ac_29_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_29_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_29_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_29_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_79_perm_0 = const()[name = tensor<string, []>("transpose_79_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_80_perm_0 = const()[name = tensor<string, []>("transpose_80_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 87]> transpose_80 = transpose(perm = transpose_80_perm_0, x = k_57_cast_fp16)[name = tensor<string, []>("transpose_109")];
+            tensor<fp16, [1, 8, 17, 64]> transpose_79 = transpose(perm = transpose_79_perm_0, x = var_3212_cast_fp16)[name = tensor<string, []>("transpose_110")];
+            tensor<fp16, [1, 8, 17, 87]> matrix_ac_29_cast_fp16 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_79, y = transpose_80)[name = tensor<string, []>("matrix_ac_29_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_59_begin_0 = const()[name = tensor<string, []>("matrix_bd_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_59_end_0 = const()[name = tensor<string, []>("matrix_bd_59_end_0"), val = tensor<int32, [4]>([1, 8, 17, 87])];
+            tensor<bool, [4]> matrix_bd_59_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 17, 87]> matrix_bd_59_cast_fp16 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57_cast_fp16)[name = tensor<string, []>("matrix_bd_59_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_3238_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = tensor<string, []>("op_3238_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_57_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_57_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 17, 87]> _inversed_scores_57_cast_fp16 = mul(x = var_3238_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_57_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> scores_59_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_59_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_3244_cast_fp16 = softmax(axis = var_62, x = scores_59_cast_fp16)[name = tensor<string, []>("op_3244_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> input_769_cast_fp16 = select(a = var_40_to_fp16, b = var_3244_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_769_cast_fp16")];
+            tensor<bool, []> x_377_transpose_x_0 = const()[name = tensor<string, []>("x_377_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_377_transpose_y_0 = const()[name = tensor<string, []>("x_377_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 87, 64]> value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_29_cast_fp16)[name = tensor<string, []>("transpose_112")];
+            tensor<fp16, [1, 8, 17, 64]> x_377_cast_fp16 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769_cast_fp16, y = value_31_cast_fp16)[name = tensor<string, []>("x_377_cast_fp16")];
+            tensor<int32, [4]> var_3248_perm_0 = const()[name = tensor<string, []>("op_3248_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3249 = const()[name = tensor<string, []>("op_3249"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 17, 8, 64]> var_3248_cast_fp16 = transpose(perm = var_3248_perm_0, x = x_377_cast_fp16)[name = tensor<string, []>("transpose_108")];
+            tensor<fp16, [1, 17, 512]> input_771_cast_fp16 = reshape(shape = var_3249, x = var_3248_cast_fp16)[name = tensor<string, []>("input_771_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_14_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182347456)))];
+            tensor<fp16, [1, 17, 512]> linear_133_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16, x = input_771_cast_fp16)[name = tensor<string, []>("linear_133_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = tensor<string, []>("input_775_cast_fp16")];
+            tensor<int32, [1]> x_381_axes_0 = const()[name = tensor<string, []>("x_381_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182871808)))];
+            tensor<fp16, [512]> encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182872896)))];
+            tensor<fp16, [1, 17, 512]> x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = tensor<string, []>("x_381_cast_fp16")];
+            tensor<int32, [3]> input_777_perm_0 = const()[name = tensor<string, []>("input_777_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_779_pad_type_0 = const()[name = tensor<string, []>("input_779_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_779_strides_0 = const()[name = tensor<string, []>("input_779_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_779_pad_0 = const()[name = tensor<string, []>("input_779_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_779_dilations_0 = const()[name = tensor<string, []>("input_779_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_779_groups_0 = const()[name = tensor<string, []>("input_779_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_14_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182873984)))];
+            tensor<fp16, [1, 512, 17]> input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = tensor<string, []>("transpose_107")];
+            tensor<fp16, [1, 1024, 17]> input_779_cast_fp16 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16, x = input_777_cast_fp16)[name = tensor<string, []>("input_779_cast_fp16")];
+            tensor<int32, []> x_383_split_num_splits_0 = const()[name = tensor<string, []>("x_383_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_383_split_axis_0 = const()[name = tensor<string, []>("x_383_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 17]> x_383_split_cast_fp16_0, tensor<fp16, [1, 512, 17]> x_383_split_cast_fp16_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779_cast_fp16)[name = tensor<string, []>("x_383_split_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = tensor<string, []>("x_383_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_383_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> input_781_cast_fp16 = select(a = var_40_to_fp16, b = x_383_cast_fp16, cond = var_418)[name = tensor<string, []>("input_781_cast_fp16")];
+            tensor<bool, []> new_x_59_interleave_0 = const()[name = tensor<string, []>("new_x_59_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 25]> new_x_59_cast_fp16 = concat(axis = var_62, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = tensor<string, []>("new_x_59_cast_fp16")];
+            tensor<int32, [3]> var_3287_begin_0 = const()[name = tensor<string, []>("op_3287_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
+            tensor<int32, [3]> var_3287_end_0 = const()[name = tensor<string, []>("op_3287_end_0"), val = tensor<int32, [3]>([1, 512, 25])];
+            tensor<bool, [3]> var_3287_end_mask_0 = const()[name = tensor<string, []>("op_3287_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_3287_cast_fp16 = slice_by_index(begin = var_3287_begin_0, end = var_3287_end_0, end_mask = var_3287_end_mask_0, x = new_x_59_cast_fp16)[name = tensor<string, []>("op_3287_cast_fp16")];
+            tensor<string, []> x_385_pad_type_0 = const()[name = tensor<string, []>("x_385_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_385_groups_0 = const()[name = tensor<string, []>("x_385_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_385_strides_0 = const()[name = tensor<string, []>("x_385_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_385_pad_0 = const()[name = tensor<string, []>("x_385_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_385_dilations_0 = const()[name = tensor<string, []>("x_385_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_14_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183922624)))];
+            tensor<fp16, [1, 512, 17]> x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16, x = new_x_59_cast_fp16)[name = tensor<string, []>("x_385_cast_fp16")];
+            tensor<int32, [3]> input_783_perm_0 = const()[name = tensor<string, []>("input_783_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_387_axes_0 = const()[name = tensor<string, []>("x_387_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183931904)))];
+            tensor<fp16, [512]> encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183932992)))];
+            tensor<fp16, [1, 17, 512]> input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = tensor<string, []>("transpose_106")];
+            tensor<fp16, [1, 17, 512]> x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = tensor<string, []>("x_387_cast_fp16")];
+            tensor<int32, [3]> input_785_perm_0 = const()[name = tensor<string, []>("input_785_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 17]> input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = tensor<string, []>("transpose_105")];
+            tensor<fp16, [1, 512, 17]> input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = tensor<string, []>("input_787_cast_fp16")];
+            tensor<string, []> x_389_pad_type_0 = const()[name = tensor<string, []>("x_389_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_389_strides_0 = const()[name = tensor<string, []>("x_389_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_389_pad_0 = const()[name = tensor<string, []>("x_389_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_389_dilations_0 = const()[name = tensor<string, []>("x_389_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_389_groups_0 = const()[name = tensor<string, []>("x_389_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_14_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183934080)))];
+            tensor<fp16, [1, 512, 17]> x_389_cast_fp16 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16, x = input_787_cast_fp16)[name = tensor<string, []>("x_389_cast_fp16")];
+            tensor<int32, [3]> input_789_perm_0 = const()[name = tensor<string, []>("input_789_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 17, 512]> input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = tensor<string, []>("transpose_104")];
+            tensor<fp16, [1, 17, 512]> input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = tensor<string, []>("input_791_cast_fp16")];
+            tensor<int32, [1]> input_793_axes_0 = const()[name = tensor<string, []>("input_793_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184458432)))];
+            tensor<fp16, [512]> encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184459520)))];
+            tensor<fp16, [1, 17, 512]> input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = tensor<string, []>("input_793_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_14_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184460608)))];
+            tensor<fp16, [1, 17, 2048]> linear_134_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16, x = input_793_cast_fp16)[name = tensor<string, []>("linear_134_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = tensor<string, []>("input_797_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_14_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186557824)))];
+            tensor<fp16, [1, 17, 512]> linear_135_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16, x = input_797_cast_fp16)[name = tensor<string, []>("linear_135_cast_fp16")];
+            tensor<fp16, []> var_3328_to_fp16 = const()[name = tensor<string, []>("op_3328_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_3329_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3328_to_fp16)[name = tensor<string, []>("op_3329_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3329_cast_fp16)[name = tensor<string, []>("input_803_cast_fp16")];
+            tensor<int32, [1]> input_805_axes_0 = const()[name = tensor<string, []>("input_805_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_14_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188655040)))];
+            tensor<fp16, [512]> encoder_layers_14_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188656128)))];
+            tensor<fp16, [1, 17, 512]> input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = tensor<string, []>("input_805_cast_fp16")];
+            tensor<int32, [4]> cache_61_begin_0 = const()[name = tensor<string, []>("cache_61_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_61_end_0 = const()[name = tensor<string, []>("cache_61_end_0"), val = tensor<int32, [4]>([16, 1, 70, 512])];
+            tensor<bool, [4]> cache_61_end_mask_0 = const()[name = tensor<string, []>("cache_61_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_61_squeeze_mask_0 = const()[name = tensor<string, []>("cache_61_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_61_cast_fp16 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_61_cast_fp16")];
+            tensor<int32, [4]> cache_63_begin_0 = const()[name = tensor<string, []>("cache_63_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_63_end_0 = const()[name = tensor<string, []>("cache_63_end_0"), val = tensor<int32, [4]>([16, 1, 512, 8])];
+            tensor<bool, [4]> cache_63_end_mask_0 = const()[name = tensor<string, []>("cache_63_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_63_squeeze_mask_0 = const()[name = tensor<string, []>("cache_63_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_63_cast_fp16 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_63_cast_fp16")];
+            tensor<int32, [1]> input_807_axes_0 = const()[name = tensor<string, []>("input_807_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188657216)))];
+            tensor<fp16, [512]> encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188658304)))];
+            tensor<fp16, [1, 17, 512]> input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = tensor<string, []>("input_807_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_15_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188659392)))];
+            tensor<fp16, [1, 17, 2048]> linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16, x = input_807_cast_fp16)[name = tensor<string, []>("linear_136_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = tensor<string, []>("input_811_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_15_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190756608)))];
+            tensor<fp16, [1, 17, 512]> linear_137_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16, x = input_811_cast_fp16)[name = tensor<string, []>("linear_137_cast_fp16")];
+            tensor<fp16, []> var_3363_to_fp16 = const()[name = tensor<string, []>("op_3363_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_3364_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3363_to_fp16)[name = tensor<string, []>("op_3364_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3364_cast_fp16)[name = tensor<string, []>("input_817_cast_fp16")];
+            tensor<int32, [1]> key_31_axes_0 = const()[name = tensor<string, []>("key_31_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192853824)))];
+            tensor<fp16, [512]> encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192854912)))];
+            tensor<fp16, [1, 17, 512]> key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = tensor<string, []>("key_31_cast_fp16")];
+            tensor<bool, []> input_819_interleave_0 = const()[name = tensor<string, []>("input_819_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 87, 512]> input_819_cast_fp16 = concat(axis = var_64, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = tensor<string, []>("input_819_cast_fp16")];
+            tensor<int32, [3]> var_3386_begin_0 = const()[name = tensor<string, []>("op_3386_begin_0"), val = tensor<int32, [3]>([0, 17, 0])];
+            tensor<int32, [3]> var_3386_end_0 = const()[name = tensor<string, []>("op_3386_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_3386_end_mask_0 = const()[name = tensor<string, []>("op_3386_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 53, 512]> var_3386_cast_fp16 = slice_by_index(begin = var_3386_begin_0, end = var_3386_end_0, end_mask = var_3386_end_mask_0, x = cache_61_cast_fp16)[name = tensor<string, []>("op_3386_cast_fp16")];
+            tensor<bool, []> var_3392_interleave_0 = const()[name = tensor<string, []>("op_3392_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_3392_cast_fp16 = concat(axis = var_64, interleave = var_3392_interleave_0, values = (var_3386_cast_fp16, key_31_cast_fp16))[name = tensor<string, []>("op_3392_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_15_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192856000)))];
+            tensor<fp16, [1, 17, 512]> linear_138_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16, x = key_31_cast_fp16)[name = tensor<string, []>("linear_138_cast_fp16")];
+            tensor<int32, [4]> var_3396 = const()[name = tensor<string, []>("op_3396"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 17, 8, 64]> q_91_cast_fp16 = reshape(shape = var_3396, x = linear_138_cast_fp16)[name = tensor<string, []>("q_91_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_15_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193380352)))];
+            tensor<fp16, [1, 87, 512]> linear_139_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16, x = input_819_cast_fp16)[name = tensor<string, []>("linear_139_cast_fp16")];
+            tensor<int32, [4]> var_3400 = const()[name = tensor<string, []>("op_3400"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> k_61_cast_fp16 = reshape(shape = var_3400, x = linear_139_cast_fp16)[name = tensor<string, []>("k_61_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_15_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193904704)))];
+            tensor<fp16, [1, 87, 512]> linear_140_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16, x = input_819_cast_fp16)[name = tensor<string, []>("linear_140_cast_fp16")];
+            tensor<int32, [4]> var_3404 = const()[name = tensor<string, []>("op_3404"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> v_31_cast_fp16 = reshape(shape = var_3404, x = linear_140_cast_fp16)[name = tensor<string, []>("v_31_cast_fp16")];
+            tensor<int32, [4]> value_33_perm_0 = const()[name = tensor<string, []>("value_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194429056)))];
+            tensor<fp16, [1, 17, 8, 64]> var_3416_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3416_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194430144)))];
+            tensor<fp16, [1, 17, 8, 64]> var_3418_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3418_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_31_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_397_transpose_x_0 = const()[name = tensor<string, []>("x_397_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_397_transpose_y_0 = const()[name = tensor<string, []>("x_397_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 173]> var_3420_to_fp16 = const()[name = tensor<string, []>("op_3420_to_fp16"), val = tensor<fp16, [1, 8, 64, 173]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194431232)))];
+            tensor<fp16, [1, 8, 17, 64]> q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3418_cast_fp16)[name = tensor<string, []>("transpose_102")];
+            tensor<fp16, [1, 8, 17, 173]> x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = var_3420_to_fp16)[name = tensor<string, []>("x_397_cast_fp16")];
+            tensor<int32, [8]> x_399_pad_0 = const()[name = tensor<string, []>("x_399_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_399_mode_0 = const()[name = tensor<string, []>("x_399_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_218_to_fp16 = const()[name = tensor<string, []>("const_218_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 17, 174]> x_399_cast_fp16 = pad(constant_val = const_218_to_fp16, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397_cast_fp16)[name = tensor<string, []>("x_399_cast_fp16")];
+            tensor<int32, [4]> var_3428 = const()[name = tensor<string, []>("op_3428"), val = tensor<int32, [4]>([1, 8, -1, 17])];
+            tensor<fp16, [1, 8, 174, 17]> x_401_cast_fp16 = reshape(shape = var_3428, x = x_399_cast_fp16)[name = tensor<string, []>("x_401_cast_fp16")];
+            tensor<int32, [4]> var_3432_begin_0 = const()[name = tensor<string, []>("op_3432_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3432_end_0 = const()[name = tensor<string, []>("op_3432_end_0"), val = tensor<int32, [4]>([1, 8, 174, 17])];
+            tensor<bool, [4]> var_3432_end_mask_0 = const()[name = tensor<string, []>("op_3432_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 173, 17]> var_3432_cast_fp16 = slice_by_index(begin = var_3432_begin_0, end = var_3432_end_0, end_mask = var_3432_end_mask_0, x = x_401_cast_fp16)[name = tensor<string, []>("op_3432_cast_fp16")];
+            tensor<int32, [4]> var_3433 = const()[name = tensor<string, []>("op_3433"), val = tensor<int32, [4]>([1, 8, 17, 173])];
+            tensor<fp16, [1, 8, 17, 173]> matrix_bd_61_cast_fp16 = reshape(shape = var_3433, x = var_3432_cast_fp16)[name = tensor<string, []>("matrix_bd_61_cast_fp16")];
+            tensor<bool, []> matrix_ac_31_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_31_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_31_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_31_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_81_perm_0 = const()[name = tensor<string, []>("transpose_81_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_82_perm_0 = const()[name = tensor<string, []>("transpose_82_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 87]> transpose_82 = transpose(perm = transpose_82_perm_0, x = k_61_cast_fp16)[name = tensor<string, []>("transpose_100")];
+            tensor<fp16, [1, 8, 17, 64]> transpose_81 = transpose(perm = transpose_81_perm_0, x = var_3416_cast_fp16)[name = tensor<string, []>("transpose_101")];
+            tensor<fp16, [1, 8, 17, 87]> matrix_ac_31_cast_fp16 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_81, y = transpose_82)[name = tensor<string, []>("matrix_ac_31_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_63_begin_0 = const()[name = tensor<string, []>("matrix_bd_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_63_end_0 = const()[name = tensor<string, []>("matrix_bd_63_end_0"), val = tensor<int32, [4]>([1, 8, 17, 87])];
+            tensor<bool, [4]> matrix_bd_63_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 17, 87]> matrix_bd_63_cast_fp16 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61_cast_fp16)[name = tensor<string, []>("matrix_bd_63_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_3442_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = tensor<string, []>("op_3442_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_61_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_61_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 17, 87]> _inversed_scores_61_cast_fp16 = mul(x = var_3442_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_61_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> scores_63_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_63_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_3448_cast_fp16 = softmax(axis = var_62, x = scores_63_cast_fp16)[name = tensor<string, []>("op_3448_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> input_821_cast_fp16 = select(a = var_40_to_fp16, b = var_3448_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_821_cast_fp16")];
+            tensor<bool, []> x_403_transpose_x_0 = const()[name = tensor<string, []>("x_403_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_403_transpose_y_0 = const()[name = tensor<string, []>("x_403_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 87, 64]> value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_31_cast_fp16)[name = tensor<string, []>("transpose_103")];
+            tensor<fp16, [1, 8, 17, 64]> x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821_cast_fp16, y = value_33_cast_fp16)[name = tensor<string, []>("x_403_cast_fp16")];
+            tensor<int32, [4]> var_3452_perm_0 = const()[name = tensor<string, []>("op_3452_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3453 = const()[name = tensor<string, []>("op_3453"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 17, 8, 64]> var_3452_cast_fp16 = transpose(perm = var_3452_perm_0, x = x_403_cast_fp16)[name = tensor<string, []>("transpose_99")];
+            tensor<fp16, [1, 17, 512]> input_823_cast_fp16 = reshape(shape = var_3453, x = var_3452_cast_fp16)[name = tensor<string, []>("input_823_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_15_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194608448)))];
+            tensor<fp16, [1, 17, 512]> linear_142_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16, x = input_823_cast_fp16)[name = tensor<string, []>("linear_142_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = tensor<string, []>("input_827_cast_fp16")];
+            tensor<int32, [1]> x_407_axes_0 = const()[name = tensor<string, []>("x_407_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195132800)))];
+            tensor<fp16, [512]> encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195133888)))];
+            tensor<fp16, [1, 17, 512]> x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = tensor<string, []>("x_407_cast_fp16")];
+            tensor<int32, [3]> input_829_perm_0 = const()[name = tensor<string, []>("input_829_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_831_pad_type_0 = const()[name = tensor<string, []>("input_831_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_831_strides_0 = const()[name = tensor<string, []>("input_831_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_831_pad_0 = const()[name = tensor<string, []>("input_831_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_831_dilations_0 = const()[name = tensor<string, []>("input_831_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_831_groups_0 = const()[name = tensor<string, []>("input_831_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_15_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195134976)))];
+            tensor<fp16, [1, 512, 17]> input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = tensor<string, []>("transpose_98")];
+            tensor<fp16, [1, 1024, 17]> input_831_cast_fp16 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16, x = input_829_cast_fp16)[name = tensor<string, []>("input_831_cast_fp16")];
+            tensor<int32, []> x_409_split_num_splits_0 = const()[name = tensor<string, []>("x_409_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_409_split_axis_0 = const()[name = tensor<string, []>("x_409_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 17]> x_409_split_cast_fp16_0, tensor<fp16, [1, 512, 17]> x_409_split_cast_fp16_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831_cast_fp16)[name = tensor<string, []>("x_409_split_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = tensor<string, []>("x_409_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_409_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> input_833_cast_fp16 = select(a = var_40_to_fp16, b = x_409_cast_fp16, cond = var_418)[name = tensor<string, []>("input_833_cast_fp16")];
+            tensor<bool, []> new_x_63_interleave_0 = const()[name = tensor<string, []>("new_x_63_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 25]> new_x_63_cast_fp16 = concat(axis = var_62, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = tensor<string, []>("new_x_63_cast_fp16")];
+            tensor<int32, [3]> var_3491_begin_0 = const()[name = tensor<string, []>("op_3491_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
+            tensor<int32, [3]> var_3491_end_0 = const()[name = tensor<string, []>("op_3491_end_0"), val = tensor<int32, [3]>([1, 512, 25])];
+            tensor<bool, [3]> var_3491_end_mask_0 = const()[name = tensor<string, []>("op_3491_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_3491_cast_fp16 = slice_by_index(begin = var_3491_begin_0, end = var_3491_end_0, end_mask = var_3491_end_mask_0, x = new_x_63_cast_fp16)[name = tensor<string, []>("op_3491_cast_fp16")];
+            tensor<string, []> x_411_pad_type_0 = const()[name = tensor<string, []>("x_411_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_411_groups_0 = const()[name = tensor<string, []>("x_411_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_411_strides_0 = const()[name = tensor<string, []>("x_411_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_411_pad_0 = const()[name = tensor<string, []>("x_411_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_411_dilations_0 = const()[name = tensor<string, []>("x_411_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_15_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196183616)))];
+            tensor<fp16, [1, 512, 17]> x_411_cast_fp16 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16, x = new_x_63_cast_fp16)[name = tensor<string, []>("x_411_cast_fp16")];
+            tensor<int32, [3]> input_835_perm_0 = const()[name = tensor<string, []>("input_835_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_413_axes_0 = const()[name = tensor<string, []>("x_413_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196192896)))];
+            tensor<fp16, [512]> encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196193984)))];
+            tensor<fp16, [1, 17, 512]> input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = tensor<string, []>("transpose_97")];
+            tensor<fp16, [1, 17, 512]> x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = tensor<string, []>("x_413_cast_fp16")];
+            tensor<int32, [3]> input_837_perm_0 = const()[name = tensor<string, []>("input_837_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 17]> input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = tensor<string, []>("transpose_96")];
+            tensor<fp16, [1, 512, 17]> input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = tensor<string, []>("input_839_cast_fp16")];
+            tensor<string, []> x_415_pad_type_0 = const()[name = tensor<string, []>("x_415_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_415_strides_0 = const()[name = tensor<string, []>("x_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_415_pad_0 = const()[name = tensor<string, []>("x_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_415_dilations_0 = const()[name = tensor<string, []>("x_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_415_groups_0 = const()[name = tensor<string, []>("x_415_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_15_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196195072)))];
+            tensor<fp16, [1, 512, 17]> x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16, x = input_839_cast_fp16)[name = tensor<string, []>("x_415_cast_fp16")];
+            tensor<int32, [3]> input_841_perm_0 = const()[name = tensor<string, []>("input_841_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 17, 512]> input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = tensor<string, []>("transpose_95")];
+            tensor<fp16, [1, 17, 512]> input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = tensor<string, []>("input_843_cast_fp16")];
+            tensor<int32, [1]> input_845_axes_0 = const()[name = tensor<string, []>("input_845_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196719424)))];
+            tensor<fp16, [512]> encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196720512)))];
+            tensor<fp16, [1, 17, 512]> input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = tensor<string, []>("input_845_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_15_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196721600)))];
+            tensor<fp16, [1, 17, 2048]> linear_143_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16, x = input_845_cast_fp16)[name = tensor<string, []>("linear_143_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = tensor<string, []>("input_849_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_15_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198818816)))];
+            tensor<fp16, [1, 17, 512]> linear_144_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16, x = input_849_cast_fp16)[name = tensor<string, []>("linear_144_cast_fp16")];
+            tensor<fp16, []> var_3532_to_fp16 = const()[name = tensor<string, []>("op_3532_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_3533_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3532_to_fp16)[name = tensor<string, []>("op_3533_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3533_cast_fp16)[name = tensor<string, []>("input_855_cast_fp16")];
+            tensor<int32, [1]> input_857_axes_0 = const()[name = tensor<string, []>("input_857_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_15_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200916032)))];
+            tensor<fp16, [512]> encoder_layers_15_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200917120)))];
+            tensor<fp16, [1, 17, 512]> input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = tensor<string, []>("input_857_cast_fp16")];
+            tensor<int32, [4]> cache_65_begin_0 = const()[name = tensor<string, []>("cache_65_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_65_end_0 = const()[name = tensor<string, []>("cache_65_end_0"), val = tensor<int32, [4]>([17, 1, 70, 512])];
+            tensor<bool, [4]> cache_65_end_mask_0 = const()[name = tensor<string, []>("cache_65_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_65_squeeze_mask_0 = const()[name = tensor<string, []>("cache_65_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_65_cast_fp16 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_65_cast_fp16")];
+            tensor<int32, [4]> cache_begin_0 = const()[name = tensor<string, []>("cache_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_end_0 = const()[name = tensor<string, []>("cache_end_0"), val = tensor<int32, [4]>([17, 1, 512, 8])];
+            tensor<bool, [4]> cache_end_mask_0 = const()[name = tensor<string, []>("cache_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_squeeze_mask_0 = const()[name = tensor<string, []>("cache_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_cast_fp16 = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_cast_fp16")];
+            tensor<int32, [1]> input_859_axes_0 = const()[name = tensor<string, []>("input_859_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200918208)))];
+            tensor<fp16, [512]> encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200919296)))];
+            tensor<fp16, [1, 17, 512]> input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = tensor<string, []>("input_859_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_16_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200920384)))];
+            tensor<fp16, [1, 17, 2048]> linear_145_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16, x = input_859_cast_fp16)[name = tensor<string, []>("linear_145_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = tensor<string, []>("input_863_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_16_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203017600)))];
+            tensor<fp16, [1, 17, 512]> linear_146_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16, x = input_863_cast_fp16)[name = tensor<string, []>("linear_146_cast_fp16")];
+            tensor<fp16, []> var_3567_to_fp16 = const()[name = tensor<string, []>("op_3567_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_3568_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3567_to_fp16)[name = tensor<string, []>("op_3568_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3568_cast_fp16)[name = tensor<string, []>("input_869_cast_fp16")];
+            tensor<int32, [1]> key_axes_0 = const()[name = tensor<string, []>("key_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205114816)))];
+            tensor<fp16, [512]> encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205115904)))];
+            tensor<fp16, [1, 17, 512]> key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = tensor<string, []>("key_cast_fp16")];
+            tensor<bool, []> input_871_interleave_0 = const()[name = tensor<string, []>("input_871_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 87, 512]> input_871_cast_fp16 = concat(axis = var_64, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_cast_fp16))[name = tensor<string, []>("input_871_cast_fp16")];
+            tensor<int32, [3]> var_3590_begin_0 = const()[name = tensor<string, []>("op_3590_begin_0"), val = tensor<int32, [3]>([0, 17, 0])];
+            tensor<int32, [3]> var_3590_end_0 = const()[name = tensor<string, []>("op_3590_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_3590_end_mask_0 = const()[name = tensor<string, []>("op_3590_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 53, 512]> var_3590_cast_fp16 = slice_by_index(begin = var_3590_begin_0, end = var_3590_end_0, end_mask = var_3590_end_mask_0, x = cache_65_cast_fp16)[name = tensor<string, []>("op_3590_cast_fp16")];
+            tensor<bool, []> cache_last_channel_cur_interleave_0 = const()[name = tensor<string, []>("cache_last_channel_cur_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> cache_last_channel_cur_cast_fp16 = concat(axis = var_64, interleave = cache_last_channel_cur_interleave_0, values = (var_3590_cast_fp16, key_cast_fp16))[name = tensor<string, []>("cache_last_channel_cur_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_16_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205116992)))];
+            tensor<fp16, [1, 17, 512]> linear_147_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16, x = key_cast_fp16)[name = tensor<string, []>("linear_147_cast_fp16")];
+            tensor<int32, [4]> var_3600 = const()[name = tensor<string, []>("op_3600"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 17, 8, 64]> q_97_cast_fp16 = reshape(shape = var_3600, x = linear_147_cast_fp16)[name = tensor<string, []>("q_97_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_16_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205641344)))];
+            tensor<fp16, [1, 87, 512]> linear_148_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16, x = input_871_cast_fp16)[name = tensor<string, []>("linear_148_cast_fp16")];
+            tensor<int32, [4]> var_3604 = const()[name = tensor<string, []>("op_3604"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> k_65_cast_fp16 = reshape(shape = var_3604, x = linear_148_cast_fp16)[name = tensor<string, []>("k_65_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_16_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206165696)))];
+            tensor<fp16, [1, 87, 512]> linear_149_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16, x = input_871_cast_fp16)[name = tensor<string, []>("linear_149_cast_fp16")];
+            tensor<int32, [4]> var_3608 = const()[name = tensor<string, []>("op_3608"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 87, 8, 64]> v_cast_fp16 = reshape(shape = var_3608, x = linear_149_cast_fp16)[name = tensor<string, []>("v_cast_fp16")];
+            tensor<int32, [4]> value_perm_0 = const()[name = tensor<string, []>("value_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206690048)))];
+            tensor<fp16, [1, 17, 8, 64]> var_3620_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3620_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206691136)))];
+            tensor<fp16, [1, 17, 8, 64]> var_3622_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3622_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_423_transpose_x_0 = const()[name = tensor<string, []>("x_423_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_423_transpose_y_0 = const()[name = tensor<string, []>("x_423_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 173]> var_3624_to_fp16 = const()[name = tensor<string, []>("op_3624_to_fp16"), val = tensor<fp16, [1, 8, 64, 173]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206692224)))];
+            tensor<fp16, [1, 8, 17, 64]> q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_3622_cast_fp16)[name = tensor<string, []>("transpose_93")];
+            tensor<fp16, [1, 8, 17, 173]> x_423_cast_fp16 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_cast_fp16, y = var_3624_to_fp16)[name = tensor<string, []>("x_423_cast_fp16")];
+            tensor<int32, [8]> x_425_pad_0 = const()[name = tensor<string, []>("x_425_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_425_mode_0 = const()[name = tensor<string, []>("x_425_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_231_to_fp16 = const()[name = tensor<string, []>("const_231_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 17, 174]> x_425_cast_fp16 = pad(constant_val = const_231_to_fp16, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423_cast_fp16)[name = tensor<string, []>("x_425_cast_fp16")];
+            tensor<int32, [4]> var_3632 = const()[name = tensor<string, []>("op_3632"), val = tensor<int32, [4]>([1, 8, -1, 17])];
+            tensor<fp16, [1, 8, 174, 17]> x_427_cast_fp16 = reshape(shape = var_3632, x = x_425_cast_fp16)[name = tensor<string, []>("x_427_cast_fp16")];
+            tensor<int32, [4]> var_3636_begin_0 = const()[name = tensor<string, []>("op_3636_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3636_end_0 = const()[name = tensor<string, []>("op_3636_end_0"), val = tensor<int32, [4]>([1, 8, 174, 17])];
+            tensor<bool, [4]> var_3636_end_mask_0 = const()[name = tensor<string, []>("op_3636_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 173, 17]> var_3636_cast_fp16 = slice_by_index(begin = var_3636_begin_0, end = var_3636_end_0, end_mask = var_3636_end_mask_0, x = x_427_cast_fp16)[name = tensor<string, []>("op_3636_cast_fp16")];
+            tensor<int32, [4]> var_3637 = const()[name = tensor<string, []>("op_3637"), val = tensor<int32, [4]>([1, 8, 17, 173])];
+            tensor<fp16, [1, 8, 17, 173]> matrix_bd_65_cast_fp16 = reshape(shape = var_3637, x = var_3636_cast_fp16)[name = tensor<string, []>("matrix_bd_65_cast_fp16")];
+            tensor<bool, []> matrix_ac_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_83_perm_0 = const()[name = tensor<string, []>("transpose_83_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_84_perm_0 = const()[name = tensor<string, []>("transpose_84_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 87]> transpose_84 = transpose(perm = transpose_84_perm_0, x = k_65_cast_fp16)[name = tensor<string, []>("transpose_91")];
+            tensor<fp16, [1, 8, 17, 64]> transpose_83 = transpose(perm = transpose_83_perm_0, x = var_3620_cast_fp16)[name = tensor<string, []>("transpose_92")];
+            tensor<fp16, [1, 8, 17, 87]> matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_83, y = transpose_84)[name = tensor<string, []>("matrix_ac_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_begin_0 = const()[name = tensor<string, []>("matrix_bd_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_end_0 = const()[name = tensor<string, []>("matrix_bd_end_0"), val = tensor<int32, [4]>([1, 8, 17, 87])];
+            tensor<bool, [4]> matrix_bd_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 17, 87]> matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_65_cast_fp16)[name = tensor<string, []>("matrix_bd_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_3646_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = tensor<string, []>("op_3646_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_65_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_65_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 17, 87]> _inversed_scores_65_cast_fp16 = mul(x = var_3646_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_65_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> scores_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> var_3652_cast_fp16 = softmax(axis = var_62, x = scores_cast_fp16)[name = tensor<string, []>("op_3652_cast_fp16")];
+            tensor<fp16, [1, 8, 17, 87]> input_873_cast_fp16 = select(a = var_40_to_fp16, b = var_3652_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_873_cast_fp16")];
+            tensor<bool, []> x_429_transpose_x_0 = const()[name = tensor<string, []>("x_429_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_429_transpose_y_0 = const()[name = tensor<string, []>("x_429_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 87, 64]> value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = tensor<string, []>("transpose_94")];
+            tensor<fp16, [1, 8, 17, 64]> x_429_cast_fp16 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873_cast_fp16, y = value_cast_fp16)[name = tensor<string, []>("x_429_cast_fp16")];
+            tensor<int32, [4]> var_3656_perm_0 = const()[name = tensor<string, []>("op_3656_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3657 = const()[name = tensor<string, []>("op_3657"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 17, 8, 64]> var_3656_cast_fp16 = transpose(perm = var_3656_perm_0, x = x_429_cast_fp16)[name = tensor<string, []>("transpose_90")];
+            tensor<fp16, [1, 17, 512]> input_875_cast_fp16 = reshape(shape = var_3657, x = var_3656_cast_fp16)[name = tensor<string, []>("input_875_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_16_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206869440)))];
+            tensor<fp16, [1, 17, 512]> linear_151_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16, x = input_875_cast_fp16)[name = tensor<string, []>("linear_151_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = tensor<string, []>("input_879_cast_fp16")];
+            tensor<int32, [1]> x_433_axes_0 = const()[name = tensor<string, []>("x_433_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207393792)))];
+            tensor<fp16, [512]> encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207394880)))];
+            tensor<fp16, [1, 17, 512]> x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = tensor<string, []>("x_433_cast_fp16")];
+            tensor<int32, [3]> input_881_perm_0 = const()[name = tensor<string, []>("input_881_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_883_pad_type_0 = const()[name = tensor<string, []>("input_883_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_883_strides_0 = const()[name = tensor<string, []>("input_883_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_883_pad_0 = const()[name = tensor<string, []>("input_883_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_883_dilations_0 = const()[name = tensor<string, []>("input_883_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_883_groups_0 = const()[name = tensor<string, []>("input_883_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_16_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207395968)))];
+            tensor<fp16, [1, 512, 17]> input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = tensor<string, []>("transpose_89")];
+            tensor<fp16, [1, 1024, 17]> input_883_cast_fp16 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16, x = input_881_cast_fp16)[name = tensor<string, []>("input_883_cast_fp16")];
+            tensor<int32, []> x_435_split_num_splits_0 = const()[name = tensor<string, []>("x_435_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_435_split_axis_0 = const()[name = tensor<string, []>("x_435_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 17]> x_435_split_cast_fp16_0, tensor<fp16, [1, 512, 17]> x_435_split_cast_fp16_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883_cast_fp16)[name = tensor<string, []>("x_435_split_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = tensor<string, []>("x_435_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_435_cast_fp16")];
+            tensor<fp16, [1, 512, 17]> input_885_cast_fp16 = select(a = var_40_to_fp16, b = x_435_cast_fp16, cond = var_418)[name = tensor<string, []>("input_885_cast_fp16")];
+            tensor<bool, []> new_x_interleave_0 = const()[name = tensor<string, []>("new_x_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 25]> new_x_cast_fp16 = concat(axis = var_62, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_885_cast_fp16))[name = tensor<string, []>("new_x_cast_fp16")];
+            tensor<int32, [3]> cache_last_time_cur_begin_0 = const()[name = tensor<string, []>("cache_last_time_cur_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
+            tensor<int32, [3]> cache_last_time_cur_end_0 = const()[name = tensor<string, []>("cache_last_time_cur_end_0"), val = tensor<int32, [3]>([1, 512, 25])];
+            tensor<bool, [3]> cache_last_time_cur_end_mask_0 = const()[name = tensor<string, []>("cache_last_time_cur_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> cache_last_time_cur_cast_fp16 = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = new_x_cast_fp16)[name = tensor<string, []>("cache_last_time_cur_cast_fp16")];
+            tensor<string, []> x_437_pad_type_0 = const()[name = tensor<string, []>("x_437_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_437_groups_0 = const()[name = tensor<string, []>("x_437_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_437_strides_0 = const()[name = tensor<string, []>("x_437_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_437_pad_0 = const()[name = tensor<string, []>("x_437_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_437_dilations_0 = const()[name = tensor<string, []>("x_437_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_16_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208444608)))];
+            tensor<fp16, [1, 512, 17]> x_437_cast_fp16 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16, x = new_x_cast_fp16)[name = tensor<string, []>("x_437_cast_fp16")];
+            tensor<int32, [3]> input_887_perm_0 = const()[name = tensor<string, []>("input_887_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_439_axes_0 = const()[name = tensor<string, []>("x_439_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208453888)))];
+            tensor<fp16, [512]> encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208454976)))];
+            tensor<fp16, [1, 17, 512]> input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = tensor<string, []>("transpose_88")];
+            tensor<fp16, [1, 17, 512]> x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = tensor<string, []>("x_439_cast_fp16")];
+            tensor<int32, [3]> input_889_perm_0 = const()[name = tensor<string, []>("input_889_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 17]> input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = tensor<string, []>("transpose_87")];
+            tensor<fp16, [1, 512, 17]> input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = tensor<string, []>("input_891_cast_fp16")];
+            tensor<string, []> x_441_pad_type_0 = const()[name = tensor<string, []>("x_441_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_441_strides_0 = const()[name = tensor<string, []>("x_441_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_441_pad_0 = const()[name = tensor<string, []>("x_441_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_441_dilations_0 = const()[name = tensor<string, []>("x_441_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_441_groups_0 = const()[name = tensor<string, []>("x_441_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_16_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208456064)))];
+            tensor<fp16, [1, 512, 17]> x_441_cast_fp16 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16, x = input_891_cast_fp16)[name = tensor<string, []>("x_441_cast_fp16")];
+            tensor<int32, [3]> input_893_perm_0 = const()[name = tensor<string, []>("input_893_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 17, 512]> input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = tensor<string, []>("transpose_86")];
+            tensor<fp16, [1, 17, 512]> input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = tensor<string, []>("input_895_cast_fp16")];
+            tensor<int32, [1]> input_897_axes_0 = const()[name = tensor<string, []>("input_897_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208980416)))];
+            tensor<fp16, [512]> encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208981504)))];
+            tensor<fp16, [1, 17, 512]> input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = tensor<string, []>("input_897_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_16_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208982592)))];
+            tensor<fp16, [1, 17, 2048]> linear_152_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16, x = input_897_cast_fp16)[name = tensor<string, []>("linear_152_cast_fp16")];
+            tensor<fp16, [1, 17, 2048]> input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = tensor<string, []>("input_901_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_16_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211079808)))];
+            tensor<fp16, [1, 17, 512]> linear_153_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16, x = input_901_cast_fp16)[name = tensor<string, []>("linear_153_cast_fp16")];
+            tensor<fp16, []> var_3736_to_fp16 = const()[name = tensor<string, []>("op_3736_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 17, 512]> var_3737_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_3736_to_fp16)[name = tensor<string, []>("op_3737_cast_fp16")];
+            tensor<fp16, [1, 17, 512]> input_cast_fp16 = add(x = input_895_cast_fp16, y = var_3737_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [1]> audio_signal_axes_0 = const()[name = tensor<string, []>("audio_signal_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_16_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213177024)))];
+            tensor<fp16, [512]> encoder_layers_16_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213178112)))];
+            tensor<fp16, [1, 17, 512]> audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("audio_signal_cast_fp16")];
+            tensor<int32, [3]> obj_1_perm_0 = const()[name = tensor<string, []>("obj_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> obj_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_1_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, []> obj_5_axis_0 = const()[name = tensor<string, []>("obj_5_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [17, 1, 70, 512]> obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_332_cast_fp16, var_536_cast_fp16, var_740_cast_fp16, var_944_cast_fp16, var_1148_cast_fp16, var_1352_cast_fp16, var_1556_cast_fp16, var_1760_cast_fp16, var_1964_cast_fp16, var_2168_cast_fp16, var_2372_cast_fp16, var_2576_cast_fp16, var_2780_cast_fp16, var_2984_cast_fp16, var_3188_cast_fp16, var_3392_cast_fp16, cache_last_channel_cur_cast_fp16))[name = tensor<string, []>("obj_5_cast_fp16")];
+            tensor<int32, []> obj_7_axis_0 = const()[name = tensor<string, []>("obj_7_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [17, 1, 512, 8]> obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_431_cast_fp16, var_635_cast_fp16, var_839_cast_fp16, var_1043_cast_fp16, var_1247_cast_fp16, var_1451_cast_fp16, var_1655_cast_fp16, var_1859_cast_fp16, var_2063_cast_fp16, var_2267_cast_fp16, var_2471_cast_fp16, var_2675_cast_fp16, var_2879_cast_fp16, var_3083_cast_fp16, var_3287_cast_fp16, var_3491_cast_fp16, cache_last_time_cur_cast_fp16))[name = tensor<string, []>("obj_7_cast_fp16")];
+            tensor<string, []> obj_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_7_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> var_3753 = add(x = cache_last_channel_len, y = max_audio_length_1)[name = tensor<string, []>("op_3753")];
+            tensor<string, []> var_3753_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3753_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> const_237_to_fp16 = const()[name = tensor<string, []>("const_237_to_fp16"), val = tensor<fp16, []>(-inf)];
+            tensor<fp16, []> var_45_promoted_to_fp16 = const()[name = tensor<string, []>("op_45_promoted_to_fp16"), val = tensor<fp16, []>(0x1.18p+6)];
+            tensor<fp16, [1]> var_3753_to_fp16 = cast(dtype = var_3753_promoted_to_fp16_dtype_0, x = var_3753)[name = tensor<string, []>("cast_184")];
+            tensor<fp16, [1]> clip_1_cast_fp16 = clip(alpha = const_237_to_fp16, beta = var_45_promoted_to_fp16, x = var_3753_to_fp16)[name = tensor<string, []>("clip_1_cast_fp16")];
+            tensor<string, []> var_3780_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3780_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<string, []> cast_179_dtype_0 = const()[name = tensor<string, []>("cast_179_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<string, []> cast_180_dtype_0 = const()[name = tensor<string, []>("cast_180_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1]> new_cache_last_channel_len = cast(dtype = cast_180_dtype_0, x = clip_1_cast_fp16)[name = tensor<string, []>("cast_181")];
+            tensor<int32, [1]> encoded_length = cast(dtype = cast_179_dtype_0, x = clip_0_cast_fp16)[name = tensor<string, []>("cast_182")];
+            tensor<fp32, [17, 1, 70, 512]> new_cache_last_channel = cast(dtype = var_3780_cast_fp16_to_fp32_dtype_0, x = obj_5_cast_fp16)[name = tensor<string, []>("cast_183")];
+            tensor<fp32, [17, 1, 512, 8]> new_cache_last_time = cast(dtype = obj_7_cast_fp16_to_fp32_dtype_0, x = obj_7_cast_fp16)[name = tensor<string, []>("cast_185")];
+            tensor<fp16, [1, 512, 17]> obj_1_cast_fp16 = transpose(perm = obj_1_perm_0, x = audio_signal_cast_fp16)[name = tensor<string, []>("transpose_85")];
+            tensor<fp32, [1, 512, 17]> encoded_output = cast(dtype = obj_1_cast_fp16_to_fp32_dtype_0, x = obj_1_cast_fp16)[name = tensor<string, []>("cast_186")];
+            tensor<fp32, [1, 128, 16]> new_pre_cache = cast(dtype = var_28_cast_fp16_to_fp32_dtype_0, x = var_28_cast_fp16)[name = tensor<string, []>("cast_192")];
+        } -> (encoded_output, encoded_length, new_pre_cache, new_cache_last_channel, new_cache_last_time, new_cache_last_channel_len);
+}
\ No newline at end of file
diff --git a/1280ms/streaming_encoder.mlmodelc/weights/weight.bin b/1280ms/streaming_encoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..6fc415d6e82641770c5d78f17d76d1e71d028e06
--- /dev/null
+++ b/1280ms/streaming_encoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c71acb590ceb2af449de5c7e3516e76057eaf4589d1f16edba774831db74b17
+size 213179200
diff --git a/1280ms/streaming_encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/1280ms/streaming_encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..7e0bb55b1d00510e94e0c0473b9231362e92c892
--- /dev/null
+++ b/1280ms/streaming_encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9181d223091aa592cb656d49346e640a38ec2426de5ec2d06edbc14e92b8968b
+size 508252
diff --git a/1280ms/streaming_encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/1280ms/streaming_encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..6fc415d6e82641770c5d78f17d76d1e71d028e06
--- /dev/null
+++ b/1280ms/streaming_encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c71acb590ceb2af449de5c7e3516e76057eaf4589d1f16edba774831db74b17
+size 213179200
diff --git a/1280ms/streaming_encoder.mlpackage/Manifest.json b/1280ms/streaming_encoder.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..e1f379019d051455a60312bf6f5664348c61be66
--- /dev/null
+++ b/1280ms/streaming_encoder.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "468B5E19-6BA9-478C-8D2A-23953ACBD5E3": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "F22DE286-FE1A-4BBD-A7ED-B0130595DAF3": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "468B5E19-6BA9-478C-8D2A-23953ACBD5E3"
+}
diff --git a/1280ms/vocab.json b/1280ms/vocab.json
new file mode 100644
index 0000000000000000000000000000000000000000..462081190dbb331a0dde13e99d5cfa99b302f0a7
--- /dev/null
+++ b/1280ms/vocab.json
@@ -0,0 +1,1028 @@
+{
+  "0": "<unk>",
+  "1": "▁t",
+  "2": "▁th",
+  "3": "▁a",
+  "4": "▁i",
+  "5": "▁the",
+  "6": "▁s",
+  "7": "re",
+  "8": "▁w",
+  "9": "▁o",
+  "10": "in",
+  "11": "at",
+  "12": "er",
+  "13": "nd",
+  "14": "ou",
+  "15": "▁c",
+  "16": "▁b",
+  "17": "▁h",
+  "18": "en",
+  "19": "on",
+  "20": "▁m",
+  "21": "▁f",
+  "22": "ing",
+  "23": "▁p",
+  "24": "▁to",
+  "25": "▁and",
+  "26": "▁d",
+  "27": "an",
+  "28": "or",
+  "29": "es",
+  "30": "▁y",
+  "31": "▁l",
+  "32": "▁of",
+  "33": "ll",
+  "34": "▁in",
+  "35": "ed",
+  "36": "it",
+  "37": "▁g",
+  "38": "is",
+  "39": "▁you",
+  "40": "▁n",
+  "41": "ar",
+  "42": "om",
+  "43": "as",
+  "44": "ve",
+  "45": "▁e",
+  "46": "ic",
+  "47": "▁it",
+  "48": "al",
+  "49": "us",
+  "50": "▁wh",
+  "51": "▁we",
+  "52": "▁be",
+  "53": "ion",
+  "54": "ow",
+  "55": "le",
+  "56": "▁is",
+  "57": "et",
+  "58": "ent",
+  "59": "ot",
+  "60": "ut",
+  "61": "▁re",
+  "62": "▁on",
+  "63": "ay",
+  "64": "▁ha",
+  "65": "ig",
+  "66": "▁so",
+  "67": "ct",
+  "68": "▁he",
+  "69": "▁for",
+  "70": "ver",
+  "71": "ke",
+  "72": "ro",
+  "73": "▁st",
+  "74": "id",
+  "75": "▁go",
+  "76": "all",
+  "77": "se",
+  "78": "ly",
+  "79": "▁u",
+  "80": "ch",
+  "81": "st",
+  "82": "ld",
+  "83": "▁k",
+  "84": "ce",
+  "85": "ur",
+  "86": "▁li",
+  "87": "am",
+  "88": "▁r",
+  "89": "ht",
+  "90": "▁j",
+  "91": "ith",
+  "92": "▁se",
+  "93": "ir",
+  "94": "▁as",
+  "95": "▁an",
+  "96": "im",
+  "97": "▁do",
+  "98": "ad",
+  "99": "▁was",
+  "100": "ight",
+  "101": "th",
+  "102": "▁are",
+  "103": "▁but",
+  "104": "▁sh",
+  "105": "ust",
+  "106": "ally",
+  "107": "▁not",
+  "108": "▁or",
+  "109": "▁com",
+  "110": "▁can",
+  "111": "▁me",
+  "112": "op",
+  "113": "▁mo",
+  "114": "▁at",
+  "115": "ill",
+  "116": "▁ch",
+  "117": "▁ne",
+  "118": "ant",
+  "119": "▁de",
+  "120": "▁kn",
+  "121": "▁one",
+  "122": "il",
+  "123": "ol",
+  "124": "▁con",
+  "125": "ter",
+  "126": "▁ab",
+  "127": "▁fr",
+  "128": "ere",
+  "129": "ck",
+  "130": "▁al",
+  "131": "▁all",
+  "132": "qu",
+  "133": "▁pro",
+  "134": "▁som",
+  "135": "ould",
+  "136": "▁tw",
+  "137": "ul",
+  "138": "ra",
+  "139": "od",
+  "140": "ers",
+  "141": "▁su",
+  "142": "ive",
+  "143": "▁v",
+  "144": "use",
+  "145": "ate",
+  "146": "ge",
+  "147": "if",
+  "148": "▁ex",
+  "149": "ess",
+  "150": "pp",
+  "151": "▁lo",
+  "152": "out",
+  "153": "▁if",
+  "154": "est",
+  "155": "ain",
+  "156": "ist",
+  "157": "and",
+  "158": "ea",
+  "159": "very",
+  "160": "art",
+  "161": "▁wor",
+  "162": "▁my",
+  "163": "ab",
+  "164": "ment",
+  "165": "▁bec",
+  "166": "un",
+  "167": "ity",
+  "168": "ri",
+  "169": "pe",
+  "170": "ions",
+  "171": "▁by",
+  "172": "ok",
+  "173": "our",
+  "174": "ort",
+  "175": "ind",
+  "176": "ink",
+  "177": "nt",
+  "178": "▁up",
+  "179": "um",
+  "180": "▁don",
+  "181": "▁get",
+  "182": "red",
+  "183": "▁out",
+  "184": "el",
+  "185": "ause",
+  "186": "res",
+  "187": "▁ma",
+  "188": "ich",
+  "189": "▁us",
+  "190": "rou",
+  "191": "▁int",
+  "192": "em",
+  "193": "os",
+  "194": "ies",
+  "195": "ie",
+  "196": "▁pl",
+  "197": "▁tr",
+  "198": "ven",
+  "199": "ous",
+  "200": "▁le",
+  "201": "▁two",
+  "202": "ard",
+  "203": "ine",
+  "204": "▁co",
+  "205": "een",
+  "206": "▁now",
+  "207": "ty",
+  "208": "her",
+  "209": "ack",
+  "210": "▁pe",
+  "211": "ame",
+  "212": "▁how",
+  "213": "▁who",
+  "214": "▁see",
+  "215": "▁tim",
+  "216": "ect",
+  "217": "ast",
+  "218": "▁our",
+  "219": "ci",
+  "220": "ree",
+  "221": "ople",
+  "222": "gh",
+  "223": "▁no",
+  "224": "▁had",
+  "225": "▁man",
+  "226": "▁qu",
+  "227": "▁en",
+  "228": "ide",
+  "229": "ure",
+  "230": "ud",
+  "231": "so",
+  "232": "▁his",
+  "233": "▁sa",
+  "234": "▁sp",
+  "235": "▁say",
+  "236": "ose",
+  "237": "ther",
+  "238": "▁act",
+  "239": "▁ta",
+  "240": "▁cl",
+  "241": "ings",
+  "242": "pt",
+  "243": "king",
+  "244": "▁any",
+  "245": "▁has",
+  "246": "▁un",
+  "247": "iv",
+  "248": "▁im",
+  "249": "▁ag",
+  "250": "▁te",
+  "251": "▁fe",
+  "252": "one",
+  "253": "per",
+  "254": "ong",
+  "255": "▁po",
+  "256": "▁ad",
+  "257": "ff",
+  "258": "ore",
+  "259": "itt",
+  "260": "ans",
+  "261": "iz",
+  "262": "eah",
+  "263": "reat",
+  "264": "act",
+  "265": "own",
+  "266": "hing",
+  "267": "enty",
+  "268": "age",
+  "269": "ber",
+  "270": "ice",
+  "271": "▁am",
+  "272": "ple",
+  "273": "are",
+  "274": "▁per",
+  "275": "und",
+  "276": "ite",
+  "277": "ix",
+  "278": "pl",
+  "279": "▁way",
+  "280": "▁did",
+  "281": "▁pr",
+  "282": "▁got",
+  "283": "ars",
+  "284": "▁she",
+  "285": "▁let",
+  "286": "ag",
+  "287": "▁ac",
+  "288": "int",
+  "289": "▁ar",
+  "290": "ry",
+  "291": "ign",
+  "292": "ish",
+  "293": "▁fir",
+  "294": "ace",
+  "295": "ble",
+  "296": "og",
+  "297": "ue",
+  "298": "▁ye",
+  "299": "ap",
+  "300": "iff",
+  "301": "▁ro",
+  "302": "▁her",
+  "303": "nder",
+  "304": "▁ok",
+  "305": "▁res",
+  "306": "▁gu",
+  "307": "ence",
+  "308": "▁may",
+  "309": "ated",
+  "310": "ip",
+  "311": "▁bo",
+  "312": "▁him",
+  "313": "way",
+  "314": "ac",
+  "315": "ical",
+  "316": "ass",
+  "317": "ase",
+  "318": "▁dis",
+  "319": "able",
+  "320": "ick",
+  "321": "▁app",
+  "322": "ance",
+  "323": "▁pre",
+  "324": "▁six",
+  "325": "▁off",
+  "326": "▁new",
+  "327": "ia",
+  "328": "orm",
+  "329": "ank",
+  "330": "▁lot",
+  "331": "ach",
+  "332": "▁fo",
+  "333": "inet",
+  "334": "ire",
+  "335": "ary",
+  "336": "ult",
+  "337": "▁tal",
+  "338": "▁mu",
+  "339": "▁bl",
+  "340": "ount",
+  "341": "sel",
+  "342": "vel",
+  "343": "▁br",
+  "344": "▁imp",
+  "345": "ep",
+  "346": "cess",
+  "347": "ord",
+  "348": "▁sc",
+  "349": "▁inc",
+  "350": "ound",
+  "351": "ang",
+  "352": "be",
+  "353": "ress",
+  "354": "uct",
+  "355": "▁ind",
+  "356": "▁af",
+  "357": "ving",
+  "358": "▁oh",
+  "359": "▁bet",
+  "360": "▁use",
+  "361": "ome",
+  "362": "ens",
+  "363": "ys",
+  "364": "▁bu",
+  "365": "co",
+  "366": "ory",
+  "367": "ater",
+  "368": "ild",
+  "369": "ght",
+  "370": "ial",
+  "371": "▁day",
+  "372": "ning",
+  "373": "na",
+  "374": "ile",
+  "375": "▁spe",
+  "376": "▁mar",
+  "377": "ody",
+  "378": "ough",
+  "379": "ade",
+  "380": "vers",
+  "381": "xt",
+  "382": "▁fl",
+  "383": "▁ke",
+  "384": "ian",
+  "385": "▁sy",
+  "386": "▁put",
+  "387": "fore",
+  "388": "ub",
+  "389": "▁ph",
+  "390": "fe",
+  "391": "▁em",
+  "392": "▁ser",
+  "393": "form",
+  "394": "ting",
+  "395": "te",
+  "396": "av",
+  "397": "ious",
+  "398": "▁rec",
+  "399": "ks",
+  "400": "▁gr",
+  "401": "ces",
+  "402": "wn",
+  "403": "ors",
+  "404": "▁jo",
+  "405": "ents",
+  "406": "▁des",
+  "407": "▁try",
+  "408": "▁equ",
+  "409": "▁z",
+  "410": "▁rem",
+  "411": "▁str",
+  "412": "self",
+  "413": "▁bit",
+  "414": "ph",
+  "415": "ved",
+  "416": "▁why",
+  "417": "▁bas",
+  "418": "▁hel",
+  "419": "▁rel",
+  "420": "ath",
+  "421": "ject",
+  "422": "ail",
+  "423": "▁la",
+  "424": "ual",
+  "425": "▁god",
+  "426": "▁nat",
+  "427": "erm",
+  "428": "day",
+  "429": "▁id",
+  "430": "ft",
+  "431": "▁wr",
+  "432": "▁min",
+  "433": "ates",
+  "434": "▁gen",
+  "435": "tain",
+  "436": "▁ob",
+  "437": "ull",
+  "438": "ict",
+  "439": "▁tra",
+  "440": "▁end",
+  "441": "▁hig",
+  "442": "▁fif",
+  "443": "oth",
+  "444": "tern",
+  "445": "▁its",
+  "446": "vent",
+  "447": "▁sm",
+  "448": "ons",
+  "449": "▁add",
+  "450": "iss",
+  "451": "▁bel",
+  "452": "ful",
+  "453": "get",
+  "454": "▁ele",
+  "455": "▁rep",
+  "456": "ak",
+  "457": "▁ho",
+  "458": "▁pos",
+  "459": "▁num",
+  "460": "ange",
+  "461": "ves",
+  "462": "ific",
+  "463": "urn",
+  "464": "ise",
+  "465": "▁cr",
+  "466": "▁um",
+  "467": "ward",
+  "468": "▁reg",
+  "469": "ady",
+  "470": "ower",
+  "471": "uc",
+  "472": "▁dec",
+  "473": "lic",
+  "474": "▁set",
+  "475": "▁gon",
+  "476": "▁op",
+  "477": "▁ear",
+  "478": "▁sub",
+  "479": "▁sl",
+  "480": "les",
+  "481": "stem",
+  "482": "cial",
+  "483": "olog",
+  "484": "atch",
+  "485": "ily",
+  "486": "body",
+  "487": "nds",
+  "488": "ular",
+  "489": "ren",
+  "490": "▁own",
+  "491": "▁too",
+  "492": "cent",
+  "493": "ible",
+  "494": "pect",
+  "495": "ered",
+  "496": "ways",
+  "497": "teen",
+  "498": "▁uh",
+  "499": "▁big",
+  "500": "▁mod",
+  "501": "▁att",
+  "502": "▁car",
+  "503": "gr",
+  "504": "▁acc",
+  "505": "ied",
+  "506": "mun",
+  "507": "ib",
+  "508": "▁mon",
+  "509": "▁sch",
+  "510": "▁pol",
+  "511": "▁dat",
+  "512": "▁fin",
+  "513": "▁sim",
+  "514": "▁inv",
+  "515": "▁def",
+  "516": "ked",
+  "517": "▁ent",
+  "518": "▁yes",
+  "519": "ows",
+  "520": "ics",
+  "521": "ited",
+  "522": "ute",
+  "523": "ism",
+  "524": "ps",
+  "525": "▁ed",
+  "526": "▁el",
+  "527": "ably",
+  "528": "ppen",
+  "529": "als",
+  "530": "▁ten",
+  "531": "ract",
+  "532": "ss",
+  "533": "▁ass",
+  "534": "▁met",
+  "535": "gan",
+  "536": "▁eng",
+  "537": "▁stu",
+  "538": "ween",
+  "539": "arch",
+  "540": "▁gl",
+  "541": "▁cor",
+  "542": "▁dr",
+  "543": "vern",
+  "544": "▁ty",
+  "545": "▁run",
+  "546": "hip",
+  "547": "cus",
+  "548": "cond",
+  "549": "▁ins",
+  "550": "irty",
+  "551": "▁pub",
+  "552": "lud",
+  "553": "llow",
+  "554": "▁cou",
+  "555": "ew",
+  "556": "iew",
+  "557": "▁sur",
+  "558": "ero",
+  "559": "ood",
+  "560": "ness",
+  "561": "▁fun",
+  "562": "▁eff",
+  "563": "cept",
+  "564": "▁ca",
+  "565": "▁exp",
+  "566": "duct",
+  "567": "▁sw",
+  "568": "ize",
+  "569": "ope",
+  "570": "▁par",
+  "571": "kes",
+  "572": "cy",
+  "573": "▁ev",
+  "574": "▁ref",
+  "575": "ell",
+  "576": "▁bus",
+  "577": "ug",
+  "578": "rib",
+  "579": "▁cur",
+  "580": "mo",
+  "581": "ock",
+  "582": "ures",
+  "583": "air",
+  "584": "▁war",
+  "585": "str",
+  "586": "▁med",
+  "587": "▁wa",
+  "588": "▁val",
+  "589": "▁sin",
+  "590": "blem",
+  "591": "▁fam",
+  "592": "li",
+  "593": "▁far",
+  "594": "▁cle",
+  "595": "▁col",
+  "596": "mon",
+  "597": "▁gra",
+  "598": "led",
+  "599": "ense",
+  "600": "tin",
+  "601": "ues",
+  "602": "its",
+  "603": "▁mem",
+  "604": "▁inf",
+  "605": "▁eas",
+  "606": "ideo",
+  "607": "▁top",
+  "608": "io",
+  "609": "pan",
+  "610": "▁hum",
+  "611": "▁old",
+  "612": "ead",
+  "613": "▁ord",
+  "614": "ric",
+  "615": "ants",
+  "616": "oy",
+  "617": "esn",
+  "618": "uck",
+  "619": "ason",
+  "620": "ced",
+  "621": "ool",
+  "622": "rat",
+  "623": "ouse",
+  "624": "▁lar",
+  "625": "▁art",
+  "626": "▁wee",
+  "627": "▁cer",
+  "628": "ized",
+  "629": "▁mat",
+  "630": "con",
+  "631": "erg",
+  "632": "land",
+  "633": "ines",
+  "634": "▁chr",
+  "635": "▁aut",
+  "636": "▁lea",
+  "637": "▁sou",
+  "638": "oney",
+  "639": "tty",
+  "640": "▁ple",
+  "641": "ulat",
+  "642": "oks",
+  "643": "▁few",
+  "644": "▁sol",
+  "645": "▁che",
+  "646": "chn",
+  "647": "ird",
+  "648": "▁bre",
+  "649": "▁dur",
+  "650": "▁wom",
+  "651": "me",
+  "652": "izat",
+  "653": "eric",
+  "654": "ote",
+  "655": "▁uni",
+  "656": "eren",
+  "657": "arn",
+  "658": "ross",
+  "659": "ices",
+  "660": "ten",
+  "661": "eral",
+  "662": "ever",
+  "663": "ieve",
+  "664": "lish",
+  "665": "ash",
+  "666": "▁opp",
+  "667": "alth",
+  "668": "ger",
+  "669": "▁sk",
+  "670": "▁red",
+  "671": "peri",
+  "672": "▁det",
+  "673": "▁ext",
+  "674": "ner",
+  "675": "ah",
+  "676": "▁var",
+  "677": "▁loc",
+  "678": "gram",
+  "679": "ists",
+  "680": "ives",
+  "681": "▁es",
+  "682": "▁nor",
+  "683": "tro",
+  "684": "ale",
+  "685": "▁iss",
+  "686": "▁pri",
+  "687": "gin",
+  "688": "az",
+  "689": "oc",
+  "690": "▁pop",
+  "691": "ern",
+  "692": "▁sit",
+  "693": "ket",
+  "694": "▁pa",
+  "695": "▁law",
+  "696": "ages",
+  "697": "br",
+  "698": "▁cam",
+  "699": "▁mom",
+  "700": "osed",
+  "701": "▁bro",
+  "702": "ne",
+  "703": "bs",
+  "704": "▁cre",
+  "705": "erat",
+  "706": "▁sec",
+  "707": "▁cap",
+  "708": "▁vis",
+  "709": "▁pat",
+  "710": "ield",
+  "711": "iet",
+  "712": "▁tri",
+  "713": "up",
+  "714": "▁bra",
+  "715": "ts",
+  "716": "▁mot",
+  "717": "▁unt",
+  "718": "put",
+  "719": "bo",
+  "720": "ork",
+  "721": "mer",
+  "722": "ital",
+  "723": "▁air",
+  "724": "ined",
+  "725": "▁beh",
+  "726": "▁adv",
+  "727": "▁ret",
+  "728": "imes",
+  "729": "▁tea",
+  "730": "ural",
+  "731": "sid",
+  "732": "ters",
+  "733": "▁pur",
+  "734": "▁sci",
+  "735": "bers",
+  "736": "ient",
+  "737": "ier",
+  "738": "cc",
+  "739": "sw",
+  "740": "▁av",
+  "741": "reen",
+  "742": "ode",
+  "743": "ont",
+  "744": "▁dra",
+  "745": "ann",
+  "746": "nect",
+  "747": "▁x",
+  "748": "▁eu",
+  "749": "ton",
+  "750": "inat",
+  "751": "ene",
+  "752": "ared",
+  "753": "els",
+  "754": "▁mor",
+  "755": "▁rat",
+  "756": "cri",
+  "757": "▁men",
+  "758": "▁ah",
+  "759": "ames",
+  "760": "▁arm",
+  "761": "eak",
+  "762": "▁pay",
+  "763": "▁hal",
+  "764": "ins",
+  "765": "ilit",
+  "766": "stit",
+  "767": "▁ra",
+  "768": "▁leg",
+  "769": "cl",
+  "770": "pr",
+  "771": "▁wal",
+  "772": "▁bad",
+  "773": "▁ge",
+  "774": "roup",
+  "775": "▁mus",
+  "776": "man",
+  "777": "▁gi",
+  "778": "eds",
+  "779": "▁aw",
+  "780": "po",
+  "781": "ark",
+  "782": "row",
+  "783": "▁dep",
+  "784": "ully",
+  "785": "ral",
+  "786": "lect",
+  "787": "pend",
+  "788": "▁sev",
+  "789": "ime",
+  "790": "gest",
+  "791": "here",
+  "792": "▁yet",
+  "793": "ted",
+  "794": "▁rev",
+  "795": "ds",
+  "796": "▁ask",
+  "797": "less",
+  "798": "▁di",
+  "799": "ets",
+  "800": "line",
+  "801": "▁aff",
+  "802": "ired",
+  "803": "▁est",
+  "804": "ken",
+  "805": "vid",
+  "806": "most",
+  "807": "ivid",
+  "808": "unch",
+  "809": "par",
+  "810": "med",
+  "811": "rop",
+  "812": "ased",
+  "813": "eone",
+  "814": "▁ve",
+  "815": "▁abs",
+  "816": "ergy",
+  "817": "ret",
+  "818": "▁saw",
+  "819": "▁ey",
+  "820": "▁cal",
+  "821": "uat",
+  "822": "▁mid",
+  "823": "vat",
+  "824": "ream",
+  "825": "vice",
+  "826": "ians",
+  "827": "rent",
+  "828": "ctor",
+  "829": "err",
+  "830": "ush",
+  "831": "ases",
+  "832": "▁suc",
+  "833": "erms",
+  "834": "ave",
+  "835": "angu",
+  "836": "ries",
+  "837": "▁wo",
+  "838": "arts",
+  "839": "▁fil",
+  "840": "▁fat",
+  "841": "▁cho",
+  "842": "orts",
+  "843": "▁fre",
+  "844": "ee",
+  "845": "ught",
+  "846": "eng",
+  "847": "ump",
+  "848": "▁bar",
+  "849": "ying",
+  "850": "ane",
+  "851": "▁tem",
+  "852": "anks",
+  "853": "ury",
+  "854": "iat",
+  "855": "mit",
+  "856": "trol",
+  "857": "▁net",
+  "858": "▁maj",
+  "859": "▁cra",
+  "860": "ling",
+  "861": "▁fig",
+  "862": "orn",
+  "863": "icat",
+  "864": "pany",
+  "865": "▁occ",
+  "866": "ott",
+  "867": "ands",
+  "868": "▁exc",
+  "869": "▁mr",
+  "870": "ency",
+  "871": "rope",
+  "872": "itch",
+  "873": "▁lit",
+  "874": "abil",
+  "875": "not",
+  "876": "ma",
+  "877": "▁typ",
+  "878": "▁opt",
+  "879": "ob",
+  "880": "ser",
+  "881": "ety",
+  "882": "ms",
+  "883": "peci",
+  "884": "aces",
+  "885": "aut",
+  "886": "▁hon",
+  "887": "cuss",
+  "888": "▁sal",
+  "889": "▁sor",
+  "890": "att",
+  "891": "▁lab",
+  "892": "▁har",
+  "893": "urch",
+  "894": "nded",
+  "895": "uce",
+  "896": "ids",
+  "897": "▁hy",
+  "898": "▁fut",
+  "899": "▁ste",
+  "900": "ours",
+  "901": "ems",
+  "902": "utes",
+  "903": "ng",
+  "904": "ta",
+  "905": "▁won",
+  "906": "▁fa",
+  "907": "▁env",
+  "908": "ards",
+  "909": "▁job",
+  "910": "ium",
+  "911": "▁dot",
+  "912": "▁obv",
+  "913": "ina",
+  "914": "side",
+  "915": "elve",
+  "916": "cu",
+  "917": "▁jes",
+  "918": "▁pot",
+  "919": "▁pie",
+  "920": "▁tre",
+  "921": "▁hey",
+  "922": "▁mag",
+  "923": "ron",
+  "924": "▁key",
+  "925": "swer",
+  "926": "▁win",
+  "927": "ucat",
+  "928": "work",
+  "929": "ides",
+  "930": "▁low",
+  "931": "▁vol",
+  "932": "▁oth",
+  "933": "atic",
+  "934": "lf",
+  "935": "ads",
+  "936": "inds",
+  "937": "com",
+  "938": "ths",
+  "939": "▁ver",
+  "940": "ised",
+  "941": "lo",
+  "942": "▁squ",
+  "943": "▁cut",
+  "944": "oked",
+  "945": "irit",
+  "946": "ateg",
+  "947": "ppy",
+  "948": "mitt",
+  "949": "come",
+  "950": "hn",
+  "951": "igin",
+  "952": "mand",
+  "953": "▁dam",
+  "954": "ho",
+  "955": "▁da",
+  "956": "▁fur",
+  "957": "iron",
+  "958": "ilar",
+  "959": "▁fac",
+  "960": "▁neg",
+  "961": "▁ago",
+  "962": "ged",
+  "963": "miss",
+  "964": "enth",
+  "965": "▁dou",
+  "966": "▁hit",
+  "967": "▁guy",
+  "968": "▁bi",
+  "969": "ove",
+  "970": "fess",
+  "971": "ples",
+  "972": "owed",
+  "973": "ured",
+  "974": "▁ris",
+  "975": "ints",
+  "976": "rew",
+  "977": "▁sum",
+  "978": "▁hu",
+  "979": "ploy",
+  "980": "ude",
+  "981": "ried",
+  "982": "▁cir",
+  "983": "▁dev",
+  "984": "ear",
+  "985": "▁tot",
+  "986": "▁ann",
+  "987": "duc",
+  "988": "ik",
+  "989": "pon",
+  "990": "sted",
+  "991": "▁ide",
+  "992": "▁'",
+  "993": "ipp",
+  "994": "▁eat",
+  "995": "▁dom",
+  "996": "▁",
+  "997": "e",
+  "998": "t",
+  "999": "o",
+  "1000": "a",
+  "1001": "i",
+  "1002": "n",
+  "1003": "s",
+  "1004": "r",
+  "1005": "h",
+  "1006": "l",
+  "1007": "d",
+  "1008": "u",
+  "1009": "c",
+  "1010": "m",
+  "1011": "y",
+  "1012": "g",
+  "1013": "w",
+  "1014": "f",
+  "1015": "p",
+  "1016": "b",
+  "1017": "v",
+  "1018": "k",
+  "1019": "'",
+  "1020": "j",
+  "1021": "x",
+  "1022": "q",
+  "1023": "z",
+  "1024": "<EOU>",
+  "1025": "<EOB>"
+}
\ No newline at end of file
diff --git a/160ms/.DS_Store b/160ms/.DS_Store
new file mode 100644
index 0000000000000000000000000000000000000000..142c5d02ad1e55affd5c417a2d8abf266ebbfc38
Binary files /dev/null and b/160ms/.DS_Store differ
diff --git a/160ms/convert_parakeet_eou.py b/160ms/convert_parakeet_eou.py
new file mode 100644
index 0000000000000000000000000000000000000000..9c865f90c8132840b5755dcd62fcfac15e3f0352
--- /dev/null
+++ b/160ms/convert_parakeet_eou.py
@@ -0,0 +1,740 @@
+#!/usr/bin/env python3
+"""CLI for exporting Parakeet Realtime EOU 120M components to CoreML.
+
+This model is a cache-aware streaming FastConformer-RNNT model optimized for
+low-latency speech recognition with end-of-utterance detection.
+
+Key differences from Parakeet TDT v3:
+- Smaller model (120M vs 600M params)
+- No duration outputs (standard RNNT, not TDT)
+- Cache-aware streaming encoder (17 layers, attention context [70,1])
+- Special <EOU> token for end-of-utterance detection
+- Optimized for 80-160ms latency
+
+Reference: https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1
+"""
+from __future__ import annotations
+
+import json
+from dataclasses import asdict
+from pathlib import Path
+from typing import Dict, Optional, Tuple
+
+import coremltools as ct
+import numpy as np
+import soundfile as sf
+import torch
+import typer
+
+import nemo.collections.asr as nemo_asr
+
+from individual_components import (
+    DecoderWrapper,
+    EncoderWrapper,
+    ExportSettings,
+    JointWrapper,
+    JointDecisionWrapper,
+    JointDecisionSingleStep,
+    PreprocessorWrapper,
+    MelEncoderWrapper,
+    _coreml_convert,
+)
+
+def apply_stft_patch():
+    # Monkey patch coremltools.stft to handle extra arguments from newer torch versions
+    try:
+        import coremltools.converters.mil.frontend.torch.ops as torch_ops
+        _original_stft = torch_ops.stft
+
+        def patched_stft(context, node):
+            if len(node.inputs) > 8:
+                node.inputs = node.inputs[:8]
+            return _original_stft(context, node)
+
+        torch_ops.stft = patched_stft
+        if "stft" in torch_ops._TORCH_OPS_REGISTRY:
+            torch_ops._TORCH_OPS_REGISTRY["stft"] = patched_stft
+        print("Monkey patched coremltools.stft for compatibility.")
+    except Exception as e:
+        print(f"Warning: Could not monkey patch stft: {e}")
+
+DEFAULT_MODEL_ID = "nvidia/parakeet_realtime_eou_120m-v1"
+AUTHOR = "Fluid Inference"
+
+
+def _compute_length(seconds: float, sample_rate: int) -> int:
+    return int(round(seconds * sample_rate))
+
+
+def _prepare_audio(
+    validation_audio: Optional[Path],
+    sample_rate: int,
+    max_samples: int,
+    seed: Optional[int],
+) -> torch.Tensor:
+    if validation_audio is None:
+        if seed is not None:
+            torch.manual_seed(seed)
+        audio = torch.randn(1, max_samples, dtype=torch.float32)
+        return audio
+
+    data, sr = sf.read(str(validation_audio), dtype="float32")
+    if sr != sample_rate:
+        raise typer.BadParameter(
+            f"Validation audio sample rate {sr} does not match model rate {sample_rate}"
+        )
+
+    if data.ndim > 1:
+        data = data[:, 0]
+
+    if data.size == 0:
+        raise typer.BadParameter("Validation audio is empty")
+
+    if data.size < max_samples:
+        pad_width = max_samples - data.size
+        data = np.pad(data, (0, pad_width))
+    elif data.size > max_samples:
+        data = data[:max_samples]
+
+    audio = torch.from_numpy(data).unsqueeze(0).to(dtype=torch.float32)
+    return audio
+
+
+def _save_mlpackage(model: ct.models.MLModel, path: Path, description: str) -> None:
+    try:
+        model.minimum_deployment_target = ct.target.iOS17
+    except Exception:
+        pass
+    model.short_description = description
+    model.author = AUTHOR
+    path.parent.mkdir(parents=True, exist_ok=True)
+    model.save(str(path))
+
+
+def _tensor_shape(tensor: torch.Tensor) -> Tuple[int, ...]:
+    return tuple(int(dim) for dim in tensor.shape)
+
+
+def _parse_compute_units(name: str) -> ct.ComputeUnit:
+    """Parse a human-friendly compute units string into ct.ComputeUnit."""
+    normalized = str(name).strip().upper()
+    mapping = {
+        "ALL": ct.ComputeUnit.ALL,
+        "CPU_ONLY": ct.ComputeUnit.CPU_ONLY,
+        "CPU_AND_GPU": ct.ComputeUnit.CPU_AND_GPU,
+        "CPU_AND_NE": ct.ComputeUnit.CPU_AND_NE,
+        "CPU_AND_NEURALENGINE": ct.ComputeUnit.CPU_AND_NE,
+    }
+    if normalized not in mapping:
+        raise typer.BadParameter(
+            f"Unknown compute units '{name}'. Choose from: " + ", ".join(mapping.keys())
+        )
+    return mapping[normalized]
+
+
+def _parse_compute_precision(name: Optional[str]) -> Optional[ct.precision]:
+    """Parse compute precision string into ct.precision or None."""
+    if name is None:
+        return None
+    normalized = str(name).strip().upper()
+    if normalized == "":
+        return None
+    mapping = {
+        "FLOAT32": ct.precision.FLOAT32,
+        "FLOAT16": ct.precision.FLOAT16,
+    }
+    if normalized not in mapping:
+        raise typer.BadParameter(
+            f"Unknown compute precision '{name}'. Choose from: "
+            + ", ".join(mapping.keys())
+        )
+    return mapping[normalized]
+
+
+app = typer.Typer(add_completion=False, pretty_exceptions_show_locals=False)
+
+
+@app.command()
+def convert(
+    nemo_path: Optional[Path] = typer.Option(
+        None,
+        "--nemo-path",
+        exists=True,
+        resolve_path=True,
+        help="Path to parakeet_realtime_eou_120m-v1.nemo checkpoint (skip to auto-download)",
+    ),
+    model_id: str = typer.Option(
+        DEFAULT_MODEL_ID,
+        "--model-id",
+        help="Model identifier to download when --nemo-path is omitted",
+    ),
+    output_dir: Path = typer.Option(
+        Path("parakeet_eou_coreml"),
+        help="Directory where mlpackages and metadata will be written",
+    ),
+    preprocessor_cu: str = typer.Option(
+        "CPU_ONLY",
+        "--preprocessor-cu",
+        help="Compute units for preprocessor (default CPU_ONLY)",
+    ),
+    mel_encoder_cu: str = typer.Option(
+        "CPU_ONLY",
+        "--mel-encoder-cu",
+        help="Compute units for fused mel+encoder (default CPU_ONLY)",
+    ),
+    compute_precision: Optional[str] = typer.Option(
+        None,
+        "--compute-precision",
+        help="Export precision: FLOAT32 (default) or FLOAT16 to shrink non-quantized weights.",
+    ),
+    max_audio_seconds: float = typer.Option(
+        15.0,
+        "--max-audio-seconds",
+        help="Maximum audio duration in seconds for the fixed window export",
+    ),
+    validation_audio: Optional[Path] = typer.Option(
+        None,
+        "--validation-audio",
+        exists=True,
+        resolve_path=True,
+        help="Path to a 16kHz WAV file for tracing (uses random if not provided)",
+    ),
+) -> None:
+    """Export all Parakeet Realtime EOU sub-modules to CoreML.
+
+    This exports the cache-aware streaming FastConformer-RNNT model for
+    low-latency speech recognition with end-of-utterance detection.
+    """
+    export_settings = ExportSettings(
+        output_dir=output_dir,
+        compute_units=ct.ComputeUnit.CPU_ONLY,
+        deployment_target=ct.target.iOS17,
+        compute_precision=_parse_compute_precision(compute_precision),
+        max_audio_seconds=max_audio_seconds,
+        max_symbol_steps=1,
+    )
+
+    typer.echo("Export configuration:")
+    typer.echo(asdict(export_settings))
+
+    output_dir.mkdir(parents=True, exist_ok=True)
+    pre_cu = _parse_compute_units(preprocessor_cu)
+    melenc_cu = _parse_compute_units(mel_encoder_cu)
+
+    if nemo_path is not None:
+        typer.echo(f"Loading NeMo model from {nemo_path}…")
+        # Try loading as generic ASRModel first, then specific class
+        try:
+            asr_model = nemo_asr.models.ASRModel.restore_from(
+                str(nemo_path), map_location="cpu"
+            )
+        except Exception:
+            # Fallback to EncDecRNNTBPEModel
+            asr_model = nemo_asr.models.EncDecRNNTBPEModel.restore_from(
+                str(nemo_path), map_location="cpu"
+            )
+        checkpoint_meta = {
+            "type": "file",
+            "path": str(nemo_path),
+        }
+    else:
+        typer.echo(f"Downloading NeMo model via {model_id}…")
+        # Use ASRModel.from_pretrained as recommended for this model
+        try:
+            asr_model = nemo_asr.models.ASRModel.from_pretrained(
+                model_id, map_location="cpu"
+            )
+        except Exception:
+            # Fallback to EncDecRNNTBPEModel
+            asr_model = nemo_asr.models.EncDecRNNTBPEModel.from_pretrained(
+                model_id, map_location="cpu"
+            )
+        checkpoint_meta = {
+            "type": "pretrained",
+            "model_id": model_id,
+        }
+    asr_model.eval()
+
+    # Print model info
+    typer.echo(f"Model class: {type(asr_model).__name__}")
+    typer.echo(f"Encoder class: {type(asr_model.encoder).__name__}")
+
+    sample_rate = int(asr_model.cfg.preprocessor.sample_rate)
+    max_samples = _compute_length(export_settings.max_audio_seconds, sample_rate)
+
+    # Prepare audio for tracing
+    if validation_audio is not None:
+        typer.echo(f"Using validation audio: {validation_audio}")
+        audio_tensor = _prepare_audio(validation_audio, sample_rate, max_samples, seed=None)
+    else:
+        typer.echo("Using random audio for tracing (seed=42)")
+        audio_tensor = _prepare_audio(None, sample_rate, max_samples, seed=42)
+
+    audio_length = torch.tensor([max_samples], dtype=torch.int32)
+
+    preprocessor = PreprocessorWrapper(asr_model.preprocessor.eval())
+    encoder = EncoderWrapper(asr_model.encoder.eval())
+    decoder = DecoderWrapper(asr_model.decoder.eval())
+    joint = JointWrapper(asr_model.joint.eval())
+
+    decoder_export_flag = getattr(asr_model.decoder, "_rnnt_export", False)
+    asr_model.decoder._rnnt_export = True
+
+    try:
+        with torch.no_grad():
+            mel_ref, mel_length_ref = preprocessor(audio_tensor, audio_length)
+            mel_length_ref = mel_length_ref.to(dtype=torch.int32)
+            encoder_ref, encoder_length_ref, frame_times_ref = encoder(
+                mel_ref, mel_length_ref
+            )
+            encoder_length_ref = encoder_length_ref.to(dtype=torch.int32)
+
+            # Clone tensors to drop inference flags
+            mel_ref = mel_ref.clone().detach()
+            mel_length_ref = mel_length_ref.clone().detach()
+            encoder_ref = encoder_ref.clone().detach()
+            encoder_length_ref = encoder_length_ref.clone().detach()
+            frame_times_ref = frame_times_ref.clone().detach()
+
+        vocab_size = int(asr_model.tokenizer.vocab_size)
+        decoder_hidden = int(asr_model.decoder.pred_hidden)
+        decoder_layers = int(asr_model.decoder.pred_rnn_layers)
+
+        # Check if model has extra outputs (TDT-style duration)
+        num_extra = getattr(asr_model.joint, "num_extra_outputs", 0)
+        typer.echo(f"Vocab size: {vocab_size}, num_extra_outputs: {num_extra}")
+
+        targets = torch.full(
+            (1, export_settings.max_symbol_steps),
+            fill_value=asr_model.decoder.blank_idx,
+            dtype=torch.int32,
+        )
+        target_lengths = torch.tensor(
+            [export_settings.max_symbol_steps], dtype=torch.int32
+        )
+        zero_state = torch.zeros(
+            decoder_layers,
+            1,
+            decoder_hidden,
+            dtype=torch.float32,
+        )
+
+        with torch.no_grad():
+            decoder_ref, h_ref, c_ref = decoder(
+                targets, target_lengths, zero_state, zero_state
+            )
+            joint_ref = joint(encoder_ref, decoder_ref)
+
+        decoder_ref = decoder_ref.clone()
+        h_ref = h_ref.clone()
+        c_ref = c_ref.clone()
+        joint_ref = joint_ref.clone()
+
+        typer.echo(f"Encoder output shape: {encoder_ref.shape}")
+        typer.echo(f"Decoder output shape: {decoder_ref.shape}")
+        typer.echo(f"Joint output shape: {joint_ref.shape}")
+
+        # === Export Preprocessor ===
+        typer.echo("Tracing and converting preprocessor…")
+        preprocessor = preprocessor.cpu()
+        audio_tensor = audio_tensor.cpu()
+        audio_length = audio_length.cpu()
+        traced_preprocessor = torch.jit.trace(
+            preprocessor, (audio_tensor, audio_length), strict=False
+        )
+        traced_preprocessor.eval()
+        preprocessor_inputs = [
+            ct.TensorType(
+                name="audio_signal",
+                shape=(1, ct.RangeDim(1, max_samples)),
+                dtype=np.float32,
+            ),
+            ct.TensorType(name="audio_length", shape=(1,), dtype=np.int32),
+        ]
+        preprocessor_outputs = [
+            ct.TensorType(name="mel", dtype=np.float32),
+            ct.TensorType(name="mel_length", dtype=np.int32),
+        ]
+        preprocessor_model = _coreml_convert(
+            traced_preprocessor,
+            preprocessor_inputs,
+            preprocessor_outputs,
+            export_settings,
+            compute_units_override=pre_cu,
+        )
+        preprocessor_path = output_dir / "parakeet_eou_preprocessor.mlpackage"
+        _save_mlpackage(
+            preprocessor_model,
+            preprocessor_path,
+            f"Parakeet EOU preprocessor ({max_audio_seconds}s window)",
+        )
+
+        # === Export Encoder ===
+        typer.echo("Tracing and converting encoder…")
+        traced_encoder = torch.jit.trace(
+            encoder, (mel_ref, mel_length_ref), strict=False
+        )
+        traced_encoder.eval()
+        encoder_inputs = [
+            ct.TensorType(
+                name="mel", shape=_tensor_shape(mel_ref), dtype=np.float32
+            ),
+            ct.TensorType(name="mel_length", shape=(1,), dtype=np.int32),
+        ]
+        encoder_outputs = [
+            ct.TensorType(name="encoder", dtype=np.float32),
+            ct.TensorType(name="encoder_length", dtype=np.int32),
+            ct.TensorType(name="frame_times", dtype=np.float32),
+        ]
+        encoder_model = _coreml_convert(
+            traced_encoder,
+            encoder_inputs,
+            encoder_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        encoder_path = output_dir / "parakeet_eou_encoder.mlpackage"
+        _save_mlpackage(
+            encoder_model,
+            encoder_path,
+            f"Parakeet EOU encoder ({max_audio_seconds}s window)",
+        )
+
+        # === Export Fused Mel+Encoder ===
+        typer.echo("Tracing and converting fused mel+encoder…")
+        mel_encoder = MelEncoderWrapper(preprocessor, encoder)
+        traced_mel_encoder = torch.jit.trace(
+            mel_encoder, (audio_tensor, audio_length), strict=False
+        )
+        traced_mel_encoder.eval()
+        mel_encoder_inputs = [
+            ct.TensorType(
+                name="audio_signal", shape=(1, max_samples), dtype=np.float32
+            ),
+            ct.TensorType(name="audio_length", shape=(1,), dtype=np.int32),
+        ]
+        mel_encoder_outputs = [
+            ct.TensorType(name="encoder", dtype=np.float32),
+            ct.TensorType(name="encoder_length", dtype=np.int32),
+            ct.TensorType(name="frame_times", dtype=np.float32),
+        ]
+        mel_encoder_model = _coreml_convert(
+            traced_mel_encoder,
+            mel_encoder_inputs,
+            mel_encoder_outputs,
+            export_settings,
+            compute_units_override=melenc_cu,
+        )
+        mel_encoder_path = output_dir / "parakeet_eou_mel_encoder.mlpackage"
+        _save_mlpackage(
+            mel_encoder_model,
+            mel_encoder_path,
+            f"Parakeet EOU fused Mel+Encoder ({max_audio_seconds}s window)",
+        )
+
+        # === Export Decoder ===
+        typer.echo("Tracing and converting decoder…")
+        traced_decoder = torch.jit.trace(
+            decoder,
+            (targets, target_lengths, zero_state, zero_state),
+            strict=False,
+        )
+        traced_decoder.eval()
+        decoder_inputs = [
+            ct.TensorType(
+                name="targets", shape=_tensor_shape(targets), dtype=np.int32
+            ),
+            ct.TensorType(name="target_length", shape=(1,), dtype=np.int32),
+            ct.TensorType(
+                name="h_in", shape=_tensor_shape(zero_state), dtype=np.float32
+            ),
+            ct.TensorType(
+                name="c_in", shape=_tensor_shape(zero_state), dtype=np.float32
+            ),
+        ]
+        decoder_outputs = [
+            ct.TensorType(name="decoder", dtype=np.float32),
+            ct.TensorType(name="h_out", dtype=np.float32),
+            ct.TensorType(name="c_out", dtype=np.float32),
+        ]
+        decoder_model = _coreml_convert(
+            traced_decoder,
+            decoder_inputs,
+            decoder_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        decoder_path = output_dir / "parakeet_eou_decoder.mlpackage"
+        _save_mlpackage(
+            decoder_model,
+            decoder_path,
+            "Parakeet EOU decoder (RNNT prediction network)",
+        )
+
+        # === Export Joint ===
+        typer.echo("Tracing and converting joint…")
+        traced_joint = torch.jit.trace(
+            joint,
+            (encoder_ref, decoder_ref),
+            strict=False,
+        )
+        traced_joint.eval()
+        joint_inputs = [
+            ct.TensorType(
+                name="encoder", shape=_tensor_shape(encoder_ref), dtype=np.float32
+            ),
+            ct.TensorType(
+                name="decoder", shape=_tensor_shape(decoder_ref), dtype=np.float32
+            ),
+        ]
+        joint_outputs = [
+            ct.TensorType(name="logits", dtype=np.float32),
+        ]
+        joint_model = _coreml_convert(
+            traced_joint,
+            joint_inputs,
+            joint_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        joint_path = output_dir / "parakeet_eou_joint.mlpackage"
+        _save_mlpackage(
+            joint_model,
+            joint_path,
+            "Parakeet EOU joint network (RNNT)",
+        )
+
+        # === Export Joint Decision Head ===
+        typer.echo("Tracing and converting joint decision head…")
+        joint_decision = JointDecisionWrapper(joint, vocab_size=vocab_size)
+        traced_joint_decision = torch.jit.trace(
+            joint_decision,
+            (encoder_ref, decoder_ref),
+            strict=False,
+        )
+        traced_joint_decision.eval()
+        joint_decision_inputs = [
+            ct.TensorType(
+                name="encoder", shape=_tensor_shape(encoder_ref), dtype=np.float32
+            ),
+            ct.TensorType(
+                name="decoder", shape=_tensor_shape(decoder_ref), dtype=np.float32
+            ),
+        ]
+        joint_decision_outputs = [
+            ct.TensorType(name="token_id", dtype=np.int32),
+            ct.TensorType(name="token_prob", dtype=np.float32),
+        ]
+        joint_decision_model = _coreml_convert(
+            traced_joint_decision,
+            joint_decision_inputs,
+            joint_decision_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        joint_decision_path = output_dir / "parakeet_eou_joint_decision.mlpackage"
+        _save_mlpackage(
+            joint_decision_model,
+            joint_decision_path,
+            "Parakeet EOU joint + decision head (softmax, argmax)",
+        )
+
+        # === Export Single-Step Joint Decision ===
+        typer.echo("Tracing and converting single-step joint decision…")
+        jd_single = JointDecisionSingleStep(joint, vocab_size=vocab_size)
+        # Create single-step slices from refs
+        enc_step = encoder_ref[:, :, :1].contiguous()
+        dec_step = decoder_ref[:, :, :1].contiguous()
+        traced_jd_single = torch.jit.trace(
+            jd_single,
+            (enc_step, dec_step),
+            strict=False,
+        )
+        traced_jd_single.eval()
+        jd_single_inputs = [
+            ct.TensorType(
+                name="encoder_step",
+                shape=(1, enc_step.shape[1], 1),
+                dtype=np.float32,
+            ),
+            ct.TensorType(
+                name="decoder_step",
+                shape=(1, dec_step.shape[1], 1),
+                dtype=np.float32,
+            ),
+        ]
+        jd_single_outputs = [
+            ct.TensorType(name="token_id", dtype=np.int32),
+            ct.TensorType(name="token_prob", dtype=np.float32),
+            ct.TensorType(name="top_k_ids", dtype=np.int32),
+            ct.TensorType(name="top_k_logits", dtype=np.float32),
+        ]
+        jd_single_model = _coreml_convert(
+            traced_jd_single,
+            jd_single_inputs,
+            jd_single_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        jd_single_path = output_dir / "parakeet_eou_joint_decision_single_step.mlpackage"
+        _save_mlpackage(
+            jd_single_model,
+            jd_single_path,
+            "Parakeet EOU single-step joint decision (current frame)",
+        )
+
+        # === Save Metadata ===
+        metadata: Dict[str, object] = {
+            "model_id": model_id,
+            "model_name": "parakeet_realtime_eou_120m-v1",
+            "model_class": type(asr_model).__name__,
+            "encoder_class": type(asr_model.encoder).__name__,
+            "sample_rate": sample_rate,
+            "max_audio_seconds": export_settings.max_audio_seconds,
+            "max_audio_samples": max_samples,
+            "max_symbol_steps": export_settings.max_symbol_steps,
+            "vocab_size": vocab_size,
+            "vocab_with_blank": vocab_size + 1,
+            "decoder_hidden": decoder_hidden,
+            "decoder_layers": decoder_layers,
+            "num_extra_outputs": num_extra,
+            "has_eou_token": True,
+            "checkpoint": checkpoint_meta,
+            "coreml": {
+                "compute_units": export_settings.compute_units.name,
+                "compute_precision": (
+                    export_settings.compute_precision.name
+                    if export_settings.compute_precision is not None
+                    else "FLOAT32"
+                ),
+            },
+            "components": {
+                "preprocessor": {
+                    "inputs": {
+                        "audio_signal": [1, max_samples],
+                        "audio_length": [1],
+                    },
+                    "outputs": {
+                        "mel": list(_tensor_shape(mel_ref)),
+                        "mel_length": [1],
+                    },
+                    "path": preprocessor_path.name,
+                },
+                "encoder": {
+                    "inputs": {
+                        "mel": list(_tensor_shape(mel_ref)),
+                        "mel_length": [1],
+                    },
+                    "outputs": {
+                        "encoder": list(_tensor_shape(encoder_ref)),
+                        "encoder_length": [1],
+                        "frame_times": [1, _tensor_shape(encoder_ref)[2]],
+                    },
+                    "path": encoder_path.name,
+                },
+                "mel_encoder": {
+                    "inputs": {
+                        "audio_signal": [1, max_samples],
+                        "audio_length": [1],
+                    },
+                    "outputs": {
+                        "encoder": list(_tensor_shape(encoder_ref)),
+                        "encoder_length": [1],
+                        "frame_times": [1, _tensor_shape(encoder_ref)[2]],
+                    },
+                    "path": mel_encoder_path.name,
+                },
+                "decoder": {
+                    "inputs": {
+                        "targets": list(_tensor_shape(targets)),
+                        "target_length": [1],
+                        "h_in": list(_tensor_shape(zero_state)),
+                        "c_in": list(_tensor_shape(zero_state)),
+                    },
+                    "outputs": {
+                        "decoder": list(_tensor_shape(decoder_ref)),
+                        "h_out": list(_tensor_shape(h_ref)),
+                        "c_out": list(_tensor_shape(c_ref)),
+                    },
+                    "path": decoder_path.name,
+                },
+                "joint": {
+                    "inputs": {
+                        "encoder": list(_tensor_shape(encoder_ref)),
+                        "decoder": list(_tensor_shape(decoder_ref)),
+                    },
+                    "outputs": {
+                        "logits": list(_tensor_shape(joint_ref)),
+                    },
+                    "path": joint_path.name,
+                },
+                "joint_decision": {
+                    "inputs": {
+                        "encoder": list(_tensor_shape(encoder_ref)),
+                        "decoder": list(_tensor_shape(decoder_ref)),
+                    },
+                    "outputs": {
+                        "token_id": [
+                            _tensor_shape(encoder_ref)[0],
+                            _tensor_shape(encoder_ref)[2],
+                            _tensor_shape(decoder_ref)[2],
+                        ],
+                        "token_prob": [
+                            _tensor_shape(encoder_ref)[0],
+                            _tensor_shape(encoder_ref)[2],
+                            _tensor_shape(decoder_ref)[2],
+                        ],
+                    },
+                    "path": joint_decision_path.name,
+                },
+                "joint_decision_single_step": {
+                    "inputs": {
+                        "encoder_step": [1, _tensor_shape(encoder_ref)[1], 1],
+                        "decoder_step": [1, _tensor_shape(decoder_ref)[1], 1],
+                    },
+                    "outputs": {
+                        "token_id": [1, 1, 1],
+                        "token_prob": [1, 1, 1],
+                        "top_k_ids": [1, 1, 1, 64],
+                        "top_k_logits": [1, 1, 1, 64],
+                    },
+                    "path": jd_single_path.name,
+                },
+            },
+        }
+
+        # Export tokenizer vocab if available
+        try:
+            tokenizer = asr_model.tokenizer
+            vocab = {
+                "blank_id": int(asr_model.decoder.blank_idx),
+                "vocab_size": vocab_size,
+            }
+            # Try to get special tokens
+            if hasattr(tokenizer, "tokenizer"):
+                inner_tokenizer = tokenizer.tokenizer
+                if hasattr(inner_tokenizer, "get_vocab"):
+                    full_vocab = inner_tokenizer.get_vocab()
+                    # Find EOU token
+                    eou_token = None
+                    for token, idx in full_vocab.items():
+                        if "<EOU>" in token.upper() or "eou" in token.lower():
+                            eou_token = {"token": token, "id": idx}
+                            break
+                    if eou_token:
+                        vocab["eou_token"] = eou_token
+            metadata["tokenizer"] = vocab
+        except Exception as e:
+            typer.echo(f"Warning: Could not export tokenizer info: {e}")
+
+        metadata_path = output_dir / "metadata.json"
+        metadata_path.write_text(json.dumps(metadata, indent=2))
+        typer.echo(f"\nExport complete. Metadata written to {metadata_path}")
+        typer.echo(f"Output directory: {output_dir}")
+
+    finally:
+        asr_model.decoder._rnnt_export = decoder_export_flag
+
+
+if __name__ == "__main__":
+    app()
diff --git a/160ms/convert_streaming_encoder.py b/160ms/convert_streaming_encoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..a1c199c8a6c540dab48b973ab4232db2901b8e72
--- /dev/null
+++ b/160ms/convert_streaming_encoder.py
@@ -0,0 +1,193 @@
+
+import torch
+import torch.nn as nn
+import coremltools as ct
+import numpy as np
+import typer
+from pathlib import Path
+from typing import Tuple, List, Optional
+import json
+import shutil
+
+# Iimport torch
+import coremltools as ct
+import numpy as np
+import argparse
+from nemo.collections.asr.models import EncDecRNNTBPEModel
+
+app = typer.Typer()
+
+class LoopbackEncoderWrapper(nn.Module):
+    """
+    Wraps the entire Parakeet Encoder (PreEncode + Conformer) for CoreML Loopback Streaming.
+    
+    Inputs:
+      - audio_signal: [B, D, T] (Mel spectrogram chunk)
+      - audio_length: [B]
+      - pre_cache: [B, D, pre_cache_size] (Previous audio context)
+      - cache_last_channel: [layers, B, cache_size, hidden]
+      - cache_last_time: [layers, B, hidden, time_cache]
+      - cache_last_channel_len: [B]
+      
+    Outputs:
+      - encoded_output: [B, D_out, T_out]
+      - encoded_length: [B]
+      - new_pre_cache: [B, D, pre_cache_size]
+      - new_cache_last_channel
+      - new_cache_last_time
+      - new_cache_last_channel_len
+    """
+    def __init__(self, encoder, pre_cache_size=16):
+        super().__init__()
+        self.encoder = encoder
+        self.pre_cache_size = pre_cache_size
+        
+    def forward(
+        self, 
+        audio_signal: torch.Tensor, 
+        audio_length: torch.Tensor,
+        pre_cache: torch.Tensor,
+        cache_last_channel: torch.Tensor,
+        cache_last_time: torch.Tensor,
+        cache_last_channel_len: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        
+        # 1. Prepend pre_cache to audio_signal
+        # audio_signal: [B, D, T]
+        # pre_cache: [B, D, T_cache]
+        full_input = torch.cat([pre_cache, audio_signal], dim=2)
+        full_length = audio_length + self.pre_cache_size
+        
+        # 2. Extract NEW pre_cache (last N frames of full_input)
+        # Note: We do this BEFORE processing because we want the raw audio context
+        new_pre_cache = full_input[:, :, -self.pre_cache_size:]
+        
+        # 3. Process with Encoder
+        # Reconstruct NeMo cache object
+        current_cache = [cache_last_channel, cache_last_time, cache_last_channel_len]
+        
+        encoded, encoded_len, new_cache_channel, new_cache_time, new_cache_len = self.encoder.cache_aware_stream_step(
+            processed_signal=full_input,
+            processed_signal_length=full_length,
+            cache_last_channel=cache_last_channel,
+            cache_last_time=cache_last_time,
+            cache_last_channel_len=cache_last_channel_len
+        )
+        
+        # 4. Drop the first few frames corresponding to pre_cache?
+        # NeMo's cache_aware_stream_step usually handles the "valid" output frames.
+        # But since we manually prepended, we might get extra output frames.
+        # However, for streaming, we usually want the model to see the context but only output the new tokens.
+        # Let's trust NeMo's streaming logic for now, or check if we need to slice.
+        # Given we are using 'cache_aware_stream_step', it expects the full context window?
+        # Actually, standard usage is: input IS the new chunk, but internal convolution looks at past.
+        # But since we are stateless, we MUST provide the past.
+        # So passing (pre_cache + chunk) is correct.
+        
+        # Cast lengths to Int32 for CoreML
+        encoded_len_32 = encoded_len.to(dtype=torch.int32)
+        new_channel_len_32 = new_cache_len.to(dtype=torch.int32)
+        
+        return encoded, encoded_len_32, new_pre_cache, new_cache_channel, new_cache_time, new_channel_len_32
+
+def _coreml_convert(
+    traced_model,
+    inputs,
+    outputs,
+    compute_units=ct.ComputeUnit.CPU_ONLY
+):
+    return ct.convert(
+        traced_model,
+        inputs=inputs,
+        outputs=outputs,
+        compute_units=compute_units,
+        minimum_deployment_target=ct.target.macOS14,
+    )
+
+def main():
+    model_id: str = "nvidia/parakeet_realtime_eou_120m-v1"
+    output_dir: str = "temp_swift_models/StreamingLoopback"
+    output_path = Path(output_dir)
+    output_path.mkdir(parents=True, exist_ok=True)
+    
+    print(f"Loading model: {model_id}...")
+    asr_model = EncDecRNNTBPEModel.from_pretrained(model_name=model_id)
+    asr_model.eval()
+    
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--chunk-frames", type=int, default=17, help="Number of frames in the input chunk (e.g. 17 for 160ms, 129 for 1.28s)")
+    args = parser.parse_args()
+
+    encoder = asr_model.encoder
+    
+    # --- Configuration ---
+    # 160ms chunk = 16 frames (but preprocessor produces 17 with padding/centering)
+    # 1.28s chunk = 128 frames (preprocessor produces 129)
+    chunk_size_in = args.chunk_frames 
+    mel_dim = 128
+    hidden_dim = encoder.d_model # 512
+    num_layers = len(encoder.layers) # 17
+    
+    # Cache sizes
+    cache_channel_size = 70
+    cache_time_size = 8
+    pre_cache_size = 16
+    
+    print(f"Config: Chunk={chunk_size_in}, Mel={mel_dim}, Hidden={hidden_dim}, Layers={num_layers}")
+    print(f"Cache: Channel={cache_channel_size}, Time={cache_time_size}, Pre={pre_cache_size}")
+
+    # --- Wrapper ---
+    wrapper = LoopbackEncoderWrapper(encoder, pre_cache_size=pre_cache_size)
+    wrapper.eval()
+    
+    # --- Test Inputs (for Tracing) ---
+    batch_size = 1
+    test_mel = torch.randn(batch_size, mel_dim, chunk_size_in)
+    test_mel_len = torch.tensor([chunk_size_in], dtype=torch.int32)
+    test_pre_cache = torch.zeros(batch_size, mel_dim, pre_cache_size)
+    
+    # Initial Cache (Zeros)
+    test_cache_channel = torch.zeros(num_layers, batch_size, cache_channel_size, hidden_dim)
+    test_cache_time = torch.zeros(num_layers, batch_size, hidden_dim, cache_time_size)
+    test_cache_len = torch.zeros(batch_size, dtype=torch.int32) 
+    
+    print("Tracing model...")
+    traced_model = torch.jit.trace(
+        wrapper,
+        (test_mel, test_mel_len, test_pre_cache, test_cache_channel, test_cache_time, test_cache_len),
+        strict=False
+    )
+    
+    # --- CoreML Conversion ---
+    print("Converting to CoreML...")
+    
+    inputs = [
+        ct.TensorType(name="audio_signal", shape=(1, 128, chunk_size_in), dtype=np.float32),
+        ct.TensorType(name="audio_length", shape=(1,), dtype=np.int32),
+        ct.TensorType(name="pre_cache", shape=(1, 128, pre_cache_size), dtype=np.float32),
+        ct.TensorType(name="cache_last_channel", shape=(num_layers, 1, cache_channel_size, hidden_dim), dtype=np.float32),
+        ct.TensorType(name="cache_last_time", shape=(num_layers, 1, hidden_dim, cache_time_size), dtype=np.float32),
+        ct.TensorType(name="cache_last_channel_len", shape=(1,), dtype=np.int32),
+    ]
+    
+    outputs = [
+        ct.TensorType(name="encoded_output", dtype=np.float32),
+        ct.TensorType(name="encoded_length", dtype=np.int32),
+        ct.TensorType(name="new_pre_cache", dtype=np.float32),
+        ct.TensorType(name="new_cache_last_channel", dtype=np.float32),
+        ct.TensorType(name="new_cache_last_time", dtype=np.float32),
+        ct.TensorType(name="new_cache_last_channel_len", dtype=np.int32),
+    ]
+    
+    mlmodel = _coreml_convert(traced_model, inputs, outputs)
+    
+    save_path = output_path / "streaming_encoder.mlpackage"
+    mlmodel.save(str(save_path))
+    print(f"Saved: {save_path}")
+    
+    # Also export Preprocessor, Decoder, Joint for completeness? 
+    # For now, let's assume we reuse the existing ones or export them separately if needed.
+    # But the user asked specifically for the Encoder loopback.
+
+if __name__ == "__main__":
+    main()
diff --git a/160ms/decoder.mlmodelc/analytics/coremldata.bin b/160ms/decoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..c6f51dedd2638184c5a2d71512339ceea086abc0
--- /dev/null
+++ b/160ms/decoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3996975a8cbc1949159c55605b3132b39b2484f51acbd55d796d93c70de02b49
+size 243
diff --git a/160ms/decoder.mlmodelc/coremldata.bin b/160ms/decoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..71936e61ffe7697e4af55bae8870d95c1fbb66c5
--- /dev/null
+++ b/160ms/decoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3ccbff963d8cf07e2be2bd56ea3384a89ea49628922c6bd95ff62e2ae57dc34
+size 497
diff --git a/160ms/decoder.mlmodelc/metadata.json b/160ms/decoder.mlmodelc/metadata.json
new file mode 100644
index 0000000000000000000000000000000000000000..3581673bcb062f7e0e1e0efa9f59edc5f218d50e
--- /dev/null
+++ b/160ms/decoder.mlmodelc/metadata.json
@@ -0,0 +1,118 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Parakeet EOU decoder (RNNT prediction network)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 640 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 640, 1]",
+        "name" : "decoder",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 640]",
+        "name" : "h_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 640]",
+        "name" : "c_out",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.squeeze" : 2,
+      "Ios17.gather" : 1,
+      "Ios17.cast" : 6,
+      "Ios17.lstm" : 1,
+      "Ios17.transpose" : 2,
+      "Identity" : 1,
+      "Ios17.expandDims" : 2
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int16, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1]",
+        "name" : "targets",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "target_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 640]",
+        "name" : "h_in",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 640]",
+        "name" : "c_in",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.4.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "parakeet_eou_decoder",
+    "method" : "predict"
+  }
+]
\ No newline at end of file
diff --git a/160ms/decoder.mlmodelc/model.mil b/160ms/decoder.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..5ead417fbdc13f8bf0c05a26039b2828503d4eca
--- /dev/null
+++ b/160ms/decoder.mlmodelc/model.mil
@@ -0,0 +1,45 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.4.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<fp32, [1, 1, 640]> c_in, tensor<fp32, [1, 1, 640]> h_in, tensor<int32, [1]> target_length, tensor<int32, [1, 1]> targets) {
+            tensor<int32, []> y_axis_0 = const()[name = tensor<string, []>("y_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> y_batch_dims_0 = const()[name = tensor<string, []>("y_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> y_validate_indices_0 = const()[name = tensor<string, []>("y_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1027, 640]> module_prediction_embed_weight_to_fp16 = const()[name = tensor<string, []>("module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [1027, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<string, []> targets_to_int16_dtype_0 = const()[name = tensor<string, []>("targets_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 1]> targets_to_int16 = cast(dtype = targets_to_int16_dtype_0, x = targets)[name = tensor<string, []>("cast_8")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = targets_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = tensor<string, []>("y_cast_fp16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [1]> input_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<string, []> h_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("h_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = tensor<string, []>("cast_7")];
+            tensor<fp16, [1, 640]> input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = h_in_to_fp16)[name = tensor<string, []>("input_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<string, []> c_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("c_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = tensor<string, []>("cast_6")];
+            tensor<fp16, [1, 640]> input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = c_in_to_fp16)[name = tensor<string, []>("input_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> input_direction_0 = const()[name = tensor<string, []>("input_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_output_sequence_0 = const()[name = tensor<string, []>("input_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_recurrent_activation_0 = const()[name = tensor<string, []>("input_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_cell_activation_0 = const()[name = tensor<string, []>("input_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_activation_0 = const()[name = tensor<string, []>("input_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = tensor<string, []>("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1314688)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = tensor<string, []>("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4591552)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = tensor<string, []>("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7868416)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_2")];
+            tensor<fp16, [1, 1, 640]> input_cast_fp16_0, tensor<fp16, [1, 640]> input_cast_fp16_1, tensor<fp16, [1, 640]> input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_0_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [1]> obj_3_axes_0 = const()[name = tensor<string, []>("obj_3_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 640]> obj_3_cast_fp16 = expand_dims(axes = obj_3_axes_0, x = input_cast_fp16_1)[name = tensor<string, []>("obj_3_cast_fp16")];
+            tensor<string, []> obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_3_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> obj_axes_0 = const()[name = tensor<string, []>("obj_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 640]> obj_cast_fp16 = expand_dims(axes = obj_axes_0, x = input_cast_fp16_2)[name = tensor<string, []>("obj_cast_fp16")];
+            tensor<string, []> obj_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = tensor<string, []>("transpose_0_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
+            tensor<string, []> transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("transpose_0_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp16, [1, 640, 1]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = tensor<string, []>("transpose_1")];
+            tensor<fp32, [1, 640, 1]> decoder = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = tensor<string, []>("cast_3")];
+            tensor<fp32, [1, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = tensor<string, []>("cast_4")];
+            tensor<fp32, [1, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = tensor<string, []>("cast_5")];
+            tensor<int32, [1]> target_length_tmp = identity(x = target_length)[name = tensor<string, []>("target_length_tmp")];
+        } -> (decoder, h_out, c_out);
+}
\ No newline at end of file
diff --git a/160ms/decoder.mlmodelc/weights/weight.bin b/160ms/decoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..cde1618cbed235b421984a302289e3bdd7e3df02
--- /dev/null
+++ b/160ms/decoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b4cacecdcd9df79ab1e56de67230baf5a8664d2afe0bb8f3408eefa972cb2f4
+size 7873600
diff --git a/160ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/160ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..0631f4ecb012ab42d13312be15475aff506055c5
--- /dev/null
+++ b/160ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:09f2dbd1f6a06faa6995f71d4b25d7c446996b6059cfac5ecc889853bdc7c6e5
+size 6728
diff --git a/160ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/160ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..cde1618cbed235b421984a302289e3bdd7e3df02
--- /dev/null
+++ b/160ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b4cacecdcd9df79ab1e56de67230baf5a8664d2afe0bb8f3408eefa972cb2f4
+size 7873600
diff --git a/160ms/decoder.mlpackage/Manifest.json b/160ms/decoder.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..4e498499a0a4026095bbc46d4e1a47ecbc39b0ca
--- /dev/null
+++ b/160ms/decoder.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "8201D73A-2B5D-488C-9C2B-7E2E75DF700D": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "F8EEBE8D-F17D-4556-B8DF-9BC11701B36D": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "F8EEBE8D-F17D-4556-B8DF-9BC11701B36D"
+}
diff --git a/160ms/individual_components.py b/160ms/individual_components.py
new file mode 100644
index 0000000000000000000000000000000000000000..47271397bc8d9d17cc0fabcf4bb63be7e7c2109c
--- /dev/null
+++ b/160ms/individual_components.py
@@ -0,0 +1,250 @@
+#!/usr/bin/env python3
+"""Export Parakeet Realtime EOU 120M RNNT components into CoreML."""
+from __future__ import annotations
+
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Optional, Tuple
+
+import coremltools as ct
+import torch
+
+
+@dataclass
+class ExportSettings:
+    output_dir: Path
+    compute_units: ct.ComputeUnit
+    deployment_target: Optional[ct.target]
+    compute_precision: Optional[ct.precision]
+    max_audio_seconds: float
+    max_symbol_steps: int
+
+
+class PreprocessorWrapper(torch.nn.Module):
+    """Wrapper for the audio preprocessor (mel spectrogram extraction)."""
+
+    def __init__(self, module: torch.nn.Module) -> None:
+        super().__init__()
+        self.module = module
+
+    def forward(
+        self, audio_signal: torch.Tensor, length: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        mel, mel_length = self.module(
+            input_signal=audio_signal, length=length.to(dtype=torch.long)
+        )
+        return mel, mel_length
+
+
+class EncoderWrapper(torch.nn.Module):
+    """Wrapper for the cache-aware FastConformer encoder.
+
+    Note: For the realtime EOU model, the encoder is cache-aware which means
+    it can operate in a streaming fashion. For CoreML export, we export
+    without cache state for simplicity (full-context mode).
+    """
+
+    def __init__(self, module: torch.nn.Module) -> None:
+        super().__init__()
+        self.module = module
+
+    def forward(
+        self, features: torch.Tensor, length: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        encoded, encoded_lengths = self.module(
+            audio_signal=features, length=length.to(dtype=torch.long)
+        )
+        # Synthesize per-frame timestamps (seconds) using the 80 ms encoder stride.
+        # Shape: [B, T_enc]
+        frame_times = (
+            torch.arange(encoded.shape[-1], device=encoded.device, dtype=torch.float32)
+            * 0.08
+        )
+        return encoded, encoded_lengths, frame_times
+
+
+class DecoderWrapper(torch.nn.Module):
+    """Wrapper for the RNNT prediction network (decoder)."""
+
+    def __init__(self, module: torch.nn.Module) -> None:
+        super().__init__()
+        self.module = module
+
+    def forward(
+        self,
+        targets: torch.Tensor,
+        target_lengths: torch.Tensor,
+        h_in: torch.Tensor,
+        c_in: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        state = [h_in, c_in]
+        decoder_output, _, new_state = self.module(
+            targets=targets.to(dtype=torch.long),
+            target_length=target_lengths.to(dtype=torch.long),
+            states=state,
+        )
+        return decoder_output, new_state[0], new_state[1]
+
+
+class JointWrapper(torch.nn.Module):
+    """Wrapper for the RNNT joint network.
+
+    Note: Unlike Parakeet TDT v3, the realtime EOU model does NOT have
+    duration outputs (num_extra_outputs). The joint network outputs only
+    token logits over the vocabulary + blank.
+    """
+
+    def __init__(self, module: torch.nn.Module) -> None:
+        super().__init__()
+        self.module = module
+
+    def forward(
+        self, encoder_outputs: torch.Tensor, decoder_outputs: torch.Tensor
+    ) -> torch.Tensor:
+        # Input: encoder_outputs [B, D, T], decoder_outputs [B, D, U]
+        # Transpose to match what projection layers expect
+        encoder_outputs = encoder_outputs.transpose(1, 2)  # [B, T, D]
+        decoder_outputs = decoder_outputs.transpose(1, 2)  # [B, U, D]
+
+        # Apply projections
+        enc_proj = self.module.enc(encoder_outputs)  # [B, T, joint_hidden]
+        dec_proj = self.module.pred(decoder_outputs)  # [B, U, joint_hidden]
+
+        # Explicit broadcasting along T and U to avoid converter ambiguity
+        x = enc_proj.unsqueeze(2) + dec_proj.unsqueeze(1)  # [B, T, U, joint_hidden]
+        x = self.module.joint_net[0](x)  # ReLU
+        x = self.module.joint_net[1](x)  # Dropout (no-op in eval)
+        out = self.module.joint_net[2](x)  # Linear -> logits [B, T, U, vocab+blank]
+        return out
+
+
+class MelEncoderWrapper(torch.nn.Module):
+    """Fused wrapper: waveform -> mel -> encoder.
+
+    Inputs:
+      - audio_signal: [B, S]
+      - audio_length: [B]
+
+    Outputs:
+      - encoder: [B, D, T_enc]
+      - encoder_length: [B]
+      - frame_times: [T_enc]
+    """
+
+    def __init__(
+        self, preprocessor: PreprocessorWrapper, encoder: EncoderWrapper
+    ) -> None:
+        super().__init__()
+        self.preprocessor = preprocessor
+        self.encoder = encoder
+
+    def forward(
+        self, audio_signal: torch.Tensor, audio_length: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        mel, mel_length = self.preprocessor(audio_signal, audio_length)
+        encoded, enc_len, frame_times = self.encoder(mel, mel_length.to(dtype=torch.int32))
+        return encoded, enc_len, frame_times
+
+
+class JointDecisionWrapper(torch.nn.Module):
+    """Joint + decision head: outputs label id and label prob.
+
+    Unlike Parakeet TDT v3, this model does NOT have duration outputs.
+
+    Inputs:
+      - encoder_outputs: [B, D, T]
+      - decoder_outputs: [B, D, U]
+
+    Returns:
+      - token_id: [B, T, U] int32
+      - token_prob: [B, T, U] float32
+    """
+
+    def __init__(self, joint: JointWrapper, vocab_size: int) -> None:
+        super().__init__()
+        self.joint = joint
+        self.vocab_with_blank = int(vocab_size) + 1
+
+    def forward(self, encoder_outputs: torch.Tensor, decoder_outputs: torch.Tensor):
+        logits = self.joint(encoder_outputs, decoder_outputs)
+
+        # Token selection
+        token_ids = torch.argmax(logits, dim=-1).to(dtype=torch.int32)
+        token_probs_all = torch.softmax(logits, dim=-1)
+        # gather expects int64 (long) indices; cast only for gather
+        token_prob = torch.gather(
+            token_probs_all, dim=-1, index=token_ids.long().unsqueeze(-1)
+        ).squeeze(-1)
+
+        return token_ids, token_prob
+
+
+class JointDecisionSingleStep(torch.nn.Module):
+    """Single-step variant for streaming: encoder_step -> token decision.
+
+    Inputs:
+      - encoder_step: [B=1, D, T=1]
+      - decoder_step: [B=1, D, U=1]
+
+    Returns:
+      - token_id: [1, 1, 1] int32
+      - token_prob: [1, 1, 1] float32
+      - top_k_ids: [1, 1, 1, K] int32
+      - top_k_logits: [1, 1, 1, K] float32
+    """
+
+    def __init__(self, joint: JointWrapper, vocab_size: int, top_k: int = 64) -> None:
+        super().__init__()
+        self.joint = joint
+        self.vocab_with_blank = int(vocab_size) + 1
+        self.top_k = int(top_k)
+
+    def forward(self, encoder_step: torch.Tensor, decoder_step: torch.Tensor):
+        # Reuse JointWrapper which expects [B, D, T] and [B, D, U]
+        logits = self.joint(encoder_step, decoder_step)  # [1, 1, 1, V+blank]
+
+        token_ids = torch.argmax(logits, dim=-1, keepdim=False).to(dtype=torch.int32)
+        token_probs_all = torch.softmax(logits, dim=-1)
+        token_prob = torch.gather(
+            token_probs_all, dim=-1, index=token_ids.long().unsqueeze(-1)
+        ).squeeze(-1)
+
+        # Also expose top-K candidates for host-side processing
+        topk_logits, topk_ids_long = torch.topk(
+            logits, k=min(self.top_k, logits.shape[-1]), dim=-1
+        )
+        topk_ids = topk_ids_long.to(dtype=torch.int32)
+        return token_ids, token_prob, topk_ids, topk_logits
+
+
+def _coreml_convert(
+    traced: torch.jit.ScriptModule,
+    inputs,
+    outputs,
+    settings: ExportSettings,
+    compute_units_override: Optional[ct.ComputeUnit] = None,
+    compute_precision: Optional[ct.precision] = None,
+) -> ct.models.MLModel:
+    cu = (
+        compute_units_override
+        if compute_units_override is not None
+        else settings.compute_units
+    )
+    kwargs = {
+        "convert_to": "mlprogram",
+        "inputs": inputs,
+        "outputs": outputs,
+        "compute_units": cu,
+    }
+    print("Converting:", traced.__class__.__name__)
+    print("Conversion kwargs:", kwargs)
+    if settings.deployment_target is not None:
+        kwargs["minimum_deployment_target"] = settings.deployment_target
+    
+    # Priority: explicit argument > settings
+    if compute_precision is not None:
+        kwargs["compute_precision"] = compute_precision
+    elif settings.compute_precision is not None:
+        kwargs["compute_precision"] = settings.compute_precision
+        
+    return ct.convert(traced, **kwargs)
diff --git a/160ms/joint_decision.mlmodelc/analytics/coremldata.bin b/160ms/joint_decision.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..ff9929e12bae25c509655bda0ebda6b9fe13fade
--- /dev/null
+++ b/160ms/joint_decision.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5bca32ad130dcad6605cc00044c752aa5b45ef57d14c17f2d1a2fa49d6cf55b5
+size 243
diff --git a/160ms/joint_decision.mlmodelc/coremldata.bin b/160ms/joint_decision.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..1aa5afe7c02e93f9baba2757cad5c21f563157bf
--- /dev/null
+++ b/160ms/joint_decision.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22d4abc4625b935ee035b5f8ce7cb28d1041b9b01c12173e287bf4b5f5d99625
+size 493
diff --git a/160ms/joint_decision.mlmodelc/metadata.json b/160ms/joint_decision.mlmodelc/metadata.json
new file mode 100644
index 0000000000000000000000000000000000000000..a0a319857f429f5ec4b106397cd52ec354ef2e96
--- /dev/null
+++ b/160ms/joint_decision.mlmodelc/metadata.json
@@ -0,0 +1,112 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Parakeet EOU single-step joint decision",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1]",
+        "name" : "token_id",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1]",
+        "name" : "token_prob",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1 × 1 × 64)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1, 64]",
+        "name" : "top_k_ids",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 1 × 64)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1, 64]",
+        "name" : "top_k_logits",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.reduceArgmax" : 1,
+      "Ios17.squeeze" : 1,
+      "Ios17.cast" : 6,
+      "Ios17.linear" : 3,
+      "Ios17.transpose" : 2,
+      "Ios17.add" : 1,
+      "Ios16.relu" : 1,
+      "Ios16.softmax" : 1,
+      "Ios17.gatherAlongAxis" : 1,
+      "Ios17.topk" : 1,
+      "Ios17.expandDims" : 3
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int16, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 1]",
+        "name" : "encoder_step",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 640 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 640, 1]",
+        "name" : "decoder_step",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.4.0"
+    },
+    "generatedClassName" : "parakeet_eou_joint_decision_single_step",
+    "method" : "predict"
+  }
+]
\ No newline at end of file
diff --git a/160ms/joint_decision.mlmodelc/model.mil b/160ms/joint_decision.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..172a7579866c84407b32b5f746b7f1ae132599d8
--- /dev/null
+++ b/160ms/joint_decision.mlmodelc/model.mil
@@ -0,0 +1,57 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.4.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<fp32, [1, 640, 1]> decoder_step, tensor<fp32, [1, 512, 1]> encoder_step) {
+            tensor<int32, [3]> input_1_perm_0 = const()[name = tensor<string, []>("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> encoder_step_to_fp16_dtype_0 = const()[name = tensor<string, []>("encoder_step_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> decoder_step_to_fp16_dtype_0 = const()[name = tensor<string, []>("decoder_step_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [640, 512]> joint_module_enc_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_enc_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [640]> joint_module_enc_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(655488)))];
+            tensor<fp16, [1, 512, 1]> encoder_step_to_fp16 = cast(dtype = encoder_step_to_fp16_dtype_0, x = encoder_step)[name = tensor<string, []>("cast_8")];
+            tensor<fp16, [1, 1, 512]> input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_step_to_fp16)[name = tensor<string, []>("transpose_1")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(656832)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1476096)))];
+            tensor<fp16, [1, 640, 1]> decoder_step_to_fp16 = cast(dtype = decoder_step_to_fp16_dtype_0, x = decoder_step)[name = tensor<string, []>("cast_7")];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_step_to_fp16)[name = tensor<string, []>("transpose_0")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_23_axes_0 = const()[name = tensor<string, []>("op_23_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_23_cast_fp16")];
+            tensor<int32, [1]> var_24_axes_0 = const()[name = tensor<string, []>("op_24_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_24_cast_fp16 = expand_dims(axes = var_24_axes_0, x = linear_1_cast_fp16)[name = tensor<string, []>("op_24_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_24_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<fp16, [1027, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [1027, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1477440)))];
+            tensor<fp16, [1027]> joint_module_joint_net_2_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [1027]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2792064)))];
+            tensor<fp16, [1, 1, 1, 1027]> linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<int32, []> var_38_axis_0 = const()[name = tensor<string, []>("op_38_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> var_38_keep_dims_0 = const()[name = tensor<string, []>("op_38_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> var_38_output_dtype_0 = const()[name = tensor<string, []>("op_38_output_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 1, 1]> token_id = reduce_argmax(axis = var_38_axis_0, keep_dims = var_38_keep_dims_0, output_dtype = var_38_output_dtype_0, x = linear_2_cast_fp16)[name = tensor<string, []>("op_38_cast_fp16")];
+            tensor<int32, []> var_44 = const()[name = tensor<string, []>("op_44"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [1, 1, 1, 1027]> token_probs_all_cast_fp16 = softmax(axis = var_44, x = linear_2_cast_fp16)[name = tensor<string, []>("token_probs_all_cast_fp16")];
+            tensor<int32, [1]> var_53_axes_0 = const()[name = tensor<string, []>("op_53_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 1, 1, 1]> var_53 = expand_dims(axes = var_53_axes_0, x = token_id)[name = tensor<string, []>("op_53")];
+            tensor<int32, []> var_54 = const()[name = tensor<string, []>("op_54"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> var_56_validate_indices_0 = const()[name = tensor<string, []>("op_56_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> var_53_to_int16_dtype_0 = const()[name = tensor<string, []>("op_53_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 1, 1, 1]> var_53_to_int16 = cast(dtype = var_53_to_int16_dtype_0, x = var_53)[name = tensor<string, []>("cast_6")];
+            tensor<fp16, [1, 1, 1, 1]> var_56_cast_fp16_cast_int16 = gather_along_axis(axis = var_54, indices = var_53_to_int16, validate_indices = var_56_validate_indices_0, x = token_probs_all_cast_fp16)[name = tensor<string, []>("op_56_cast_fp16_cast_int16")];
+            tensor<int32, [1]> var_58_axes_0 = const()[name = tensor<string, []>("op_58_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 1, 1]> var_58_cast_fp16 = squeeze(axes = var_58_axes_0, x = var_56_cast_fp16_cast_int16)[name = tensor<string, []>("op_58_cast_fp16")];
+            tensor<string, []> var_58_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_58_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, []> var_59 = const()[name = tensor<string, []>("op_59"), val = tensor<int32, []>(64)];
+            tensor<int32, []> var_63_axis_0 = const()[name = tensor<string, []>("op_63_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> var_63_ascending_0 = const()[name = tensor<string, []>("op_63_ascending_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_63_sort_0 = const()[name = tensor<string, []>("op_63_sort_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_63_return_indices_0 = const()[name = tensor<string, []>("op_63_return_indices_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> var_63_cast_fp16_cast_int16_output_indices_dtype_0 = const()[name = tensor<string, []>("op_63_cast_fp16_cast_int16_output_indices_dtype_0"), val = tensor<string, []>("uint16")];
+            tensor<fp16, [1, 1, 1, 64]> var_63_cast_fp16_cast_int16_0, tensor<uint16, [1, 1, 1, 64]> var_63_cast_fp16_cast_int16_1 = topk(ascending = var_63_ascending_0, axis = var_63_axis_0, k = var_59, output_indices_dtype = var_63_cast_fp16_cast_int16_output_indices_dtype_0, return_indices = var_63_return_indices_0, sort = var_63_sort_0, x = linear_2_cast_fp16)[name = tensor<string, []>("op_63_cast_fp16_cast_int16")];
+            tensor<string, []> var_63_cast_fp16_cast_int16_1_to_int32_dtype_0 = const()[name = tensor<string, []>("op_63_cast_fp16_cast_int16_1_to_int32_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<string, []> var_63_cast_fp16_0_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_63_cast_fp16_0_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 1, 1, 64]> top_k_logits = cast(dtype = var_63_cast_fp16_0_to_fp32_dtype_0, x = var_63_cast_fp16_cast_int16_0)[name = tensor<string, []>("cast_3")];
+            tensor<int32, [1, 1, 1, 64]> top_k_ids = cast(dtype = var_63_cast_fp16_cast_int16_1_to_int32_dtype_0, x = var_63_cast_fp16_cast_int16_1)[name = tensor<string, []>("cast_4")];
+            tensor<fp32, [1, 1, 1]> token_prob = cast(dtype = var_58_cast_fp16_to_fp32_dtype_0, x = var_58_cast_fp16)[name = tensor<string, []>("cast_5")];
+        } -> (token_id, token_prob, top_k_ids, top_k_logits);
+}
\ No newline at end of file
diff --git a/160ms/joint_decision.mlmodelc/weights/weight.bin b/160ms/joint_decision.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..ec3d2c1dea18b332fa1a3cd8d31981bae4e3f649
--- /dev/null
+++ b/160ms/joint_decision.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7039b2010a269153f5a96edf28637f921a86ef8822f248f2d6712f7a6bce84b4
+size 2794182
diff --git a/160ms/joint_decision.mlpackage/Data/com.apple.CoreML/model.mlmodel b/160ms/joint_decision.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..678be6fae0d594ff4bb923f973b0459139c8b370
--- /dev/null
+++ b/160ms/joint_decision.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25d4d7be6eeb60c7de1d3a1278a5a4700cbe34017e1a8c1cab33204ddb2e4d5e
+size 8701
diff --git a/160ms/joint_decision.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/160ms/joint_decision.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..ec3d2c1dea18b332fa1a3cd8d31981bae4e3f649
--- /dev/null
+++ b/160ms/joint_decision.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7039b2010a269153f5a96edf28637f921a86ef8822f248f2d6712f7a6bce84b4
+size 2794182
diff --git a/160ms/joint_decision.mlpackage/Manifest.json b/160ms/joint_decision.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..fffd90a77691d335c8153fd9c38d49adb725c949
--- /dev/null
+++ b/160ms/joint_decision.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "634E266B-4447-41D3-879E-F3611888F54B": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "C7F40527-180B-45CD-BC12-4F054F2E5D9A": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "C7F40527-180B-45CD-BC12-4F054F2E5D9A"
+}
diff --git a/160ms/parakeet_eou_preprocessor.mlmodelc/analytics/coremldata.bin b/160ms/parakeet_eou_preprocessor.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..ca93efecf9fbf12e37943f7e192cd324a39e9f94
--- /dev/null
+++ b/160ms/parakeet_eou_preprocessor.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4ada8b0b99ac1d2ba7acbffacfbbf1a06cb69d30e9410d237ee0aa4c2b0ad63
+size 243
diff --git a/160ms/parakeet_eou_preprocessor.mlmodelc/coremldata.bin b/160ms/parakeet_eou_preprocessor.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..ca69399b73b039b35bb1b4967b25b7a7a12e8178
--- /dev/null
+++ b/160ms/parakeet_eou_preprocessor.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc7252fa47622fe39577361233627062019a3bb740fdbb5366a7bae09df0ec5e
+size 422
diff --git a/160ms/parakeet_eou_preprocessor.mlmodelc/metadata.json b/160ms/parakeet_eou_preprocessor.mlmodelc/metadata.json
new file mode 100644
index 0000000000000000000000000000000000000000..98b70ffcae507184b68f87d7adaad706e5c68c9a
--- /dev/null
+++ b/160ms/parakeet_eou_preprocessor.mlmodelc/metadata.json
@@ -0,0 +1,105 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Parakeet EOU preprocessor",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "mel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "mel_length",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float32",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Range1d" : 1,
+      "Ios17.reshape" : 2,
+      "Identity" : 1,
+      "Ios17.matmul" : 1,
+      "Ios17.expandDims" : 5,
+      "Select" : 1,
+      "Ios17.add" : 3,
+      "Ios17.sliceByIndex" : 3,
+      "Ios16.reduceSum" : 1,
+      "Shape" : 1,
+      "Ios17.gather" : 1,
+      "Pad" : 1,
+      "Ios17.log" : 1,
+      "Ios17.conv" : 2,
+      "Ios17.sub" : 2,
+      "Ios17.pow" : 1,
+      "Ios17.cast" : 2,
+      "Stack" : 1,
+      "Ios17.concat" : 1,
+      "Ios17.floorDiv" : 1,
+      "Ios17.greaterEqual" : 1,
+      "Ios17.mul" : 1
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "dataType" : "Float32",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 1...32000",
+        "shapeRange" : "[[1, 1], [1, 32000]]",
+        "formattedType" : "MultiArray (Float32 1 × 1)",
+        "type" : "MultiArray",
+        "shape" : "[1, 1]",
+        "name" : "audio_signal",
+        "shortDescription" : ""
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "audio_length",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.4.0",
+      "com.github.apple.coremltools.version" : "8.3.0"
+    },
+    "generatedClassName" : "parakeet_eou_preprocessor",
+    "method" : "predict"
+  }
+]
\ No newline at end of file
diff --git a/160ms/parakeet_eou_preprocessor.mlmodelc/model.mil b/160ms/parakeet_eou_preprocessor.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..25db6a1374c1718340459f374949bfba2bff85bb
--- /dev/null
+++ b/160ms/parakeet_eou_preprocessor.mlmodelc/model.mil
@@ -0,0 +1,96 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.4.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<int32, [1]> audio_length, tensor<fp32, [1, ?]> audio_signal) [FlexibleShapeInformation = tuple<tuple<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>, tuple<tensor<string, []>, dict<tensor<string, []>, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"audio_signal", [1, 1]}}), ("RangeDims", {{"audio_signal", [[1, 1], [1, 32000]]}})))] {
+            tensor<int32, []> var_9 = const()[name = tensor<string, []>("op_9"), val = tensor<int32, []>(1)];
+            tensor<int32, []> var_10 = const()[name = tensor<string, []>("op_10"), val = tensor<int32, []>(160)];
+            tensor<int32, []> var_32 = const()[name = tensor<string, []>("op_32"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> var_33 = add(x = audio_length, y = var_32)[name = tensor<string, []>("op_33")];
+            tensor<int32, []> var_34 = const()[name = tensor<string, []>("op_34"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> var_35 = sub(x = var_33, y = var_34)[name = tensor<string, []>("op_35")];
+            tensor<int32, [1]> floor_div_0 = floor_div(x = var_35, y = var_10)[name = tensor<string, []>("floor_div_0")];
+            tensor<string, []> var_36_dtype_0 = const()[name = tensor<string, []>("op_36_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, []> var_37_promoted = const()[name = tensor<string, []>("op_37_promoted"), val = tensor<fp32, []>(0x1p+0)];
+            tensor<fp32, [1]> var_36 = cast(dtype = var_36_dtype_0, x = floor_div_0)[name = tensor<string, []>("cast_11")];
+            tensor<fp32, [1]> seq_len_1 = add(x = var_36, y = var_37_promoted)[name = tensor<string, []>("seq_len_1")];
+            tensor<string, []> cast_2_dtype_0 = const()[name = tensor<string, []>("cast_2_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [2]> var_41_begin_0 = const()[name = tensor<string, []>("op_41_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_41_end_0 = const()[name = tensor<string, []>("op_41_end_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<bool, [2]> var_41_end_mask_0 = const()[name = tensor<string, []>("op_41_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<bool, [2]> var_41_squeeze_mask_0 = const()[name = tensor<string, []>("op_41_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
+            tensor<fp32, [1]> var_41 = slice_by_index(begin = var_41_begin_0, end = var_41_end_0, end_mask = var_41_end_mask_0, squeeze_mask = var_41_squeeze_mask_0, x = audio_signal)[name = tensor<string, []>("op_41")];
+            tensor<int32, [1]> var_42_axes_0 = const()[name = tensor<string, []>("op_42_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1]> var_42 = expand_dims(axes = var_42_axes_0, x = var_41)[name = tensor<string, []>("op_42")];
+            tensor<int32, [2]> var_44_begin_0 = const()[name = tensor<string, []>("op_44_begin_0"), val = tensor<int32, [2]>([0, 1])];
+            tensor<int32, [2]> var_44_end_0 = const()[name = tensor<string, []>("op_44_end_0"), val = tensor<int32, [2]>([1, 0])];
+            tensor<bool, [2]> var_44_end_mask_0 = const()[name = tensor<string, []>("op_44_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<fp32, [1, ?]> var_44 = slice_by_index(begin = var_44_begin_0, end = var_44_end_0, end_mask = var_44_end_mask_0, x = audio_signal)[name = tensor<string, []>("op_44")];
+            tensor<int32, [2]> var_46_begin_0 = const()[name = tensor<string, []>("op_46_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_46_end_0 = const()[name = tensor<string, []>("op_46_end_0"), val = tensor<int32, [2]>([1, -1])];
+            tensor<bool, [2]> var_46_end_mask_0 = const()[name = tensor<string, []>("op_46_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp32, [1, ?]> var_46 = slice_by_index(begin = var_46_begin_0, end = var_46_end_0, end_mask = var_46_end_mask_0, x = audio_signal)[name = tensor<string, []>("op_46")];
+            tensor<fp32, []> var_47 = const()[name = tensor<string, []>("op_47"), val = tensor<fp32, []>(0x1.f0a3d8p-1)];
+            tensor<fp32, [1, ?]> var_48 = mul(x = var_46, y = var_47)[name = tensor<string, []>("op_48")];
+            tensor<fp32, [1, ?]> var_49 = sub(x = var_44, y = var_48)[name = tensor<string, []>("op_49")];
+            tensor<bool, []> input_1_interleave_0 = const()[name = tensor<string, []>("input_1_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, ?]> input_1 = concat(axis = var_9, interleave = input_1_interleave_0, values = (var_42, var_49))[name = tensor<string, []>("input_1")];
+            tensor<int32, [3]> concat_0x = const()[name = tensor<string, []>("concat_0x"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp32, [1, 1, ?]> input_3 = reshape(shape = concat_0x, x = input_1)[name = tensor<string, []>("input_3")];
+            tensor<fp32, []> const_1 = const()[name = tensor<string, []>("const_1"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [6]> input_5_pad_0 = const()[name = tensor<string, []>("input_5_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 256, 256])];
+            tensor<string, []> input_5_mode_0 = const()[name = tensor<string, []>("input_5_mode_0"), val = tensor<string, []>("reflect")];
+            tensor<fp32, [1, 1, ?]> input_5 = pad(constant_val = const_1, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3)[name = tensor<string, []>("input_5")];
+            tensor<int32, [2]> concat_1x = const()[name = tensor<string, []>("concat_1x"), val = tensor<int32, [2]>([1, -1])];
+            tensor<fp32, [1, ?]> input = reshape(shape = concat_1x, x = input_5)[name = tensor<string, []>("input")];
+            tensor<fp32, [257, 1, 512]> expand_dims_1 = const()[name = tensor<string, []>("expand_dims_1"), val = tensor<fp32, [257, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp32, [257, 1, 512]> expand_dims_2 = const()[name = tensor<string, []>("expand_dims_2"), val = tensor<fp32, [257, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526464)))];
+            tensor<int32, [1]> expand_dims_3 = const()[name = tensor<string, []>("expand_dims_3"), val = tensor<int32, [1]>([160])];
+            tensor<int32, [1]> expand_dims_4_axes_0 = const()[name = tensor<string, []>("expand_dims_4_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1, ?]> expand_dims_4 = expand_dims(axes = expand_dims_4_axes_0, x = input)[name = tensor<string, []>("expand_dims_4")];
+            tensor<string, []> conv_0_pad_type_0 = const()[name = tensor<string, []>("conv_0_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> conv_0_pad_0 = const()[name = tensor<string, []>("conv_0_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_0_dilations_0 = const()[name = tensor<string, []>("conv_0_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> conv_0_groups_0 = const()[name = tensor<string, []>("conv_0_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 257, ?]> conv_0 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1, x = expand_dims_4)[name = tensor<string, []>("conv_0")];
+            tensor<string, []> conv_1_pad_type_0 = const()[name = tensor<string, []>("conv_1_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> conv_1_pad_0 = const()[name = tensor<string, []>("conv_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_1_dilations_0 = const()[name = tensor<string, []>("conv_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> conv_1_groups_0 = const()[name = tensor<string, []>("conv_1_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 257, ?]> conv_1 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2, x = expand_dims_4)[name = tensor<string, []>("conv_1")];
+            tensor<int32, []> stack_0_axis_0 = const()[name = tensor<string, []>("stack_0_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<fp32, [1, 257, ?, 2]> stack_0 = stack(axis = stack_0_axis_0, values = (conv_0, conv_1))[name = tensor<string, []>("stack_0")];
+            tensor<fp32, []> var_17_promoted = const()[name = tensor<string, []>("op_17_promoted"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1, 257, ?, 2]> var_65 = pow(x = stack_0, y = var_17_promoted)[name = tensor<string, []>("op_65")];
+            tensor<int32, [1]> var_67_axes_0 = const()[name = tensor<string, []>("op_67_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> var_67_keep_dims_0 = const()[name = tensor<string, []>("op_67_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 257, ?]> var_67 = reduce_sum(axes = var_67_axes_0, keep_dims = var_67_keep_dims_0, x = var_65)[name = tensor<string, []>("op_67")];
+            tensor<fp32, [1, 257, ?]> x_9 = identity(x = var_67)[name = tensor<string, []>("x_9")];
+            tensor<fp32, [1, 128, 257]> const_2 = const()[name = tensor<string, []>("const_2"), val = tensor<fp32, [1, 128, 257]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1052864)))];
+            tensor<bool, []> x_11_transpose_x_0 = const()[name = tensor<string, []>("x_11_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_11_transpose_y_0 = const()[name = tensor<string, []>("x_11_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 128, ?]> x_11 = matmul(transpose_x = x_11_transpose_x_0, transpose_y = x_11_transpose_y_0, x = const_2, y = x_9)[name = tensor<string, []>("x_11")];
+            tensor<fp32, []> var_74 = const()[name = tensor<string, []>("op_74"), val = tensor<fp32, []>(0x1p-24)];
+            tensor<fp32, [1, 128, ?]> var_75 = add(x = x_11, y = var_74)[name = tensor<string, []>("op_75")];
+            tensor<fp32, []> x_epsilon_0 = const()[name = tensor<string, []>("x_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
+            tensor<fp32, [1, 128, ?]> x = log(epsilon = x_epsilon_0, x = var_75)[name = tensor<string, []>("x")];
+            tensor<int32, [3]> var_77_shape = shape(x = x)[name = tensor<string, []>("op_77_shape")];
+            tensor<int32, []> select_4 = const()[name = tensor<string, []>("select_4"), val = tensor<int32, []>(2)];
+            tensor<int32, []> gather_4_axis_0 = const()[name = tensor<string, []>("gather_4_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> gather_4_batch_dims_0 = const()[name = tensor<string, []>("gather_4_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_4_validate_indices_0 = const()[name = tensor<string, []>("gather_4_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<int32, []> gather_4 = gather(axis = gather_4_axis_0, batch_dims = gather_4_batch_dims_0, indices = select_4, validate_indices = gather_4_validate_indices_0, x = var_77_shape)[name = tensor<string, []>("gather_4")];
+            tensor<int32, []> const_3 = const()[name = tensor<string, []>("const_3"), val = tensor<int32, []>(0)];
+            tensor<int32, []> const_4 = const()[name = tensor<string, []>("const_4"), val = tensor<int32, []>(1)];
+            tensor<int32, [?]> mask_1 = range_1d(end = gather_4, start = const_3, step = const_4)[name = tensor<string, []>("mask_1")];
+            tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = tensor<string, []>("expand_dims_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = tensor<string, []>("expand_dims_0")];
+            tensor<int32, [1]> var_82_axes_0 = const()[name = tensor<string, []>("op_82_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> mel_length = cast(dtype = cast_2_dtype_0, x = seq_len_1)[name = tensor<string, []>("cast_10")];
+            tensor<int32, [1, 1]> var_82 = expand_dims(axes = var_82_axes_0, x = mel_length)[name = tensor<string, []>("op_82")];
+            tensor<bool, [1, ?]> mask = greater_equal(x = expand_dims_0, y = var_82)[name = tensor<string, []>("mask")];
+            tensor<int32, [1]> var_84_axes_0 = const()[name = tensor<string, []>("op_84_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, ?]> var_84 = expand_dims(axes = var_84_axes_0, x = mask)[name = tensor<string, []>("op_84")];
+            tensor<fp32, []> cast_7 = const()[name = tensor<string, []>("cast_7"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<fp32, [1, 128, ?]> mel = select(a = cast_7, b = x, cond = var_84)[name = tensor<string, []>("processed_signal")];
+        } -> (mel, mel_length);
+}
\ No newline at end of file
diff --git a/160ms/parakeet_eou_preprocessor.mlmodelc/weights/weight.bin b/160ms/parakeet_eou_preprocessor.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..99f6476377d3e7474aa570db365b7429f9d7b5b8
--- /dev/null
+++ b/160ms/parakeet_eou_preprocessor.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:009bba4fde82dc55db9b55d77cf3ba5f791ce366c49f079285fe25a3b6e2291d
+size 1184512
diff --git a/160ms/parakeet_eou_preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel b/160ms/parakeet_eou_preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..3e2bc094c7b810cced59b083c5713339db85e068
--- /dev/null
+++ b/160ms/parakeet_eou_preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0849137cc7b828554ddf85d6bc7923b76e8ade3a8f111e72dcedecc9889b21ba
+size 16111
diff --git a/160ms/parakeet_eou_preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/160ms/parakeet_eou_preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..117eb770f48b9e96d96462009988c22143d0c615
--- /dev/null
+++ b/160ms/parakeet_eou_preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f257ad1ac11575d73a6ffda555319b2c96b0a224f0dc03ddd8c62950e9b18e53
+size 592384
diff --git a/160ms/parakeet_eou_preprocessor.mlpackage/Manifest.json b/160ms/parakeet_eou_preprocessor.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..13fb6413b0fb5c9e552cea893a70ea84bb6ffefa
--- /dev/null
+++ b/160ms/parakeet_eou_preprocessor.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "07C051EE-C19D-4B6F-A62D-88BF3ADE779A": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "CBBDA0B6-FD16-43BF-914C-B518684DCC1B": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "07C051EE-C19D-4B6F-A62D-88BF3ADE779A"
+}
diff --git a/160ms/streaming_encoder.mlmodelc/analytics/coremldata.bin b/160ms/streaming_encoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..042b43c0828ad617d3510bdaea4d7e87e0d9d2d8
--- /dev/null
+++ b/160ms/streaming_encoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a981b257db79b4f86e6fa06a92562160a0ae71554746c24af24d8634b85f0356
+size 243
diff --git a/160ms/streaming_encoder.mlmodelc/coremldata.bin b/160ms/streaming_encoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..7c32b6d8b90a4cfb081aa7a8531b4a8dc5765a4b
--- /dev/null
+++ b/160ms/streaming_encoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e762abc60d999bcd10aab985b68191a602f2e8e03165cf08671c60f93936037a
+size 670
diff --git a/160ms/streaming_encoder.mlmodelc/metadata.json b/160ms/streaming_encoder.mlmodelc/metadata.json
new file mode 100644
index 0000000000000000000000000000000000000000..32148a71ed327b665c0b11778846e2378749683d
--- /dev/null
+++ b/160ms/streaming_encoder.mlmodelc/metadata.json
@@ -0,0 +1,187 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512 × 2)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 2]",
+        "name" : "encoded_output",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "encoded_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128 × 16)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 16]",
+        "name" : "new_pre_cache",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 17 × 1 × 70 × 512)",
+        "shortDescription" : "",
+        "shape" : "[17, 1, 70, 512]",
+        "name" : "new_cache_last_channel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 17 × 1 × 512 × 8)",
+        "shortDescription" : "",
+        "shape" : "[17, 1, 512, 8]",
+        "name" : "new_cache_last_time",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "new_cache_last_channel_len",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.floor" : 3,
+      "Ios17.logicalAnd" : 3,
+      "Ios17.reshape" : 103,
+      "Ios16.softmax" : 17,
+      "Ios17.matmul" : 51,
+      "Ios17.transpose" : 157,
+      "Split" : 17,
+      "Ios17.expandDims" : 6,
+      "Select" : 51,
+      "Ios17.add" : 126,
+      "Tile" : 1,
+      "Ios17.sliceByIndex" : 141,
+      "Ios16.sigmoid" : 17,
+      "Pad" : 20,
+      "Ios17.logicalNot" : 2,
+      "Ios17.layerNorm" : 102,
+      "Ios17.less" : 1,
+      "Ios17.sub" : 1,
+      "Ios17.conv" : 56,
+      "Ios17.clip" : 3,
+      "Ios16.relu" : 3,
+      "Ios17.linear" : 137,
+      "Ios17.concat" : 52,
+      "Ios17.greaterEqual" : 1,
+      "Ios17.cast" : 16,
+      "Ios16.silu" : 51,
+      "Stack" : 2,
+      "Ios17.mul" : 72
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.4.0"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128 × 17)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 17]",
+        "name" : "audio_signal",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "audio_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128 × 16)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 16]",
+        "name" : "pre_cache",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 17 × 1 × 70 × 512)",
+        "shortDescription" : "",
+        "shape" : "[17, 1, 70, 512]",
+        "name" : "cache_last_channel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 17 × 1 × 512 × 8)",
+        "shortDescription" : "",
+        "shape" : "[17, 1, 512, 8]",
+        "name" : "cache_last_time",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "cache_last_channel_len",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "streaming_encoder",
+    "method" : "predict"
+  }
+]
\ No newline at end of file
diff --git a/160ms/streaming_encoder.mlmodelc/model.mil b/160ms/streaming_encoder.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..52cb37c36ef8f9be82cd6744ed43a2e13a911b40
--- /dev/null
+++ b/160ms/streaming_encoder.mlmodelc/model.mil
@@ -0,0 +1,3124 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.4.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<int32, [1]> audio_length, tensor<fp32, [1, 128, 17]> audio_signal, tensor<fp32, [17, 1, 70, 512]> cache_last_channel, tensor<int32, [1]> cache_last_channel_len, tensor<fp32, [17, 1, 512, 8]> cache_last_time, tensor<fp32, [1, 128, 16]> pre_cache) {
+            tensor<int32, []> var_9 = const()[name = tensor<string, []>("op_9"), val = tensor<int32, []>(2)];
+            tensor<bool, []> full_input_interleave_0 = const()[name = tensor<string, []>("full_input_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> pre_cache_to_fp16_dtype_0 = const()[name = tensor<string, []>("pre_cache_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<string, []> audio_signal_to_fp16_dtype_0 = const()[name = tensor<string, []>("audio_signal_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 128, 17]> audio_signal_to_fp16 = cast(dtype = audio_signal_to_fp16_dtype_0, x = audio_signal)[name = tensor<string, []>("cast_195")];
+            tensor<fp16, [1, 128, 16]> pre_cache_to_fp16 = cast(dtype = pre_cache_to_fp16_dtype_0, x = pre_cache)[name = tensor<string, []>("cast_196")];
+            tensor<fp16, [1, 128, 33]> full_input_cast_fp16 = concat(axis = var_9, interleave = full_input_interleave_0, values = (pre_cache_to_fp16, audio_signal_to_fp16))[name = tensor<string, []>("full_input_cast_fp16")];
+            tensor<int32, []> var_12 = const()[name = tensor<string, []>("op_12"), val = tensor<int32, []>(16)];
+            tensor<int32, [1]> value_1 = add(x = audio_length, y = var_12)[name = tensor<string, []>("value_1")];
+            tensor<int32, [3]> var_28_begin_0 = const()[name = tensor<string, []>("op_28_begin_0"), val = tensor<int32, [3]>([0, 0, 17])];
+            tensor<int32, [3]> var_28_end_0 = const()[name = tensor<string, []>("op_28_end_0"), val = tensor<int32, [3]>([1, 128, 33])];
+            tensor<bool, [3]> var_28_end_mask_0 = const()[name = tensor<string, []>("op_28_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 128, 16]> var_28_cast_fp16 = slice_by_index(begin = var_28_begin_0, end = var_28_end_0, end_mask = var_28_end_mask_0, x = full_input_cast_fp16)[name = tensor<string, []>("op_28_cast_fp16")];
+            tensor<string, []> var_28_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_28_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, []> var_62 = const()[name = tensor<string, []>("op_62"), val = tensor<int32, []>(-1)];
+            tensor<int32, []> var_64 = const()[name = tensor<string, []>("op_64"), val = tensor<int32, []>(1)];
+            tensor<int32, [3]> x_1_perm_0 = const()[name = tensor<string, []>("x_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> cast_0_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_0_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> _inversed_108_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_108_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1]> value_1_to_fp16 = cast(dtype = cast_0_to_fp16_dtype_0, x = value_1)[name = tensor<string, []>("cast_193")];
+            tensor<fp16, [1]> _inversed_108_cast_fp16 = mul(x = value_1_to_fp16, y = _inversed_108_y_0_to_fp16)[name = tensor<string, []>("_inversed_108_cast_fp16")];
+            tensor<fp16, []> var_109_to_fp16 = const()[name = tensor<string, []>("op_109_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> lengths_1_cast_fp16 = add(x = _inversed_108_cast_fp16, y = var_109_to_fp16)[name = tensor<string, []>("lengths_1_cast_fp16")];
+            tensor<fp16, [1]> lengths_3_cast_fp16 = floor(x = lengths_1_cast_fp16)[name = tensor<string, []>("lengths_3_cast_fp16")];
+            tensor<fp16, []> _inversed_116_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_116_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1]> _inversed_116_cast_fp16 = mul(x = lengths_3_cast_fp16, y = _inversed_116_y_0_to_fp16)[name = tensor<string, []>("_inversed_116_cast_fp16")];
+            tensor<fp16, []> var_117_to_fp16 = const()[name = tensor<string, []>("op_117_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> lengths_7_cast_fp16 = add(x = _inversed_116_cast_fp16, y = var_117_to_fp16)[name = tensor<string, []>("lengths_7_cast_fp16")];
+            tensor<fp16, [1]> lengths_9_cast_fp16 = floor(x = lengths_7_cast_fp16)[name = tensor<string, []>("lengths_9_cast_fp16")];
+            tensor<fp16, []> _inversed_124_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_124_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1]> _inversed_124_cast_fp16 = mul(x = lengths_9_cast_fp16, y = _inversed_124_y_0_to_fp16)[name = tensor<string, []>("_inversed_124_cast_fp16")];
+            tensor<fp16, []> var_125_to_fp16 = const()[name = tensor<string, []>("op_125_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> lengths_13_cast_fp16 = add(x = _inversed_124_cast_fp16, y = var_125_to_fp16)[name = tensor<string, []>("lengths_13_cast_fp16")];
+            tensor<fp16, [1]> lengths_cast_fp16 = floor(x = lengths_13_cast_fp16)[name = tensor<string, []>("lengths_cast_fp16")];
+            tensor<string, []> cast_9_dtype_0 = const()[name = tensor<string, []>("cast_9_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1]> input_1_axes_0 = const()[name = tensor<string, []>("input_1_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 33, 128]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = full_input_cast_fp16)[name = tensor<string, []>("transpose_241")];
+            tensor<fp16, [1, 1, 33, 128]> input_1_cast_fp16 = expand_dims(axes = input_1_axes_0, x = x_1_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
+            tensor<int32, [8]> input_3_pad_0 = const()[name = tensor<string, []>("input_3_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            tensor<string, []> input_3_mode_0 = const()[name = tensor<string, []>("input_3_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_0_to_fp16 = const()[name = tensor<string, []>("const_0_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 1, 36, 131]> input_3_cast_fp16 = pad(constant_val = const_0_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<string, []> input_5_pad_type_0 = const()[name = tensor<string, []>("input_5_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_5_strides_0 = const()[name = tensor<string, []>("input_5_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, [4]> input_5_pad_0 = const()[name = tensor<string, []>("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_5_dilations_0 = const()[name = tensor<string, []>("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_5_groups_0 = const()[name = tensor<string, []>("input_5_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_0_weight_to_fp16"), val = tensor<fp16, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4736)))];
+            tensor<fp16, [1, 256, 17, 65]> input_5_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [1, 256, 17, 65]> input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<int32, [8]> input_9_pad_0 = const()[name = tensor<string, []>("input_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            tensor<string, []> input_9_mode_0 = const()[name = tensor<string, []>("input_9_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_1_to_fp16 = const()[name = tensor<string, []>("const_1_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 256, 20, 68]> input_9_cast_fp16 = pad(constant_val = const_1_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
+            tensor<string, []> input_11_pad_type_0 = const()[name = tensor<string, []>("input_11_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_11_strides_0 = const()[name = tensor<string, []>("input_11_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, []> input_11_groups_0 = const()[name = tensor<string, []>("input_11_groups_0"), val = tensor<int32, []>(256)];
+            tensor<int32, [4]> input_11_pad_0 = const()[name = tensor<string, []>("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_11_dilations_0 = const()[name = tensor<string, []>("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_2_weight_to_fp16"), val = tensor<fp16, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5312)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9984)))];
+            tensor<fp16, [1, 256, 9, 33]> input_11_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<string, []> input_13_pad_type_0 = const()[name = tensor<string, []>("input_13_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_13_strides_0 = const()[name = tensor<string, []>("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_13_pad_0 = const()[name = tensor<string, []>("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_13_dilations_0 = const()[name = tensor<string, []>("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_13_groups_0 = const()[name = tensor<string, []>("input_13_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_3_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_3_weight_to_fp16"), val = tensor<fp16, [256, 256, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10560)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141696)))];
+            tensor<fp16, [1, 256, 9, 33]> input_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
+            tensor<fp16, [1, 256, 9, 33]> input_15_cast_fp16 = relu(x = input_13_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
+            tensor<int32, [8]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            tensor<string, []> input_17_mode_0 = const()[name = tensor<string, []>("input_17_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_2_to_fp16 = const()[name = tensor<string, []>("const_2_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 256, 12, 36]> input_17_cast_fp16 = pad(constant_val = const_2_to_fp16, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
+            tensor<string, []> input_19_pad_type_0 = const()[name = tensor<string, []>("input_19_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_19_strides_0 = const()[name = tensor<string, []>("input_19_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, []> input_19_groups_0 = const()[name = tensor<string, []>("input_19_groups_0"), val = tensor<int32, []>(256)];
+            tensor<int32, [4]> input_19_pad_0 = const()[name = tensor<string, []>("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_19_dilations_0 = const()[name = tensor<string, []>("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_5_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_5_weight_to_fp16"), val = tensor<fp16, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142272)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146944)))];
+            tensor<fp16, [1, 256, 5, 17]> input_19_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
+            tensor<string, []> input_21_pad_type_0 = const()[name = tensor<string, []>("input_21_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_21_strides_0 = const()[name = tensor<string, []>("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_21_pad_0 = const()[name = tensor<string, []>("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_21_dilations_0 = const()[name = tensor<string, []>("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_21_groups_0 = const()[name = tensor<string, []>("input_21_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_6_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_6_weight_to_fp16"), val = tensor<fp16, [256, 256, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147520)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278656)))];
+            tensor<fp16, [1, 256, 5, 17]> input_21_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
+            tensor<fp16, [1, 256, 5, 17]> x_3_cast_fp16 = relu(x = input_21_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
+            tensor<int32, [4]> var_181_perm_0 = const()[name = tensor<string, []>("op_181_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_182 = const()[name = tensor<string, []>("op_182"), val = tensor<int32, [3]>([1, 5, -1])];
+            tensor<fp16, [1, 5, 256, 17]> var_181_cast_fp16 = transpose(perm = var_181_perm_0, x = x_3_cast_fp16)[name = tensor<string, []>("transpose_240")];
+            tensor<fp16, [1, 5, 4352]> input_23_cast_fp16 = reshape(shape = var_182, x = var_181_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
+            tensor<fp16, [512, 4352]> encoder_pre_encode_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_out_weight_to_fp16"), val = tensor<fp16, [512, 4352]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279232)))];
+            tensor<fp16, [512]> encoder_pre_encode_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4735744)))];
+            tensor<fp16, [1, 5, 512]> linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<int32, [3]> var_192_begin_0 = const()[name = tensor<string, []>("op_192_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_192_end_0 = const()[name = tensor<string, []>("op_192_end_0"), val = tensor<int32, [3]>([1, 5, 512])];
+            tensor<bool, [3]> var_192_end_mask_0 = const()[name = tensor<string, []>("op_192_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 3, 512]> var_192_cast_fp16 = slice_by_index(begin = var_192_begin_0, end = var_192_end_0, end_mask = var_192_end_mask_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_192_cast_fp16")];
+            tensor<int32, []> var_194 = const()[name = tensor<string, []>("op_194"), val = tensor<int32, []>(2)];
+            tensor<int32, [1]> lengths_cast_fp16_to_int32 = cast(dtype = cast_9_dtype_0, x = lengths_cast_fp16)[name = tensor<string, []>("cast_192")];
+            tensor<int32, [1]> var_195 = sub(x = lengths_cast_fp16_to_int32, y = var_194)[name = tensor<string, []>("op_195")];
+            tensor<string, []> var_195_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_195_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> var_60_promoted_to_fp16 = const()[name = tensor<string, []>("op_60_promoted_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, []> const_5_to_fp16 = const()[name = tensor<string, []>("const_5_to_fp16"), val = tensor<fp16, []>(inf)];
+            tensor<fp16, [1]> var_195_to_fp16 = cast(dtype = var_195_promoted_to_fp16_dtype_0, x = var_195)[name = tensor<string, []>("cast_191")];
+            tensor<fp16, [1]> clip_0_cast_fp16 = clip(alpha = var_60_promoted_to_fp16, beta = const_5_to_fp16, x = var_195_to_fp16)[name = tensor<string, []>("clip_0_cast_fp16")];
+            tensor<int32, [1]> cache_keep_size = const()[name = tensor<string, []>("cache_keep_size"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, []> var_211_promoted_to_fp16 = const()[name = tensor<string, []>("op_211_promoted_to_fp16"), val = tensor<fp16, []>(0x1.18p+6)];
+            tensor<fp16, [1]> padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_211_promoted_to_fp16)[name = tensor<string, []>("padding_length_cast_fp16")];
+            tensor<int32, []> const_7 = const()[name = tensor<string, []>("const_7"), val = tensor<int32, []>(-1)];
+            tensor<int32, [1]> var_213 = mul(x = cache_last_channel_len, y = const_7)[name = tensor<string, []>("op_213")];
+            tensor<int32, []> var_214 = const()[name = tensor<string, []>("op_214"), val = tensor<int32, []>(70)];
+            tensor<int32, [1]> offset = add(x = var_213, y = var_214)[name = tensor<string, []>("offset")];
+            tensor<int32, [1]> var_254_axes_0 = const()[name = tensor<string, []>("op_254_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 1]> var_254_cast_fp16 = expand_dims(axes = var_254_axes_0, x = padding_length_cast_fp16)[name = tensor<string, []>("op_254_cast_fp16")];
+            tensor<fp16, [1, 73]> var_253_promoted_to_fp16 = const()[name = tensor<string, []>("op_253_promoted_to_fp16"), val = tensor<fp16, [1, 73]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4736832)))];
+            tensor<bool, [1, 73]> pad_mask_1_cast_fp16 = less(x = var_253_promoted_to_fp16, y = var_254_cast_fp16)[name = tensor<string, []>("pad_mask_1_cast_fp16")];
+            tensor<int32, [1, 73]> expand_dims_1 = const()[name = tensor<string, []>("expand_dims_1"), val = tensor<int32, [1, 73]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72]])];
+            tensor<int32, [1]> var_260_axes_0 = const()[name = tensor<string, []>("op_260_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 1]> var_260 = expand_dims(axes = var_260_axes_0, x = offset)[name = tensor<string, []>("op_260")];
+            tensor<bool, [1, 73]> pad_mask_off = greater_equal(x = expand_dims_1, y = var_260)[name = tensor<string, []>("pad_mask_off")];
+            tensor<bool, [1, 73]> pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = tensor<string, []>("pad_mask_3")];
+            tensor<int32, [1]> var_263_axes_0 = const()[name = tensor<string, []>("op_263_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 73]> var_263 = expand_dims(axes = var_263_axes_0, x = pad_mask_3)[name = tensor<string, []>("op_263")];
+            tensor<int32, [3]> var_264 = const()[name = tensor<string, []>("op_264"), val = tensor<int32, [3]>([1, 73, 1])];
+            tensor<bool, [1, 73, 73]> pad_mask_for_att_mask_1 = tile(reps = var_264, x = var_263)[name = tensor<string, []>("pad_mask_for_att_mask_1")];
+            tensor<int32, [3]> var_266_perm_0 = const()[name = tensor<string, []>("op_266_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<bool, [1, 73, 73]> var_266 = transpose(perm = var_266_perm_0, x = pad_mask_for_att_mask_1)[name = tensor<string, []>("transpose_239")];
+            tensor<bool, [1, 73, 73]> pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_266)[name = tensor<string, []>("pad_mask_for_att_mask")];
+            tensor<bool, [1, 73, 73]> const_15 = const()[name = tensor<string, []>("const_15"), val = tensor<bool, [1, 73, 73]>([[[true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false], [false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])];
+            tensor<bool, [1, 73, 73]> att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_15)[name = tensor<string, []>("att_mask_9")];
+            tensor<bool, [1, 73, 73]> att_mask = logical_not(x = att_mask_9)[name = tensor<string, []>("att_mask")];
+            tensor<bool, [1, 73]> pad_mask_5 = logical_not(x = pad_mask_3)[name = tensor<string, []>("pad_mask_5")];
+            tensor<int32, [2]> pad_mask_begin_0 = const()[name = tensor<string, []>("pad_mask_begin_0"), val = tensor<int32, [2]>([0, 70])];
+            tensor<int32, [2]> pad_mask_end_0 = const()[name = tensor<string, []>("pad_mask_end_0"), val = tensor<int32, [2]>([1, 73])];
+            tensor<bool, [2]> pad_mask_end_mask_0 = const()[name = tensor<string, []>("pad_mask_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<bool, [1, 3]> pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = tensor<string, []>("pad_mask")];
+            tensor<int32, [3]> mask_1_begin_0 = const()[name = tensor<string, []>("mask_1_begin_0"), val = tensor<int32, [3]>([0, 70, 0])];
+            tensor<int32, [3]> mask_1_end_0 = const()[name = tensor<string, []>("mask_1_end_0"), val = tensor<int32, [3]>([1, 73, 73])];
+            tensor<bool, [3]> mask_1_end_mask_0 = const()[name = tensor<string, []>("mask_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<bool, [1, 3, 73]> mask_1 = slice_by_index(begin = mask_1_begin_0, end = mask_1_end_0, end_mask = mask_1_end_mask_0, x = att_mask)[name = tensor<string, []>("mask_1")];
+            tensor<int32, [4]> cache_1_begin_0 = const()[name = tensor<string, []>("cache_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_1_end_0 = const()[name = tensor<string, []>("cache_1_end_0"), val = tensor<int32, [4]>([1, 1, 70, 512])];
+            tensor<bool, [4]> cache_1_end_mask_0 = const()[name = tensor<string, []>("cache_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_1_squeeze_mask_0 = const()[name = tensor<string, []>("cache_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<string, []> cache_last_channel_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_last_channel_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [17, 1, 70, 512]> cache_last_channel_to_fp16 = cast(dtype = cache_last_channel_to_fp16_dtype_0, x = cache_last_channel)[name = tensor<string, []>("cast_190")];
+            tensor<fp16, [1, 70, 512]> cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_1_cast_fp16")];
+            tensor<int32, [4]> cache_3_begin_0 = const()[name = tensor<string, []>("cache_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_3_end_0 = const()[name = tensor<string, []>("cache_3_end_0"), val = tensor<int32, [4]>([1, 1, 512, 8])];
+            tensor<bool, [4]> cache_3_end_mask_0 = const()[name = tensor<string, []>("cache_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_3_squeeze_mask_0 = const()[name = tensor<string, []>("cache_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<string, []> cache_last_time_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_last_time_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [17, 1, 512, 8]> cache_last_time_to_fp16 = cast(dtype = cache_last_time_to_fp16_dtype_0, x = cache_last_time)[name = tensor<string, []>("cast_189")];
+            tensor<fp16, [1, 512, 8]> cache_3_cast_fp16 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_3_cast_fp16")];
+            tensor<int32, [1]> input_27_axes_0 = const()[name = tensor<string, []>("input_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4737088)))];
+            tensor<fp16, [512]> encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4738176)))];
+            tensor<fp16, []> var_38_to_fp16 = const()[name = tensor<string, []>("op_38_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 3, 512]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_192_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_0_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4739264)))];
+            tensor<fp16, [2048]> linear_1_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_1_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6836480)))];
+            tensor<fp16, [1, 3, 2048]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_0_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6840640)))];
+            tensor<fp16, [512]> linear_2_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_2_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8937856)))];
+            tensor<fp16, [1, 3, 512]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<fp16, []> var_303_to_fp16 = const()[name = tensor<string, []>("op_303_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_304_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_303_to_fp16)[name = tensor<string, []>("op_304_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_37_cast_fp16 = add(x = var_192_cast_fp16, y = var_304_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
+            tensor<int32, [1]> key_1_axes_0 = const()[name = tensor<string, []>("key_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8938944)))];
+            tensor<fp16, [512]> encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8940032)))];
+            tensor<fp16, [1, 3, 512]> key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")];
+            tensor<bool, []> input_39_interleave_0 = const()[name = tensor<string, []>("input_39_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 73, 512]> input_39_cast_fp16 = concat(axis = var_64, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = tensor<string, []>("input_39_cast_fp16")];
+            tensor<int32, [3]> var_326_begin_0 = const()[name = tensor<string, []>("op_326_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_326_end_0 = const()[name = tensor<string, []>("op_326_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_326_end_mask_0 = const()[name = tensor<string, []>("op_326_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 69, 512]> var_326_cast_fp16 = slice_by_index(begin = var_326_begin_0, end = var_326_end_0, end_mask = var_326_end_mask_0, x = cache_1_cast_fp16)[name = tensor<string, []>("op_326_cast_fp16")];
+            tensor<int32, [3]> var_329_begin_0 = const()[name = tensor<string, []>("op_329_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_329_end_0 = const()[name = tensor<string, []>("op_329_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<bool, [3]> var_329_end_mask_0 = const()[name = tensor<string, []>("op_329_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 1, 512]> var_329_cast_fp16 = slice_by_index(begin = var_329_begin_0, end = var_329_end_0, end_mask = var_329_end_mask_0, x = key_1_cast_fp16)[name = tensor<string, []>("op_329_cast_fp16")];
+            tensor<bool, []> var_332_interleave_0 = const()[name = tensor<string, []>("op_332_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_332_cast_fp16 = concat(axis = var_64, interleave = var_332_interleave_0, values = (var_326_cast_fp16, var_329_cast_fp16))[name = tensor<string, []>("op_332_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_0_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8941120)))];
+            tensor<fp16, [1, 3, 512]> linear_3_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16, x = key_1_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
+            tensor<int32, [4]> var_336 = const()[name = tensor<string, []>("op_336"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 3, 8, 64]> q_1_cast_fp16 = reshape(shape = var_336, x = linear_3_cast_fp16)[name = tensor<string, []>("q_1_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_0_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9465472)))];
+            tensor<fp16, [1, 73, 512]> linear_4_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
+            tensor<int32, [4]> var_340 = const()[name = tensor<string, []>("op_340"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> k_1_cast_fp16 = reshape(shape = var_340, x = linear_4_cast_fp16)[name = tensor<string, []>("k_1_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_0_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9989824)))];
+            tensor<fp16, [1, 73, 512]> linear_5_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
+            tensor<int32, [4]> var_344 = const()[name = tensor<string, []>("op_344"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> v_1_cast_fp16 = reshape(shape = var_344, x = linear_5_cast_fp16)[name = tensor<string, []>("v_1_cast_fp16")];
+            tensor<int32, [4]> value_3_perm_0 = const()[name = tensor<string, []>("value_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10514176)))];
+            tensor<fp16, [1, 3, 8, 64]> var_356_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_356_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10515264)))];
+            tensor<fp16, [1, 3, 8, 64]> var_358_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_358_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_1_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_7_transpose_x_0 = const()[name = tensor<string, []>("x_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_7_transpose_y_0 = const()[name = tensor<string, []>("x_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 145]> var_360_to_fp16 = const()[name = tensor<string, []>("op_360_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10516352)))];
+            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_358_cast_fp16)[name = tensor<string, []>("transpose_237")];
+            tensor<fp16, [1, 8, 3, 145]> x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = var_360_to_fp16)[name = tensor<string, []>("x_7_cast_fp16")];
+            tensor<int32, [8]> x_9_pad_0 = const()[name = tensor<string, []>("x_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_9_mode_0 = const()[name = tensor<string, []>("x_9_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_23_to_fp16 = const()[name = tensor<string, []>("const_23_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 3, 146]> x_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7_cast_fp16)[name = tensor<string, []>("x_9_cast_fp16")];
+            tensor<int32, [4]> var_368 = const()[name = tensor<string, []>("op_368"), val = tensor<int32, [4]>([1, 8, -1, 3])];
+            tensor<fp16, [1, 8, 146, 3]> x_11_cast_fp16 = reshape(shape = var_368, x = x_9_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
+            tensor<int32, [4]> var_372_begin_0 = const()[name = tensor<string, []>("op_372_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_372_end_0 = const()[name = tensor<string, []>("op_372_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
+            tensor<bool, [4]> var_372_end_mask_0 = const()[name = tensor<string, []>("op_372_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 145, 3]> var_372_cast_fp16 = slice_by_index(begin = var_372_begin_0, end = var_372_end_0, end_mask = var_372_end_mask_0, x = x_11_cast_fp16)[name = tensor<string, []>("op_372_cast_fp16")];
+            tensor<int32, [4]> var_373 = const()[name = tensor<string, []>("op_373"), val = tensor<int32, [4]>([1, 8, 3, 145])];
+            tensor<fp16, [1, 8, 3, 145]> matrix_bd_1_cast_fp16 = reshape(shape = var_373, x = var_372_cast_fp16)[name = tensor<string, []>("matrix_bd_1_cast_fp16")];
+            tensor<bool, []> matrix_ac_1_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_1_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_51_perm_0 = const()[name = tensor<string, []>("transpose_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_52_perm_0 = const()[name = tensor<string, []>("transpose_52_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 73]> transpose_52 = transpose(perm = transpose_52_perm_0, x = k_1_cast_fp16)[name = tensor<string, []>("transpose_235")];
+            tensor<fp16, [1, 8, 3, 64]> transpose_51 = transpose(perm = transpose_51_perm_0, x = var_356_cast_fp16)[name = tensor<string, []>("transpose_236")];
+            tensor<fp16, [1, 8, 3, 73]> matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_51, y = transpose_52)[name = tensor<string, []>("matrix_ac_1_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_3_begin_0 = const()[name = tensor<string, []>("matrix_bd_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_3_end_0 = const()[name = tensor<string, []>("matrix_bd_3_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
+            tensor<bool, [4]> matrix_bd_3_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 3, 73]> matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1_cast_fp16)[name = tensor<string, []>("matrix_bd_3_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_382_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = tensor<string, []>("op_382_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_1_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_1_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_1_cast_fp16 = mul(x = var_382_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_1_cast_fp16")];
+            tensor<int32, [1]> mask_3_axes_0 = const()[name = tensor<string, []>("mask_3_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 3, 73]> mask_3 = expand_dims(axes = mask_3_axes_0, x = mask_1)[name = tensor<string, []>("mask_3")];
+            tensor<fp16, []> var_41_to_fp16 = const()[name = tensor<string, []>("op_41_to_fp16"), val = tensor<fp16, []>(-0x1.388p+13)];
+            tensor<fp16, [1, 8, 3, 73]> scores_3_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_3_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_388_cast_fp16 = softmax(axis = var_62, x = scores_3_cast_fp16)[name = tensor<string, []>("op_388_cast_fp16")];
+            tensor<fp16, []> var_40_to_fp16 = const()[name = tensor<string, []>("op_40_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 3, 73]> input_41_cast_fp16 = select(a = var_40_to_fp16, b = var_388_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_41_cast_fp16")];
+            tensor<bool, []> x_13_transpose_x_0 = const()[name = tensor<string, []>("x_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_13_transpose_y_0 = const()[name = tensor<string, []>("x_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 73, 64]> value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = v_1_cast_fp16)[name = tensor<string, []>("transpose_238")];
+            tensor<fp16, [1, 8, 3, 64]> x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41_cast_fp16, y = value_3_cast_fp16)[name = tensor<string, []>("x_13_cast_fp16")];
+            tensor<int32, [4]> var_392_perm_0 = const()[name = tensor<string, []>("op_392_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_393 = const()[name = tensor<string, []>("op_393"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 3, 8, 64]> var_392_cast_fp16 = transpose(perm = var_392_perm_0, x = x_13_cast_fp16)[name = tensor<string, []>("transpose_234")];
+            tensor<fp16, [1, 3, 512]> input_43_cast_fp16 = reshape(shape = var_393, x = var_392_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_0_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10664896)))];
+            tensor<fp16, [1, 3, 512]> linear_7_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
+            tensor<int32, [1]> x_17_axes_0 = const()[name = tensor<string, []>("x_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11189248)))];
+            tensor<fp16, [512]> encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11190336)))];
+            tensor<fp16, [1, 3, 512]> x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("x_17_cast_fp16")];
+            tensor<int32, [3]> input_49_perm_0 = const()[name = tensor<string, []>("input_49_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_51_pad_type_0 = const()[name = tensor<string, []>("input_51_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_51_strides_0 = const()[name = tensor<string, []>("input_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_51_pad_0 = const()[name = tensor<string, []>("input_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_51_dilations_0 = const()[name = tensor<string, []>("input_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_51_groups_0 = const()[name = tensor<string, []>("input_51_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_0_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11191424)))];
+            tensor<fp16, [1, 512, 3]> input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = tensor<string, []>("transpose_233")];
+            tensor<fp16, [1, 1024, 3]> input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
+            tensor<int32, []> x_19_split_num_splits_0 = const()[name = tensor<string, []>("x_19_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_19_split_axis_0 = const()[name = tensor<string, []>("x_19_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 3]> x_19_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51_cast_fp16)[name = tensor<string, []>("x_19_split_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = tensor<string, []>("x_19_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")];
+            tensor<int32, [1]> var_418_axes_0 = const()[name = tensor<string, []>("op_418_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 3]> var_418 = expand_dims(axes = var_418_axes_0, x = pad_mask)[name = tensor<string, []>("op_418")];
+            tensor<fp16, [1, 512, 3]> input_53_cast_fp16 = select(a = var_40_to_fp16, b = x_19_cast_fp16, cond = var_418)[name = tensor<string, []>("input_53_cast_fp16")];
+            tensor<bool, []> new_x_3_interleave_0 = const()[name = tensor<string, []>("new_x_3_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 11]> new_x_3_cast_fp16 = concat(axis = var_62, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = tensor<string, []>("new_x_3_cast_fp16")];
+            tensor<int32, [3]> next_cache_1_begin_0 = const()[name = tensor<string, []>("next_cache_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_1_end_0 = const()[name = tensor<string, []>("next_cache_1_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> next_cache_1_end_mask_0 = const()[name = tensor<string, []>("next_cache_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 9]> next_cache_1_cast_fp16 = slice_by_index(begin = next_cache_1_begin_0, end = next_cache_1_end_0, end_mask = next_cache_1_end_mask_0, x = new_x_3_cast_fp16)[name = tensor<string, []>("next_cache_1_cast_fp16")];
+            tensor<int32, [3]> var_434_begin_0 = const()[name = tensor<string, []>("op_434_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> var_434_end_0 = const()[name = tensor<string, []>("op_434_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> var_434_end_mask_0 = const()[name = tensor<string, []>("op_434_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_434_cast_fp16 = slice_by_index(begin = var_434_begin_0, end = var_434_end_0, end_mask = var_434_end_mask_0, x = next_cache_1_cast_fp16)[name = tensor<string, []>("op_434_cast_fp16")];
+            tensor<string, []> x_21_pad_type_0 = const()[name = tensor<string, []>("x_21_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_21_groups_0 = const()[name = tensor<string, []>("x_21_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_21_strides_0 = const()[name = tensor<string, []>("x_21_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_21_pad_0 = const()[name = tensor<string, []>("x_21_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_21_dilations_0 = const()[name = tensor<string, []>("x_21_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_0_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12240064)))];
+            tensor<fp16, [1, 512, 3]> x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16, x = new_x_3_cast_fp16)[name = tensor<string, []>("x_21_cast_fp16")];
+            tensor<int32, [3]> input_55_perm_0 = const()[name = tensor<string, []>("input_55_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_23_axes_0 = const()[name = tensor<string, []>("x_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12249344)))];
+            tensor<fp16, [512]> encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12250432)))];
+            tensor<fp16, [1, 3, 512]> input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = tensor<string, []>("transpose_232")];
+            tensor<fp16, [1, 3, 512]> x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")];
+            tensor<int32, [3]> input_57_perm_0 = const()[name = tensor<string, []>("input_57_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 3]> input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = tensor<string, []>("transpose_231")];
+            tensor<fp16, [1, 512, 3]> input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
+            tensor<string, []> x_25_pad_type_0 = const()[name = tensor<string, []>("x_25_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_25_strides_0 = const()[name = tensor<string, []>("x_25_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_25_pad_0 = const()[name = tensor<string, []>("x_25_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_25_dilations_0 = const()[name = tensor<string, []>("x_25_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_25_groups_0 = const()[name = tensor<string, []>("x_25_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_0_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12251520)))];
+            tensor<fp16, [1, 512, 3]> x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("x_25_cast_fp16")];
+            tensor<int32, [3]> input_61_perm_0 = const()[name = tensor<string, []>("input_61_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 3, 512]> input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = tensor<string, []>("transpose_230")];
+            tensor<fp16, [1, 3, 512]> input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
+            tensor<int32, [1]> input_65_axes_0 = const()[name = tensor<string, []>("input_65_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12775872)))];
+            tensor<fp16, [512]> encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12776960)))];
+            tensor<fp16, [1, 3, 512]> input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_0_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12778048)))];
+            tensor<fp16, [1, 3, 2048]> linear_8_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_0_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14875264)))];
+            tensor<fp16, [1, 3, 512]> linear_9_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
+            tensor<fp16, []> var_475_to_fp16 = const()[name = tensor<string, []>("op_475_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_476_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_475_to_fp16)[name = tensor<string, []>("op_476_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_476_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
+            tensor<int32, [1]> input_77_axes_0 = const()[name = tensor<string, []>("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_0_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16972480)))];
+            tensor<fp16, [512]> encoder_layers_0_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16973568)))];
+            tensor<fp16, [1, 3, 512]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
+            tensor<int32, [4]> cache_5_begin_0 = const()[name = tensor<string, []>("cache_5_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_5_end_0 = const()[name = tensor<string, []>("cache_5_end_0"), val = tensor<int32, [4]>([2, 1, 70, 512])];
+            tensor<bool, [4]> cache_5_end_mask_0 = const()[name = tensor<string, []>("cache_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_5_squeeze_mask_0 = const()[name = tensor<string, []>("cache_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_5_cast_fp16 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_5_cast_fp16")];
+            tensor<int32, [4]> cache_7_begin_0 = const()[name = tensor<string, []>("cache_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_7_end_0 = const()[name = tensor<string, []>("cache_7_end_0"), val = tensor<int32, [4]>([2, 1, 512, 8])];
+            tensor<bool, [4]> cache_7_end_mask_0 = const()[name = tensor<string, []>("cache_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_7_squeeze_mask_0 = const()[name = tensor<string, []>("cache_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_7_cast_fp16 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_7_cast_fp16")];
+            tensor<int32, [1]> input_79_axes_0 = const()[name = tensor<string, []>("input_79_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16974656)))];
+            tensor<fp16, [512]> encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16975744)))];
+            tensor<fp16, [1, 3, 512]> input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_1_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16976832)))];
+            tensor<fp16, [1, 3, 2048]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_1_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19074048)))];
+            tensor<fp16, [1, 3, 512]> linear_11_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
+            tensor<fp16, []> var_510_to_fp16 = const()[name = tensor<string, []>("op_510_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_511_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_510_to_fp16)[name = tensor<string, []>("op_511_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_511_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
+            tensor<int32, [1]> key_3_axes_0 = const()[name = tensor<string, []>("key_3_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21171264)))];
+            tensor<fp16, [512]> encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21172352)))];
+            tensor<fp16, [1, 3, 512]> key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("key_3_cast_fp16")];
+            tensor<bool, []> input_91_interleave_0 = const()[name = tensor<string, []>("input_91_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 73, 512]> input_91_cast_fp16 = concat(axis = var_64, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = tensor<string, []>("input_91_cast_fp16")];
+            tensor<int32, [3]> var_533_begin_0 = const()[name = tensor<string, []>("op_533_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_533_end_0 = const()[name = tensor<string, []>("op_533_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_533_end_mask_0 = const()[name = tensor<string, []>("op_533_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 69, 512]> var_533_cast_fp16 = slice_by_index(begin = var_533_begin_0, end = var_533_end_0, end_mask = var_533_end_mask_0, x = cache_5_cast_fp16)[name = tensor<string, []>("op_533_cast_fp16")];
+            tensor<int32, [3]> var_536_begin_0 = const()[name = tensor<string, []>("op_536_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_536_end_0 = const()[name = tensor<string, []>("op_536_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<bool, [3]> var_536_end_mask_0 = const()[name = tensor<string, []>("op_536_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 1, 512]> var_536_cast_fp16 = slice_by_index(begin = var_536_begin_0, end = var_536_end_0, end_mask = var_536_end_mask_0, x = key_3_cast_fp16)[name = tensor<string, []>("op_536_cast_fp16")];
+            tensor<bool, []> var_539_interleave_0 = const()[name = tensor<string, []>("op_539_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_539_cast_fp16 = concat(axis = var_64, interleave = var_539_interleave_0, values = (var_533_cast_fp16, var_536_cast_fp16))[name = tensor<string, []>("op_539_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_1_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21173440)))];
+            tensor<fp16, [1, 3, 512]> linear_12_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16, x = key_3_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
+            tensor<int32, [4]> var_543 = const()[name = tensor<string, []>("op_543"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 3, 8, 64]> q_7_cast_fp16 = reshape(shape = var_543, x = linear_12_cast_fp16)[name = tensor<string, []>("q_7_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_1_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21697792)))];
+            tensor<fp16, [1, 73, 512]> linear_13_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
+            tensor<int32, [4]> var_547 = const()[name = tensor<string, []>("op_547"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> k_5_cast_fp16 = reshape(shape = var_547, x = linear_13_cast_fp16)[name = tensor<string, []>("k_5_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_1_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22222144)))];
+            tensor<fp16, [1, 73, 512]> linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
+            tensor<int32, [4]> var_551 = const()[name = tensor<string, []>("op_551"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> v_3_cast_fp16 = reshape(shape = var_551, x = linear_14_cast_fp16)[name = tensor<string, []>("v_3_cast_fp16")];
+            tensor<int32, [4]> value_5_perm_0 = const()[name = tensor<string, []>("value_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22746496)))];
+            tensor<fp16, [1, 3, 8, 64]> var_563_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_563_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22747584)))];
+            tensor<fp16, [1, 3, 8, 64]> var_565_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_565_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_3_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_33_transpose_x_0 = const()[name = tensor<string, []>("x_33_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_33_transpose_y_0 = const()[name = tensor<string, []>("x_33_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 145]> var_567_to_fp16 = const()[name = tensor<string, []>("op_567_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22748672)))];
+            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_565_cast_fp16)[name = tensor<string, []>("transpose_228")];
+            tensor<fp16, [1, 8, 3, 145]> x_33_cast_fp16 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = var_567_to_fp16)[name = tensor<string, []>("x_33_cast_fp16")];
+            tensor<int32, [8]> x_35_pad_0 = const()[name = tensor<string, []>("x_35_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_35_mode_0 = const()[name = tensor<string, []>("x_35_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_36_to_fp16 = const()[name = tensor<string, []>("const_36_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 3, 146]> x_35_cast_fp16 = pad(constant_val = const_36_to_fp16, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")];
+            tensor<int32, [4]> var_575 = const()[name = tensor<string, []>("op_575"), val = tensor<int32, [4]>([1, 8, -1, 3])];
+            tensor<fp16, [1, 8, 146, 3]> x_37_cast_fp16 = reshape(shape = var_575, x = x_35_cast_fp16)[name = tensor<string, []>("x_37_cast_fp16")];
+            tensor<int32, [4]> var_579_begin_0 = const()[name = tensor<string, []>("op_579_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_579_end_0 = const()[name = tensor<string, []>("op_579_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
+            tensor<bool, [4]> var_579_end_mask_0 = const()[name = tensor<string, []>("op_579_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 145, 3]> var_579_cast_fp16 = slice_by_index(begin = var_579_begin_0, end = var_579_end_0, end_mask = var_579_end_mask_0, x = x_37_cast_fp16)[name = tensor<string, []>("op_579_cast_fp16")];
+            tensor<int32, [4]> var_580 = const()[name = tensor<string, []>("op_580"), val = tensor<int32, [4]>([1, 8, 3, 145])];
+            tensor<fp16, [1, 8, 3, 145]> matrix_bd_5_cast_fp16 = reshape(shape = var_580, x = var_579_cast_fp16)[name = tensor<string, []>("matrix_bd_5_cast_fp16")];
+            tensor<bool, []> matrix_ac_3_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_3_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_3_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_3_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_53_perm_0 = const()[name = tensor<string, []>("transpose_53_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_54_perm_0 = const()[name = tensor<string, []>("transpose_54_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 73]> transpose_54 = transpose(perm = transpose_54_perm_0, x = k_5_cast_fp16)[name = tensor<string, []>("transpose_226")];
+            tensor<fp16, [1, 8, 3, 64]> transpose_53 = transpose(perm = transpose_53_perm_0, x = var_563_cast_fp16)[name = tensor<string, []>("transpose_227")];
+            tensor<fp16, [1, 8, 3, 73]> matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_53, y = transpose_54)[name = tensor<string, []>("matrix_ac_3_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_7_begin_0 = const()[name = tensor<string, []>("matrix_bd_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_7_end_0 = const()[name = tensor<string, []>("matrix_bd_7_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
+            tensor<bool, [4]> matrix_bd_7_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 3, 73]> matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5_cast_fp16)[name = tensor<string, []>("matrix_bd_7_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_589_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = tensor<string, []>("op_589_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_5_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_5_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_5_cast_fp16 = mul(x = var_589_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_5_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> scores_7_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_7_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_595_cast_fp16 = softmax(axis = var_62, x = scores_7_cast_fp16)[name = tensor<string, []>("op_595_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> input_93_cast_fp16 = select(a = var_40_to_fp16, b = var_595_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_93_cast_fp16")];
+            tensor<bool, []> x_39_transpose_x_0 = const()[name = tensor<string, []>("x_39_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_39_transpose_y_0 = const()[name = tensor<string, []>("x_39_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 73, 64]> value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = v_3_cast_fp16)[name = tensor<string, []>("transpose_229")];
+            tensor<fp16, [1, 8, 3, 64]> x_39_cast_fp16 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93_cast_fp16, y = value_5_cast_fp16)[name = tensor<string, []>("x_39_cast_fp16")];
+            tensor<int32, [4]> var_599_perm_0 = const()[name = tensor<string, []>("op_599_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_600 = const()[name = tensor<string, []>("op_600"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 3, 8, 64]> var_599_cast_fp16 = transpose(perm = var_599_perm_0, x = x_39_cast_fp16)[name = tensor<string, []>("transpose_225")];
+            tensor<fp16, [1, 3, 512]> input_95_cast_fp16 = reshape(shape = var_600, x = var_599_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_1_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22897216)))];
+            tensor<fp16, [1, 3, 512]> linear_16_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
+            tensor<int32, [1]> x_43_axes_0 = const()[name = tensor<string, []>("x_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23421568)))];
+            tensor<fp16, [512]> encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23422656)))];
+            tensor<fp16, [1, 3, 512]> x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")];
+            tensor<int32, [3]> input_101_perm_0 = const()[name = tensor<string, []>("input_101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_103_pad_type_0 = const()[name = tensor<string, []>("input_103_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_103_strides_0 = const()[name = tensor<string, []>("input_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_103_pad_0 = const()[name = tensor<string, []>("input_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_103_dilations_0 = const()[name = tensor<string, []>("input_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_103_groups_0 = const()[name = tensor<string, []>("input_103_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_1_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23423744)))];
+            tensor<fp16, [1, 512, 3]> input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = tensor<string, []>("transpose_224")];
+            tensor<fp16, [1, 1024, 3]> input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
+            tensor<int32, []> x_45_split_num_splits_0 = const()[name = tensor<string, []>("x_45_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_45_split_axis_0 = const()[name = tensor<string, []>("x_45_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 3]> x_45_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_45_split_cast_fp16_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103_cast_fp16)[name = tensor<string, []>("x_45_split_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = tensor<string, []>("x_45_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_45_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> input_105_cast_fp16 = select(a = var_40_to_fp16, b = x_45_cast_fp16, cond = var_418)[name = tensor<string, []>("input_105_cast_fp16")];
+            tensor<bool, []> new_x_7_interleave_0 = const()[name = tensor<string, []>("new_x_7_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 11]> new_x_7_cast_fp16 = concat(axis = var_62, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = tensor<string, []>("new_x_7_cast_fp16")];
+            tensor<int32, [3]> next_cache_3_begin_0 = const()[name = tensor<string, []>("next_cache_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_3_end_0 = const()[name = tensor<string, []>("next_cache_3_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> next_cache_3_end_mask_0 = const()[name = tensor<string, []>("next_cache_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 9]> next_cache_3_cast_fp16 = slice_by_index(begin = next_cache_3_begin_0, end = next_cache_3_end_0, end_mask = next_cache_3_end_mask_0, x = new_x_7_cast_fp16)[name = tensor<string, []>("next_cache_3_cast_fp16")];
+            tensor<int32, [3]> var_641_begin_0 = const()[name = tensor<string, []>("op_641_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> var_641_end_0 = const()[name = tensor<string, []>("op_641_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> var_641_end_mask_0 = const()[name = tensor<string, []>("op_641_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_641_cast_fp16 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = next_cache_3_cast_fp16)[name = tensor<string, []>("op_641_cast_fp16")];
+            tensor<string, []> x_47_pad_type_0 = const()[name = tensor<string, []>("x_47_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_47_groups_0 = const()[name = tensor<string, []>("x_47_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_47_strides_0 = const()[name = tensor<string, []>("x_47_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_47_pad_0 = const()[name = tensor<string, []>("x_47_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_47_dilations_0 = const()[name = tensor<string, []>("x_47_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_1_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24472384)))];
+            tensor<fp16, [1, 512, 3]> x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16, x = new_x_7_cast_fp16)[name = tensor<string, []>("x_47_cast_fp16")];
+            tensor<int32, [3]> input_107_perm_0 = const()[name = tensor<string, []>("input_107_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_49_axes_0 = const()[name = tensor<string, []>("x_49_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24481664)))];
+            tensor<fp16, [512]> encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24482752)))];
+            tensor<fp16, [1, 3, 512]> input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = tensor<string, []>("transpose_223")];
+            tensor<fp16, [1, 3, 512]> x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = tensor<string, []>("x_49_cast_fp16")];
+            tensor<int32, [3]> input_109_perm_0 = const()[name = tensor<string, []>("input_109_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 3]> input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = tensor<string, []>("transpose_222")];
+            tensor<fp16, [1, 512, 3]> input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
+            tensor<string, []> x_51_pad_type_0 = const()[name = tensor<string, []>("x_51_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_51_strides_0 = const()[name = tensor<string, []>("x_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_51_pad_0 = const()[name = tensor<string, []>("x_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_51_dilations_0 = const()[name = tensor<string, []>("x_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_51_groups_0 = const()[name = tensor<string, []>("x_51_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_1_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24483840)))];
+            tensor<fp16, [1, 512, 3]> x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("x_51_cast_fp16")];
+            tensor<int32, [3]> input_113_perm_0 = const()[name = tensor<string, []>("input_113_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 3, 512]> input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = tensor<string, []>("transpose_221")];
+            tensor<fp16, [1, 3, 512]> input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
+            tensor<int32, [1]> input_117_axes_0 = const()[name = tensor<string, []>("input_117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25008192)))];
+            tensor<fp16, [512]> encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25009280)))];
+            tensor<fp16, [1, 3, 512]> input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_1_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25010368)))];
+            tensor<fp16, [1, 3, 2048]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16, x = input_117_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_1_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27107584)))];
+            tensor<fp16, [1, 3, 512]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
+            tensor<fp16, []> var_682_to_fp16 = const()[name = tensor<string, []>("op_682_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_683_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_682_to_fp16)[name = tensor<string, []>("op_683_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_683_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
+            tensor<int32, [1]> input_129_axes_0 = const()[name = tensor<string, []>("input_129_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_1_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29204800)))];
+            tensor<fp16, [512]> encoder_layers_1_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29205888)))];
+            tensor<fp16, [1, 3, 512]> input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
+            tensor<int32, [4]> cache_9_begin_0 = const()[name = tensor<string, []>("cache_9_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_9_end_0 = const()[name = tensor<string, []>("cache_9_end_0"), val = tensor<int32, [4]>([3, 1, 70, 512])];
+            tensor<bool, [4]> cache_9_end_mask_0 = const()[name = tensor<string, []>("cache_9_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_9_squeeze_mask_0 = const()[name = tensor<string, []>("cache_9_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_9_cast_fp16 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_9_cast_fp16")];
+            tensor<int32, [4]> cache_11_begin_0 = const()[name = tensor<string, []>("cache_11_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_11_end_0 = const()[name = tensor<string, []>("cache_11_end_0"), val = tensor<int32, [4]>([3, 1, 512, 8])];
+            tensor<bool, [4]> cache_11_end_mask_0 = const()[name = tensor<string, []>("cache_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_11_squeeze_mask_0 = const()[name = tensor<string, []>("cache_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_11_cast_fp16 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_11_cast_fp16")];
+            tensor<int32, [1]> input_131_axes_0 = const()[name = tensor<string, []>("input_131_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29206976)))];
+            tensor<fp16, [512]> encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29208064)))];
+            tensor<fp16, [1, 3, 512]> input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("input_131_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_2_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29209152)))];
+            tensor<fp16, [1, 3, 2048]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16, x = input_131_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_2_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31306368)))];
+            tensor<fp16, [1, 3, 512]> linear_20_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16, x = input_135_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
+            tensor<fp16, []> var_717_to_fp16 = const()[name = tensor<string, []>("op_717_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_718_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_717_to_fp16)[name = tensor<string, []>("op_718_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_718_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
+            tensor<int32, [1]> key_5_axes_0 = const()[name = tensor<string, []>("key_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33403584)))];
+            tensor<fp16, [512]> encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33404672)))];
+            tensor<fp16, [1, 3, 512]> key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")];
+            tensor<bool, []> input_143_interleave_0 = const()[name = tensor<string, []>("input_143_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 73, 512]> input_143_cast_fp16 = concat(axis = var_64, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = tensor<string, []>("input_143_cast_fp16")];
+            tensor<int32, [3]> var_740_begin_0 = const()[name = tensor<string, []>("op_740_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_740_end_0 = const()[name = tensor<string, []>("op_740_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_740_end_mask_0 = const()[name = tensor<string, []>("op_740_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 69, 512]> var_740_cast_fp16 = slice_by_index(begin = var_740_begin_0, end = var_740_end_0, end_mask = var_740_end_mask_0, x = cache_9_cast_fp16)[name = tensor<string, []>("op_740_cast_fp16")];
+            tensor<int32, [3]> var_743_begin_0 = const()[name = tensor<string, []>("op_743_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_743_end_0 = const()[name = tensor<string, []>("op_743_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<bool, [3]> var_743_end_mask_0 = const()[name = tensor<string, []>("op_743_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 1, 512]> var_743_cast_fp16 = slice_by_index(begin = var_743_begin_0, end = var_743_end_0, end_mask = var_743_end_mask_0, x = key_5_cast_fp16)[name = tensor<string, []>("op_743_cast_fp16")];
+            tensor<bool, []> var_746_interleave_0 = const()[name = tensor<string, []>("op_746_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_746_cast_fp16 = concat(axis = var_64, interleave = var_746_interleave_0, values = (var_740_cast_fp16, var_743_cast_fp16))[name = tensor<string, []>("op_746_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_2_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33405760)))];
+            tensor<fp16, [1, 3, 512]> linear_21_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16, x = key_5_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
+            tensor<int32, [4]> var_750 = const()[name = tensor<string, []>("op_750"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 3, 8, 64]> q_13_cast_fp16 = reshape(shape = var_750, x = linear_21_cast_fp16)[name = tensor<string, []>("q_13_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_2_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33930112)))];
+            tensor<fp16, [1, 73, 512]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
+            tensor<int32, [4]> var_754 = const()[name = tensor<string, []>("op_754"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> k_9_cast_fp16 = reshape(shape = var_754, x = linear_22_cast_fp16)[name = tensor<string, []>("k_9_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_2_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34454464)))];
+            tensor<fp16, [1, 73, 512]> linear_23_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
+            tensor<int32, [4]> var_758 = const()[name = tensor<string, []>("op_758"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> v_5_cast_fp16 = reshape(shape = var_758, x = linear_23_cast_fp16)[name = tensor<string, []>("v_5_cast_fp16")];
+            tensor<int32, [4]> value_7_perm_0 = const()[name = tensor<string, []>("value_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34978816)))];
+            tensor<fp16, [1, 3, 8, 64]> var_770_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_770_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34979904)))];
+            tensor<fp16, [1, 3, 8, 64]> var_772_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_772_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_5_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_59_transpose_x_0 = const()[name = tensor<string, []>("x_59_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_59_transpose_y_0 = const()[name = tensor<string, []>("x_59_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 145]> var_774_to_fp16 = const()[name = tensor<string, []>("op_774_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34980992)))];
+            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_772_cast_fp16)[name = tensor<string, []>("transpose_219")];
+            tensor<fp16, [1, 8, 3, 145]> x_59_cast_fp16 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = var_774_to_fp16)[name = tensor<string, []>("x_59_cast_fp16")];
+            tensor<int32, [8]> x_61_pad_0 = const()[name = tensor<string, []>("x_61_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_61_mode_0 = const()[name = tensor<string, []>("x_61_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_49_to_fp16 = const()[name = tensor<string, []>("const_49_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 3, 146]> x_61_cast_fp16 = pad(constant_val = const_49_to_fp16, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59_cast_fp16)[name = tensor<string, []>("x_61_cast_fp16")];
+            tensor<int32, [4]> var_782 = const()[name = tensor<string, []>("op_782"), val = tensor<int32, [4]>([1, 8, -1, 3])];
+            tensor<fp16, [1, 8, 146, 3]> x_63_cast_fp16 = reshape(shape = var_782, x = x_61_cast_fp16)[name = tensor<string, []>("x_63_cast_fp16")];
+            tensor<int32, [4]> var_786_begin_0 = const()[name = tensor<string, []>("op_786_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_786_end_0 = const()[name = tensor<string, []>("op_786_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
+            tensor<bool, [4]> var_786_end_mask_0 = const()[name = tensor<string, []>("op_786_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 145, 3]> var_786_cast_fp16 = slice_by_index(begin = var_786_begin_0, end = var_786_end_0, end_mask = var_786_end_mask_0, x = x_63_cast_fp16)[name = tensor<string, []>("op_786_cast_fp16")];
+            tensor<int32, [4]> var_787 = const()[name = tensor<string, []>("op_787"), val = tensor<int32, [4]>([1, 8, 3, 145])];
+            tensor<fp16, [1, 8, 3, 145]> matrix_bd_9_cast_fp16 = reshape(shape = var_787, x = var_786_cast_fp16)[name = tensor<string, []>("matrix_bd_9_cast_fp16")];
+            tensor<bool, []> matrix_ac_5_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_5_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_5_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_5_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_55_perm_0 = const()[name = tensor<string, []>("transpose_55_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_56_perm_0 = const()[name = tensor<string, []>("transpose_56_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 73]> transpose_56 = transpose(perm = transpose_56_perm_0, x = k_9_cast_fp16)[name = tensor<string, []>("transpose_217")];
+            tensor<fp16, [1, 8, 3, 64]> transpose_55 = transpose(perm = transpose_55_perm_0, x = var_770_cast_fp16)[name = tensor<string, []>("transpose_218")];
+            tensor<fp16, [1, 8, 3, 73]> matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_55, y = transpose_56)[name = tensor<string, []>("matrix_ac_5_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_11_begin_0 = const()[name = tensor<string, []>("matrix_bd_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_11_end_0 = const()[name = tensor<string, []>("matrix_bd_11_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
+            tensor<bool, [4]> matrix_bd_11_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 3, 73]> matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9_cast_fp16)[name = tensor<string, []>("matrix_bd_11_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_796_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = tensor<string, []>("op_796_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_9_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_9_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_9_cast_fp16 = mul(x = var_796_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_9_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> scores_11_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_11_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_802_cast_fp16 = softmax(axis = var_62, x = scores_11_cast_fp16)[name = tensor<string, []>("op_802_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> input_145_cast_fp16 = select(a = var_40_to_fp16, b = var_802_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_145_cast_fp16")];
+            tensor<bool, []> x_65_transpose_x_0 = const()[name = tensor<string, []>("x_65_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_65_transpose_y_0 = const()[name = tensor<string, []>("x_65_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 73, 64]> value_7_cast_fp16 = transpose(perm = value_7_perm_0, x = v_5_cast_fp16)[name = tensor<string, []>("transpose_220")];
+            tensor<fp16, [1, 8, 3, 64]> x_65_cast_fp16 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145_cast_fp16, y = value_7_cast_fp16)[name = tensor<string, []>("x_65_cast_fp16")];
+            tensor<int32, [4]> var_806_perm_0 = const()[name = tensor<string, []>("op_806_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_807 = const()[name = tensor<string, []>("op_807"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 3, 8, 64]> var_806_cast_fp16 = transpose(perm = var_806_perm_0, x = x_65_cast_fp16)[name = tensor<string, []>("transpose_216")];
+            tensor<fp16, [1, 3, 512]> input_147_cast_fp16 = reshape(shape = var_807, x = var_806_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_2_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35129536)))];
+            tensor<fp16, [1, 3, 512]> linear_25_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")];
+            tensor<int32, [1]> x_69_axes_0 = const()[name = tensor<string, []>("x_69_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35653888)))];
+            tensor<fp16, [512]> encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35654976)))];
+            tensor<fp16, [1, 3, 512]> x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = tensor<string, []>("x_69_cast_fp16")];
+            tensor<int32, [3]> input_153_perm_0 = const()[name = tensor<string, []>("input_153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_155_pad_type_0 = const()[name = tensor<string, []>("input_155_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_155_strides_0 = const()[name = tensor<string, []>("input_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_155_pad_0 = const()[name = tensor<string, []>("input_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_155_dilations_0 = const()[name = tensor<string, []>("input_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_155_groups_0 = const()[name = tensor<string, []>("input_155_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_2_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35656064)))];
+            tensor<fp16, [1, 512, 3]> input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = tensor<string, []>("transpose_215")];
+            tensor<fp16, [1, 1024, 3]> input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor<string, []>("input_155_cast_fp16")];
+            tensor<int32, []> x_71_split_num_splits_0 = const()[name = tensor<string, []>("x_71_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_71_split_axis_0 = const()[name = tensor<string, []>("x_71_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 3]> x_71_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_71_split_cast_fp16_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155_cast_fp16)[name = tensor<string, []>("x_71_split_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = tensor<string, []>("x_71_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_71_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> input_157_cast_fp16 = select(a = var_40_to_fp16, b = x_71_cast_fp16, cond = var_418)[name = tensor<string, []>("input_157_cast_fp16")];
+            tensor<bool, []> new_x_11_interleave_0 = const()[name = tensor<string, []>("new_x_11_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 11]> new_x_11_cast_fp16 = concat(axis = var_62, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = tensor<string, []>("new_x_11_cast_fp16")];
+            tensor<int32, [3]> next_cache_5_begin_0 = const()[name = tensor<string, []>("next_cache_5_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_5_end_0 = const()[name = tensor<string, []>("next_cache_5_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> next_cache_5_end_mask_0 = const()[name = tensor<string, []>("next_cache_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 9]> next_cache_5_cast_fp16 = slice_by_index(begin = next_cache_5_begin_0, end = next_cache_5_end_0, end_mask = next_cache_5_end_mask_0, x = new_x_11_cast_fp16)[name = tensor<string, []>("next_cache_5_cast_fp16")];
+            tensor<int32, [3]> var_848_begin_0 = const()[name = tensor<string, []>("op_848_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> var_848_end_0 = const()[name = tensor<string, []>("op_848_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> var_848_end_mask_0 = const()[name = tensor<string, []>("op_848_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_848_cast_fp16 = slice_by_index(begin = var_848_begin_0, end = var_848_end_0, end_mask = var_848_end_mask_0, x = next_cache_5_cast_fp16)[name = tensor<string, []>("op_848_cast_fp16")];
+            tensor<string, []> x_73_pad_type_0 = const()[name = tensor<string, []>("x_73_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_73_groups_0 = const()[name = tensor<string, []>("x_73_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_73_strides_0 = const()[name = tensor<string, []>("x_73_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_73_pad_0 = const()[name = tensor<string, []>("x_73_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_73_dilations_0 = const()[name = tensor<string, []>("x_73_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_2_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36704704)))];
+            tensor<fp16, [1, 512, 3]> x_73_cast_fp16 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16, x = new_x_11_cast_fp16)[name = tensor<string, []>("x_73_cast_fp16")];
+            tensor<int32, [3]> input_159_perm_0 = const()[name = tensor<string, []>("input_159_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_75_axes_0 = const()[name = tensor<string, []>("x_75_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36713984)))];
+            tensor<fp16, [512]> encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36715072)))];
+            tensor<fp16, [1, 3, 512]> input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = tensor<string, []>("transpose_214")];
+            tensor<fp16, [1, 3, 512]> x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = tensor<string, []>("x_75_cast_fp16")];
+            tensor<int32, [3]> input_161_perm_0 = const()[name = tensor<string, []>("input_161_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 3]> input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = tensor<string, []>("transpose_213")];
+            tensor<fp16, [1, 512, 3]> input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
+            tensor<string, []> x_77_pad_type_0 = const()[name = tensor<string, []>("x_77_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_77_strides_0 = const()[name = tensor<string, []>("x_77_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_77_pad_0 = const()[name = tensor<string, []>("x_77_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_77_dilations_0 = const()[name = tensor<string, []>("x_77_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_77_groups_0 = const()[name = tensor<string, []>("x_77_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_2_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36716160)))];
+            tensor<fp16, [1, 512, 3]> x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16, x = input_163_cast_fp16)[name = tensor<string, []>("x_77_cast_fp16")];
+            tensor<int32, [3]> input_165_perm_0 = const()[name = tensor<string, []>("input_165_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 3, 512]> input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = tensor<string, []>("transpose_212")];
+            tensor<fp16, [1, 3, 512]> input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")];
+            tensor<int32, [1]> input_169_axes_0 = const()[name = tensor<string, []>("input_169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37240512)))];
+            tensor<fp16, [512]> encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37241600)))];
+            tensor<fp16, [1, 3, 512]> input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor<string, []>("input_169_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_2_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37242688)))];
+            tensor<fp16, [1, 3, 2048]> linear_26_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16, x = input_169_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_2_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39339904)))];
+            tensor<fp16, [1, 3, 512]> linear_27_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")];
+            tensor<fp16, []> var_889_to_fp16 = const()[name = tensor<string, []>("op_889_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_890_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_889_to_fp16)[name = tensor<string, []>("op_890_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_890_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
+            tensor<int32, [1]> input_181_axes_0 = const()[name = tensor<string, []>("input_181_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_2_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41437120)))];
+            tensor<fp16, [512]> encoder_layers_2_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41438208)))];
+            tensor<fp16, [1, 3, 512]> input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")];
+            tensor<int32, [4]> cache_13_begin_0 = const()[name = tensor<string, []>("cache_13_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_13_end_0 = const()[name = tensor<string, []>("cache_13_end_0"), val = tensor<int32, [4]>([4, 1, 70, 512])];
+            tensor<bool, [4]> cache_13_end_mask_0 = const()[name = tensor<string, []>("cache_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_13_squeeze_mask_0 = const()[name = tensor<string, []>("cache_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_13_cast_fp16 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_13_cast_fp16")];
+            tensor<int32, [4]> cache_15_begin_0 = const()[name = tensor<string, []>("cache_15_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_15_end_0 = const()[name = tensor<string, []>("cache_15_end_0"), val = tensor<int32, [4]>([4, 1, 512, 8])];
+            tensor<bool, [4]> cache_15_end_mask_0 = const()[name = tensor<string, []>("cache_15_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_15_squeeze_mask_0 = const()[name = tensor<string, []>("cache_15_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_15_cast_fp16 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_15_cast_fp16")];
+            tensor<int32, [1]> input_183_axes_0 = const()[name = tensor<string, []>("input_183_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41439296)))];
+            tensor<fp16, [512]> encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41440384)))];
+            tensor<fp16, [1, 3, 512]> input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor<string, []>("input_183_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_3_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41441472)))];
+            tensor<fp16, [1, 3, 2048]> linear_28_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = tensor<string, []>("input_187_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_3_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43538688)))];
+            tensor<fp16, [1, 3, 512]> linear_29_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")];
+            tensor<fp16, []> var_924_to_fp16 = const()[name = tensor<string, []>("op_924_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_925_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_924_to_fp16)[name = tensor<string, []>("op_925_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_925_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
+            tensor<int32, [1]> key_7_axes_0 = const()[name = tensor<string, []>("key_7_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45635904)))];
+            tensor<fp16, [512]> encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45636992)))];
+            tensor<fp16, [1, 3, 512]> key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = tensor<string, []>("key_7_cast_fp16")];
+            tensor<bool, []> input_195_interleave_0 = const()[name = tensor<string, []>("input_195_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 73, 512]> input_195_cast_fp16 = concat(axis = var_64, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = tensor<string, []>("input_195_cast_fp16")];
+            tensor<int32, [3]> var_947_begin_0 = const()[name = tensor<string, []>("op_947_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_947_end_0 = const()[name = tensor<string, []>("op_947_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_947_end_mask_0 = const()[name = tensor<string, []>("op_947_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 69, 512]> var_947_cast_fp16 = slice_by_index(begin = var_947_begin_0, end = var_947_end_0, end_mask = var_947_end_mask_0, x = cache_13_cast_fp16)[name = tensor<string, []>("op_947_cast_fp16")];
+            tensor<int32, [3]> var_950_begin_0 = const()[name = tensor<string, []>("op_950_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_950_end_0 = const()[name = tensor<string, []>("op_950_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<bool, [3]> var_950_end_mask_0 = const()[name = tensor<string, []>("op_950_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 1, 512]> var_950_cast_fp16 = slice_by_index(begin = var_950_begin_0, end = var_950_end_0, end_mask = var_950_end_mask_0, x = key_7_cast_fp16)[name = tensor<string, []>("op_950_cast_fp16")];
+            tensor<bool, []> var_953_interleave_0 = const()[name = tensor<string, []>("op_953_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_953_cast_fp16 = concat(axis = var_64, interleave = var_953_interleave_0, values = (var_947_cast_fp16, var_950_cast_fp16))[name = tensor<string, []>("op_953_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_3_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45638080)))];
+            tensor<fp16, [1, 3, 512]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16, x = key_7_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")];
+            tensor<int32, [4]> var_957 = const()[name = tensor<string, []>("op_957"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 3, 8, 64]> q_19_cast_fp16 = reshape(shape = var_957, x = linear_30_cast_fp16)[name = tensor<string, []>("q_19_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_3_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46162432)))];
+            tensor<fp16, [1, 73, 512]> linear_31_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")];
+            tensor<int32, [4]> var_961 = const()[name = tensor<string, []>("op_961"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> k_13_cast_fp16 = reshape(shape = var_961, x = linear_31_cast_fp16)[name = tensor<string, []>("k_13_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_3_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46686784)))];
+            tensor<fp16, [1, 73, 512]> linear_32_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")];
+            tensor<int32, [4]> var_965 = const()[name = tensor<string, []>("op_965"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> v_7_cast_fp16 = reshape(shape = var_965, x = linear_32_cast_fp16)[name = tensor<string, []>("v_7_cast_fp16")];
+            tensor<int32, [4]> value_9_perm_0 = const()[name = tensor<string, []>("value_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47211136)))];
+            tensor<fp16, [1, 3, 8, 64]> var_977_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_977_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47212224)))];
+            tensor<fp16, [1, 3, 8, 64]> var_979_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_979_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_7_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_85_transpose_x_0 = const()[name = tensor<string, []>("x_85_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_85_transpose_y_0 = const()[name = tensor<string, []>("x_85_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 145]> var_981_to_fp16 = const()[name = tensor<string, []>("op_981_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47213312)))];
+            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_979_cast_fp16)[name = tensor<string, []>("transpose_210")];
+            tensor<fp16, [1, 8, 3, 145]> x_85_cast_fp16 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = var_981_to_fp16)[name = tensor<string, []>("x_85_cast_fp16")];
+            tensor<int32, [8]> x_87_pad_0 = const()[name = tensor<string, []>("x_87_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_87_mode_0 = const()[name = tensor<string, []>("x_87_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_62_to_fp16 = const()[name = tensor<string, []>("const_62_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 3, 146]> x_87_cast_fp16 = pad(constant_val = const_62_to_fp16, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85_cast_fp16)[name = tensor<string, []>("x_87_cast_fp16")];
+            tensor<int32, [4]> var_989 = const()[name = tensor<string, []>("op_989"), val = tensor<int32, [4]>([1, 8, -1, 3])];
+            tensor<fp16, [1, 8, 146, 3]> x_89_cast_fp16 = reshape(shape = var_989, x = x_87_cast_fp16)[name = tensor<string, []>("x_89_cast_fp16")];
+            tensor<int32, [4]> var_993_begin_0 = const()[name = tensor<string, []>("op_993_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_993_end_0 = const()[name = tensor<string, []>("op_993_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
+            tensor<bool, [4]> var_993_end_mask_0 = const()[name = tensor<string, []>("op_993_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 145, 3]> var_993_cast_fp16 = slice_by_index(begin = var_993_begin_0, end = var_993_end_0, end_mask = var_993_end_mask_0, x = x_89_cast_fp16)[name = tensor<string, []>("op_993_cast_fp16")];
+            tensor<int32, [4]> var_994 = const()[name = tensor<string, []>("op_994"), val = tensor<int32, [4]>([1, 8, 3, 145])];
+            tensor<fp16, [1, 8, 3, 145]> matrix_bd_13_cast_fp16 = reshape(shape = var_994, x = var_993_cast_fp16)[name = tensor<string, []>("matrix_bd_13_cast_fp16")];
+            tensor<bool, []> matrix_ac_7_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_7_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_57_perm_0 = const()[name = tensor<string, []>("transpose_57_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_58_perm_0 = const()[name = tensor<string, []>("transpose_58_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 73]> transpose_58 = transpose(perm = transpose_58_perm_0, x = k_13_cast_fp16)[name = tensor<string, []>("transpose_208")];
+            tensor<fp16, [1, 8, 3, 64]> transpose_57 = transpose(perm = transpose_57_perm_0, x = var_977_cast_fp16)[name = tensor<string, []>("transpose_209")];
+            tensor<fp16, [1, 8, 3, 73]> matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_57, y = transpose_58)[name = tensor<string, []>("matrix_ac_7_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_15_begin_0 = const()[name = tensor<string, []>("matrix_bd_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_15_end_0 = const()[name = tensor<string, []>("matrix_bd_15_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
+            tensor<bool, [4]> matrix_bd_15_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 3, 73]> matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13_cast_fp16)[name = tensor<string, []>("matrix_bd_15_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_1003_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = tensor<string, []>("op_1003_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_13_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_13_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_13_cast_fp16 = mul(x = var_1003_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_13_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> scores_15_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_15_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_1009_cast_fp16 = softmax(axis = var_62, x = scores_15_cast_fp16)[name = tensor<string, []>("op_1009_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> input_197_cast_fp16 = select(a = var_40_to_fp16, b = var_1009_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_197_cast_fp16")];
+            tensor<bool, []> x_91_transpose_x_0 = const()[name = tensor<string, []>("x_91_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_91_transpose_y_0 = const()[name = tensor<string, []>("x_91_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 73, 64]> value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_7_cast_fp16)[name = tensor<string, []>("transpose_211")];
+            tensor<fp16, [1, 8, 3, 64]> x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197_cast_fp16, y = value_9_cast_fp16)[name = tensor<string, []>("x_91_cast_fp16")];
+            tensor<int32, [4]> var_1013_perm_0 = const()[name = tensor<string, []>("op_1013_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1014 = const()[name = tensor<string, []>("op_1014"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 3, 8, 64]> var_1013_cast_fp16 = transpose(perm = var_1013_perm_0, x = x_91_cast_fp16)[name = tensor<string, []>("transpose_207")];
+            tensor<fp16, [1, 3, 512]> input_199_cast_fp16 = reshape(shape = var_1014, x = var_1013_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_3_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47361856)))];
+            tensor<fp16, [1, 3, 512]> linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16, x = input_199_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")];
+            tensor<int32, [1]> x_95_axes_0 = const()[name = tensor<string, []>("x_95_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47886208)))];
+            tensor<fp16, [512]> encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47887296)))];
+            tensor<fp16, [1, 3, 512]> x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = tensor<string, []>("x_95_cast_fp16")];
+            tensor<int32, [3]> input_205_perm_0 = const()[name = tensor<string, []>("input_205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_207_pad_type_0 = const()[name = tensor<string, []>("input_207_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_207_strides_0 = const()[name = tensor<string, []>("input_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_207_pad_0 = const()[name = tensor<string, []>("input_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_207_dilations_0 = const()[name = tensor<string, []>("input_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_207_groups_0 = const()[name = tensor<string, []>("input_207_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_3_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47888384)))];
+            tensor<fp16, [1, 512, 3]> input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = tensor<string, []>("transpose_206")];
+            tensor<fp16, [1, 1024, 3]> input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")];
+            tensor<int32, []> x_97_split_num_splits_0 = const()[name = tensor<string, []>("x_97_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_97_split_axis_0 = const()[name = tensor<string, []>("x_97_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 3]> x_97_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207_cast_fp16)[name = tensor<string, []>("x_97_split_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = tensor<string, []>("x_97_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_97_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> input_209_cast_fp16 = select(a = var_40_to_fp16, b = x_97_cast_fp16, cond = var_418)[name = tensor<string, []>("input_209_cast_fp16")];
+            tensor<bool, []> new_x_15_interleave_0 = const()[name = tensor<string, []>("new_x_15_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 11]> new_x_15_cast_fp16 = concat(axis = var_62, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = tensor<string, []>("new_x_15_cast_fp16")];
+            tensor<int32, [3]> next_cache_7_begin_0 = const()[name = tensor<string, []>("next_cache_7_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_7_end_0 = const()[name = tensor<string, []>("next_cache_7_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> next_cache_7_end_mask_0 = const()[name = tensor<string, []>("next_cache_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 9]> next_cache_7_cast_fp16 = slice_by_index(begin = next_cache_7_begin_0, end = next_cache_7_end_0, end_mask = next_cache_7_end_mask_0, x = new_x_15_cast_fp16)[name = tensor<string, []>("next_cache_7_cast_fp16")];
+            tensor<int32, [3]> var_1055_begin_0 = const()[name = tensor<string, []>("op_1055_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> var_1055_end_0 = const()[name = tensor<string, []>("op_1055_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> var_1055_end_mask_0 = const()[name = tensor<string, []>("op_1055_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_1055_cast_fp16 = slice_by_index(begin = var_1055_begin_0, end = var_1055_end_0, end_mask = var_1055_end_mask_0, x = next_cache_7_cast_fp16)[name = tensor<string, []>("op_1055_cast_fp16")];
+            tensor<string, []> x_99_pad_type_0 = const()[name = tensor<string, []>("x_99_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_99_groups_0 = const()[name = tensor<string, []>("x_99_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_99_strides_0 = const()[name = tensor<string, []>("x_99_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_99_pad_0 = const()[name = tensor<string, []>("x_99_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_99_dilations_0 = const()[name = tensor<string, []>("x_99_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_3_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48937024)))];
+            tensor<fp16, [1, 512, 3]> x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16, x = new_x_15_cast_fp16)[name = tensor<string, []>("x_99_cast_fp16")];
+            tensor<int32, [3]> input_211_perm_0 = const()[name = tensor<string, []>("input_211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_101_axes_0 = const()[name = tensor<string, []>("x_101_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48946304)))];
+            tensor<fp16, [512]> encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48947392)))];
+            tensor<fp16, [1, 3, 512]> input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = tensor<string, []>("transpose_205")];
+            tensor<fp16, [1, 3, 512]> x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = tensor<string, []>("x_101_cast_fp16")];
+            tensor<int32, [3]> input_213_perm_0 = const()[name = tensor<string, []>("input_213_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 3]> input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = tensor<string, []>("transpose_204")];
+            tensor<fp16, [1, 512, 3]> input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = tensor<string, []>("input_215_cast_fp16")];
+            tensor<string, []> x_103_pad_type_0 = const()[name = tensor<string, []>("x_103_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_103_strides_0 = const()[name = tensor<string, []>("x_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_103_pad_0 = const()[name = tensor<string, []>("x_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_103_dilations_0 = const()[name = tensor<string, []>("x_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_103_groups_0 = const()[name = tensor<string, []>("x_103_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_3_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48948480)))];
+            tensor<fp16, [1, 512, 3]> x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor<string, []>("x_103_cast_fp16")];
+            tensor<int32, [3]> input_217_perm_0 = const()[name = tensor<string, []>("input_217_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 3, 512]> input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = tensor<string, []>("transpose_203")];
+            tensor<fp16, [1, 3, 512]> input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
+            tensor<int32, [1]> input_221_axes_0 = const()[name = tensor<string, []>("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49472832)))];
+            tensor<fp16, [512]> encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49473920)))];
+            tensor<fp16, [1, 3, 512]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = tensor<string, []>("input_221_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_3_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49475008)))];
+            tensor<fp16, [1, 3, 2048]> linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16, x = input_221_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = tensor<string, []>("input_225_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_3_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51572224)))];
+            tensor<fp16, [1, 3, 512]> linear_36_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor<string, []>("linear_36_cast_fp16")];
+            tensor<fp16, []> var_1096_to_fp16 = const()[name = tensor<string, []>("op_1096_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_1097_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1096_to_fp16)[name = tensor<string, []>("op_1097_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1097_cast_fp16)[name = tensor<string, []>("input_231_cast_fp16")];
+            tensor<int32, [1]> input_233_axes_0 = const()[name = tensor<string, []>("input_233_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_3_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53669440)))];
+            tensor<fp16, [512]> encoder_layers_3_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53670528)))];
+            tensor<fp16, [1, 3, 512]> input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = tensor<string, []>("input_233_cast_fp16")];
+            tensor<int32, [4]> cache_17_begin_0 = const()[name = tensor<string, []>("cache_17_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_17_end_0 = const()[name = tensor<string, []>("cache_17_end_0"), val = tensor<int32, [4]>([5, 1, 70, 512])];
+            tensor<bool, [4]> cache_17_end_mask_0 = const()[name = tensor<string, []>("cache_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_17_squeeze_mask_0 = const()[name = tensor<string, []>("cache_17_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_17_cast_fp16 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_17_cast_fp16")];
+            tensor<int32, [4]> cache_19_begin_0 = const()[name = tensor<string, []>("cache_19_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_19_end_0 = const()[name = tensor<string, []>("cache_19_end_0"), val = tensor<int32, [4]>([5, 1, 512, 8])];
+            tensor<bool, [4]> cache_19_end_mask_0 = const()[name = tensor<string, []>("cache_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_19_squeeze_mask_0 = const()[name = tensor<string, []>("cache_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_19_cast_fp16 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_19_cast_fp16")];
+            tensor<int32, [1]> input_235_axes_0 = const()[name = tensor<string, []>("input_235_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53671616)))];
+            tensor<fp16, [512]> encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53672704)))];
+            tensor<fp16, [1, 3, 512]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = tensor<string, []>("input_235_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_4_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53673792)))];
+            tensor<fp16, [1, 3, 2048]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16, x = input_235_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = tensor<string, []>("input_239_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_4_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55771008)))];
+            tensor<fp16, [1, 3, 512]> linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")];
+            tensor<fp16, []> var_1131_to_fp16 = const()[name = tensor<string, []>("op_1131_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_1132_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1131_to_fp16)[name = tensor<string, []>("op_1132_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1132_cast_fp16)[name = tensor<string, []>("input_245_cast_fp16")];
+            tensor<int32, [1]> key_9_axes_0 = const()[name = tensor<string, []>("key_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57868224)))];
+            tensor<fp16, [512]> encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57869312)))];
+            tensor<fp16, [1, 3, 512]> key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")];
+            tensor<bool, []> input_247_interleave_0 = const()[name = tensor<string, []>("input_247_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 73, 512]> input_247_cast_fp16 = concat(axis = var_64, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = tensor<string, []>("input_247_cast_fp16")];
+            tensor<int32, [3]> var_1154_begin_0 = const()[name = tensor<string, []>("op_1154_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_1154_end_0 = const()[name = tensor<string, []>("op_1154_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_1154_end_mask_0 = const()[name = tensor<string, []>("op_1154_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 69, 512]> var_1154_cast_fp16 = slice_by_index(begin = var_1154_begin_0, end = var_1154_end_0, end_mask = var_1154_end_mask_0, x = cache_17_cast_fp16)[name = tensor<string, []>("op_1154_cast_fp16")];
+            tensor<int32, [3]> var_1157_begin_0 = const()[name = tensor<string, []>("op_1157_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_1157_end_0 = const()[name = tensor<string, []>("op_1157_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<bool, [3]> var_1157_end_mask_0 = const()[name = tensor<string, []>("op_1157_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 1, 512]> var_1157_cast_fp16 = slice_by_index(begin = var_1157_begin_0, end = var_1157_end_0, end_mask = var_1157_end_mask_0, x = key_9_cast_fp16)[name = tensor<string, []>("op_1157_cast_fp16")];
+            tensor<bool, []> var_1160_interleave_0 = const()[name = tensor<string, []>("op_1160_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_1160_cast_fp16 = concat(axis = var_64, interleave = var_1160_interleave_0, values = (var_1154_cast_fp16, var_1157_cast_fp16))[name = tensor<string, []>("op_1160_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_4_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57870400)))];
+            tensor<fp16, [1, 3, 512]> linear_39_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16, x = key_9_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")];
+            tensor<int32, [4]> var_1164 = const()[name = tensor<string, []>("op_1164"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 3, 8, 64]> q_25_cast_fp16 = reshape(shape = var_1164, x = linear_39_cast_fp16)[name = tensor<string, []>("q_25_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_4_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58394752)))];
+            tensor<fp16, [1, 73, 512]> linear_40_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16, x = input_247_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")];
+            tensor<int32, [4]> var_1168 = const()[name = tensor<string, []>("op_1168"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> k_17_cast_fp16 = reshape(shape = var_1168, x = linear_40_cast_fp16)[name = tensor<string, []>("k_17_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_4_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58919104)))];
+            tensor<fp16, [1, 73, 512]> linear_41_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16, x = input_247_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")];
+            tensor<int32, [4]> var_1172 = const()[name = tensor<string, []>("op_1172"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> v_9_cast_fp16 = reshape(shape = var_1172, x = linear_41_cast_fp16)[name = tensor<string, []>("v_9_cast_fp16")];
+            tensor<int32, [4]> value_11_perm_0 = const()[name = tensor<string, []>("value_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59443456)))];
+            tensor<fp16, [1, 3, 8, 64]> var_1184_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1184_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59444544)))];
+            tensor<fp16, [1, 3, 8, 64]> var_1186_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1186_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_9_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_111_transpose_x_0 = const()[name = tensor<string, []>("x_111_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_111_transpose_y_0 = const()[name = tensor<string, []>("x_111_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 145]> var_1188_to_fp16 = const()[name = tensor<string, []>("op_1188_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59445632)))];
+            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1186_cast_fp16)[name = tensor<string, []>("transpose_201")];
+            tensor<fp16, [1, 8, 3, 145]> x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = var_1188_to_fp16)[name = tensor<string, []>("x_111_cast_fp16")];
+            tensor<int32, [8]> x_113_pad_0 = const()[name = tensor<string, []>("x_113_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_113_mode_0 = const()[name = tensor<string, []>("x_113_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_75_to_fp16 = const()[name = tensor<string, []>("const_75_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 3, 146]> x_113_cast_fp16 = pad(constant_val = const_75_to_fp16, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111_cast_fp16)[name = tensor<string, []>("x_113_cast_fp16")];
+            tensor<int32, [4]> var_1196 = const()[name = tensor<string, []>("op_1196"), val = tensor<int32, [4]>([1, 8, -1, 3])];
+            tensor<fp16, [1, 8, 146, 3]> x_115_cast_fp16 = reshape(shape = var_1196, x = x_113_cast_fp16)[name = tensor<string, []>("x_115_cast_fp16")];
+            tensor<int32, [4]> var_1200_begin_0 = const()[name = tensor<string, []>("op_1200_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1200_end_0 = const()[name = tensor<string, []>("op_1200_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
+            tensor<bool, [4]> var_1200_end_mask_0 = const()[name = tensor<string, []>("op_1200_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 145, 3]> var_1200_cast_fp16 = slice_by_index(begin = var_1200_begin_0, end = var_1200_end_0, end_mask = var_1200_end_mask_0, x = x_115_cast_fp16)[name = tensor<string, []>("op_1200_cast_fp16")];
+            tensor<int32, [4]> var_1201 = const()[name = tensor<string, []>("op_1201"), val = tensor<int32, [4]>([1, 8, 3, 145])];
+            tensor<fp16, [1, 8, 3, 145]> matrix_bd_17_cast_fp16 = reshape(shape = var_1201, x = var_1200_cast_fp16)[name = tensor<string, []>("matrix_bd_17_cast_fp16")];
+            tensor<bool, []> matrix_ac_9_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_9_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_9_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_9_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_59_perm_0 = const()[name = tensor<string, []>("transpose_59_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_60_perm_0 = const()[name = tensor<string, []>("transpose_60_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 73]> transpose_60 = transpose(perm = transpose_60_perm_0, x = k_17_cast_fp16)[name = tensor<string, []>("transpose_199")];
+            tensor<fp16, [1, 8, 3, 64]> transpose_59 = transpose(perm = transpose_59_perm_0, x = var_1184_cast_fp16)[name = tensor<string, []>("transpose_200")];
+            tensor<fp16, [1, 8, 3, 73]> matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_59, y = transpose_60)[name = tensor<string, []>("matrix_ac_9_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_19_begin_0 = const()[name = tensor<string, []>("matrix_bd_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_19_end_0 = const()[name = tensor<string, []>("matrix_bd_19_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
+            tensor<bool, [4]> matrix_bd_19_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 3, 73]> matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17_cast_fp16)[name = tensor<string, []>("matrix_bd_19_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_1210_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = tensor<string, []>("op_1210_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_17_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_17_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_17_cast_fp16 = mul(x = var_1210_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_17_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> scores_19_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_19_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_1216_cast_fp16 = softmax(axis = var_62, x = scores_19_cast_fp16)[name = tensor<string, []>("op_1216_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> input_249_cast_fp16 = select(a = var_40_to_fp16, b = var_1216_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_249_cast_fp16")];
+            tensor<bool, []> x_117_transpose_x_0 = const()[name = tensor<string, []>("x_117_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_117_transpose_y_0 = const()[name = tensor<string, []>("x_117_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 73, 64]> value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_9_cast_fp16)[name = tensor<string, []>("transpose_202")];
+            tensor<fp16, [1, 8, 3, 64]> x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249_cast_fp16, y = value_11_cast_fp16)[name = tensor<string, []>("x_117_cast_fp16")];
+            tensor<int32, [4]> var_1220_perm_0 = const()[name = tensor<string, []>("op_1220_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1221 = const()[name = tensor<string, []>("op_1221"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 3, 8, 64]> var_1220_cast_fp16 = transpose(perm = var_1220_perm_0, x = x_117_cast_fp16)[name = tensor<string, []>("transpose_198")];
+            tensor<fp16, [1, 3, 512]> input_251_cast_fp16 = reshape(shape = var_1221, x = var_1220_cast_fp16)[name = tensor<string, []>("input_251_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_4_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59594176)))];
+            tensor<fp16, [1, 3, 512]> linear_43_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16, x = input_251_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = tensor<string, []>("input_255_cast_fp16")];
+            tensor<int32, [1]> x_121_axes_0 = const()[name = tensor<string, []>("x_121_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60118528)))];
+            tensor<fp16, [512]> encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60119616)))];
+            tensor<fp16, [1, 3, 512]> x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = tensor<string, []>("x_121_cast_fp16")];
+            tensor<int32, [3]> input_257_perm_0 = const()[name = tensor<string, []>("input_257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_259_pad_type_0 = const()[name = tensor<string, []>("input_259_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_259_strides_0 = const()[name = tensor<string, []>("input_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_259_pad_0 = const()[name = tensor<string, []>("input_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_259_dilations_0 = const()[name = tensor<string, []>("input_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_259_groups_0 = const()[name = tensor<string, []>("input_259_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_4_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60120704)))];
+            tensor<fp16, [1, 512, 3]> input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = tensor<string, []>("transpose_197")];
+            tensor<fp16, [1, 1024, 3]> input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16, x = input_257_cast_fp16)[name = tensor<string, []>("input_259_cast_fp16")];
+            tensor<int32, []> x_123_split_num_splits_0 = const()[name = tensor<string, []>("x_123_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_123_split_axis_0 = const()[name = tensor<string, []>("x_123_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 3]> x_123_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_123_split_cast_fp16_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259_cast_fp16)[name = tensor<string, []>("x_123_split_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = tensor<string, []>("x_123_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_123_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> input_261_cast_fp16 = select(a = var_40_to_fp16, b = x_123_cast_fp16, cond = var_418)[name = tensor<string, []>("input_261_cast_fp16")];
+            tensor<bool, []> new_x_19_interleave_0 = const()[name = tensor<string, []>("new_x_19_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 11]> new_x_19_cast_fp16 = concat(axis = var_62, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = tensor<string, []>("new_x_19_cast_fp16")];
+            tensor<int32, [3]> next_cache_9_begin_0 = const()[name = tensor<string, []>("next_cache_9_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_9_end_0 = const()[name = tensor<string, []>("next_cache_9_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> next_cache_9_end_mask_0 = const()[name = tensor<string, []>("next_cache_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 9]> next_cache_9_cast_fp16 = slice_by_index(begin = next_cache_9_begin_0, end = next_cache_9_end_0, end_mask = next_cache_9_end_mask_0, x = new_x_19_cast_fp16)[name = tensor<string, []>("next_cache_9_cast_fp16")];
+            tensor<int32, [3]> var_1262_begin_0 = const()[name = tensor<string, []>("op_1262_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> var_1262_end_0 = const()[name = tensor<string, []>("op_1262_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> var_1262_end_mask_0 = const()[name = tensor<string, []>("op_1262_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_1262_cast_fp16 = slice_by_index(begin = var_1262_begin_0, end = var_1262_end_0, end_mask = var_1262_end_mask_0, x = next_cache_9_cast_fp16)[name = tensor<string, []>("op_1262_cast_fp16")];
+            tensor<string, []> x_125_pad_type_0 = const()[name = tensor<string, []>("x_125_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_125_groups_0 = const()[name = tensor<string, []>("x_125_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_125_strides_0 = const()[name = tensor<string, []>("x_125_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_125_pad_0 = const()[name = tensor<string, []>("x_125_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_125_dilations_0 = const()[name = tensor<string, []>("x_125_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_4_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61169344)))];
+            tensor<fp16, [1, 512, 3]> x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16, x = new_x_19_cast_fp16)[name = tensor<string, []>("x_125_cast_fp16")];
+            tensor<int32, [3]> input_263_perm_0 = const()[name = tensor<string, []>("input_263_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_127_axes_0 = const()[name = tensor<string, []>("x_127_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61178624)))];
+            tensor<fp16, [512]> encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61179712)))];
+            tensor<fp16, [1, 3, 512]> input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = tensor<string, []>("transpose_196")];
+            tensor<fp16, [1, 3, 512]> x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = tensor<string, []>("x_127_cast_fp16")];
+            tensor<int32, [3]> input_265_perm_0 = const()[name = tensor<string, []>("input_265_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 3]> input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = tensor<string, []>("transpose_195")];
+            tensor<fp16, [1, 512, 3]> input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = tensor<string, []>("input_267_cast_fp16")];
+            tensor<string, []> x_129_pad_type_0 = const()[name = tensor<string, []>("x_129_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_129_strides_0 = const()[name = tensor<string, []>("x_129_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_129_pad_0 = const()[name = tensor<string, []>("x_129_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_129_dilations_0 = const()[name = tensor<string, []>("x_129_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_129_groups_0 = const()[name = tensor<string, []>("x_129_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_4_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61180800)))];
+            tensor<fp16, [1, 512, 3]> x_129_cast_fp16 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16, x = input_267_cast_fp16)[name = tensor<string, []>("x_129_cast_fp16")];
+            tensor<int32, [3]> input_269_perm_0 = const()[name = tensor<string, []>("input_269_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 3, 512]> input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = tensor<string, []>("transpose_194")];
+            tensor<fp16, [1, 3, 512]> input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = tensor<string, []>("input_271_cast_fp16")];
+            tensor<int32, [1]> input_273_axes_0 = const()[name = tensor<string, []>("input_273_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61705152)))];
+            tensor<fp16, [512]> encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61706240)))];
+            tensor<fp16, [1, 3, 512]> input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor<string, []>("input_273_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_4_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61707328)))];
+            tensor<fp16, [1, 3, 2048]> linear_44_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16, x = input_273_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = tensor<string, []>("input_277_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_4_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63804544)))];
+            tensor<fp16, [1, 3, 512]> linear_45_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor<string, []>("linear_45_cast_fp16")];
+            tensor<fp16, []> var_1303_to_fp16 = const()[name = tensor<string, []>("op_1303_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_1304_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1303_to_fp16)[name = tensor<string, []>("op_1304_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1304_cast_fp16)[name = tensor<string, []>("input_283_cast_fp16")];
+            tensor<int32, [1]> input_285_axes_0 = const()[name = tensor<string, []>("input_285_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_4_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65901760)))];
+            tensor<fp16, [512]> encoder_layers_4_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65902848)))];
+            tensor<fp16, [1, 3, 512]> input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = tensor<string, []>("input_285_cast_fp16")];
+            tensor<int32, [4]> cache_21_begin_0 = const()[name = tensor<string, []>("cache_21_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_21_end_0 = const()[name = tensor<string, []>("cache_21_end_0"), val = tensor<int32, [4]>([6, 1, 70, 512])];
+            tensor<bool, [4]> cache_21_end_mask_0 = const()[name = tensor<string, []>("cache_21_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_21_squeeze_mask_0 = const()[name = tensor<string, []>("cache_21_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_21_cast_fp16 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_21_cast_fp16")];
+            tensor<int32, [4]> cache_23_begin_0 = const()[name = tensor<string, []>("cache_23_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_23_end_0 = const()[name = tensor<string, []>("cache_23_end_0"), val = tensor<int32, [4]>([6, 1, 512, 8])];
+            tensor<bool, [4]> cache_23_end_mask_0 = const()[name = tensor<string, []>("cache_23_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_23_squeeze_mask_0 = const()[name = tensor<string, []>("cache_23_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_23_cast_fp16 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_23_cast_fp16")];
+            tensor<int32, [1]> input_287_axes_0 = const()[name = tensor<string, []>("input_287_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65903936)))];
+            tensor<fp16, [512]> encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65905024)))];
+            tensor<fp16, [1, 3, 512]> input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = tensor<string, []>("input_287_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_5_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65906112)))];
+            tensor<fp16, [1, 3, 2048]> linear_46_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor<string, []>("linear_46_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = tensor<string, []>("input_291_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_5_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68003328)))];
+            tensor<fp16, [1, 3, 512]> linear_47_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16, x = input_291_cast_fp16)[name = tensor<string, []>("linear_47_cast_fp16")];
+            tensor<fp16, []> var_1338_to_fp16 = const()[name = tensor<string, []>("op_1338_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_1339_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1338_to_fp16)[name = tensor<string, []>("op_1339_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1339_cast_fp16)[name = tensor<string, []>("input_297_cast_fp16")];
+            tensor<int32, [1]> key_11_axes_0 = const()[name = tensor<string, []>("key_11_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70100544)))];
+            tensor<fp16, [512]> encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70101632)))];
+            tensor<fp16, [1, 3, 512]> key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = tensor<string, []>("key_11_cast_fp16")];
+            tensor<bool, []> input_299_interleave_0 = const()[name = tensor<string, []>("input_299_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 73, 512]> input_299_cast_fp16 = concat(axis = var_64, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = tensor<string, []>("input_299_cast_fp16")];
+            tensor<int32, [3]> var_1361_begin_0 = const()[name = tensor<string, []>("op_1361_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_1361_end_0 = const()[name = tensor<string, []>("op_1361_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_1361_end_mask_0 = const()[name = tensor<string, []>("op_1361_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 69, 512]> var_1361_cast_fp16 = slice_by_index(begin = var_1361_begin_0, end = var_1361_end_0, end_mask = var_1361_end_mask_0, x = cache_21_cast_fp16)[name = tensor<string, []>("op_1361_cast_fp16")];
+            tensor<int32, [3]> var_1364_begin_0 = const()[name = tensor<string, []>("op_1364_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_1364_end_0 = const()[name = tensor<string, []>("op_1364_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<bool, [3]> var_1364_end_mask_0 = const()[name = tensor<string, []>("op_1364_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 1, 512]> var_1364_cast_fp16 = slice_by_index(begin = var_1364_begin_0, end = var_1364_end_0, end_mask = var_1364_end_mask_0, x = key_11_cast_fp16)[name = tensor<string, []>("op_1364_cast_fp16")];
+            tensor<bool, []> var_1367_interleave_0 = const()[name = tensor<string, []>("op_1367_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_1367_cast_fp16 = concat(axis = var_64, interleave = var_1367_interleave_0, values = (var_1361_cast_fp16, var_1364_cast_fp16))[name = tensor<string, []>("op_1367_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_5_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70102720)))];
+            tensor<fp16, [1, 3, 512]> linear_48_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16, x = key_11_cast_fp16)[name = tensor<string, []>("linear_48_cast_fp16")];
+            tensor<int32, [4]> var_1371 = const()[name = tensor<string, []>("op_1371"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 3, 8, 64]> q_31_cast_fp16 = reshape(shape = var_1371, x = linear_48_cast_fp16)[name = tensor<string, []>("q_31_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_5_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70627072)))];
+            tensor<fp16, [1, 73, 512]> linear_49_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16, x = input_299_cast_fp16)[name = tensor<string, []>("linear_49_cast_fp16")];
+            tensor<int32, [4]> var_1375 = const()[name = tensor<string, []>("op_1375"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> k_21_cast_fp16 = reshape(shape = var_1375, x = linear_49_cast_fp16)[name = tensor<string, []>("k_21_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_5_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71151424)))];
+            tensor<fp16, [1, 73, 512]> linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16, x = input_299_cast_fp16)[name = tensor<string, []>("linear_50_cast_fp16")];
+            tensor<int32, [4]> var_1379 = const()[name = tensor<string, []>("op_1379"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> v_11_cast_fp16 = reshape(shape = var_1379, x = linear_50_cast_fp16)[name = tensor<string, []>("v_11_cast_fp16")];
+            tensor<int32, [4]> value_13_perm_0 = const()[name = tensor<string, []>("value_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71675776)))];
+            tensor<fp16, [1, 3, 8, 64]> var_1391_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1391_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71676864)))];
+            tensor<fp16, [1, 3, 8, 64]> var_1393_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1393_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_11_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_137_transpose_x_0 = const()[name = tensor<string, []>("x_137_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_137_transpose_y_0 = const()[name = tensor<string, []>("x_137_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 145]> var_1395_to_fp16 = const()[name = tensor<string, []>("op_1395_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71677952)))];
+            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1393_cast_fp16)[name = tensor<string, []>("transpose_192")];
+            tensor<fp16, [1, 8, 3, 145]> x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = var_1395_to_fp16)[name = tensor<string, []>("x_137_cast_fp16")];
+            tensor<int32, [8]> x_139_pad_0 = const()[name = tensor<string, []>("x_139_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_139_mode_0 = const()[name = tensor<string, []>("x_139_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_88_to_fp16 = const()[name = tensor<string, []>("const_88_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 3, 146]> x_139_cast_fp16 = pad(constant_val = const_88_to_fp16, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137_cast_fp16)[name = tensor<string, []>("x_139_cast_fp16")];
+            tensor<int32, [4]> var_1403 = const()[name = tensor<string, []>("op_1403"), val = tensor<int32, [4]>([1, 8, -1, 3])];
+            tensor<fp16, [1, 8, 146, 3]> x_141_cast_fp16 = reshape(shape = var_1403, x = x_139_cast_fp16)[name = tensor<string, []>("x_141_cast_fp16")];
+            tensor<int32, [4]> var_1407_begin_0 = const()[name = tensor<string, []>("op_1407_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1407_end_0 = const()[name = tensor<string, []>("op_1407_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
+            tensor<bool, [4]> var_1407_end_mask_0 = const()[name = tensor<string, []>("op_1407_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 145, 3]> var_1407_cast_fp16 = slice_by_index(begin = var_1407_begin_0, end = var_1407_end_0, end_mask = var_1407_end_mask_0, x = x_141_cast_fp16)[name = tensor<string, []>("op_1407_cast_fp16")];
+            tensor<int32, [4]> var_1408 = const()[name = tensor<string, []>("op_1408"), val = tensor<int32, [4]>([1, 8, 3, 145])];
+            tensor<fp16, [1, 8, 3, 145]> matrix_bd_21_cast_fp16 = reshape(shape = var_1408, x = var_1407_cast_fp16)[name = tensor<string, []>("matrix_bd_21_cast_fp16")];
+            tensor<bool, []> matrix_ac_11_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_11_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_11_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_11_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_61_perm_0 = const()[name = tensor<string, []>("transpose_61_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_62_perm_0 = const()[name = tensor<string, []>("transpose_62_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 73]> transpose_62 = transpose(perm = transpose_62_perm_0, x = k_21_cast_fp16)[name = tensor<string, []>("transpose_190")];
+            tensor<fp16, [1, 8, 3, 64]> transpose_61 = transpose(perm = transpose_61_perm_0, x = var_1391_cast_fp16)[name = tensor<string, []>("transpose_191")];
+            tensor<fp16, [1, 8, 3, 73]> matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_61, y = transpose_62)[name = tensor<string, []>("matrix_ac_11_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_23_begin_0 = const()[name = tensor<string, []>("matrix_bd_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_23_end_0 = const()[name = tensor<string, []>("matrix_bd_23_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
+            tensor<bool, [4]> matrix_bd_23_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 3, 73]> matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21_cast_fp16)[name = tensor<string, []>("matrix_bd_23_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_1417_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = tensor<string, []>("op_1417_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_21_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_21_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_21_cast_fp16 = mul(x = var_1417_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_21_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> scores_23_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_23_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_1423_cast_fp16 = softmax(axis = var_62, x = scores_23_cast_fp16)[name = tensor<string, []>("op_1423_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> input_301_cast_fp16 = select(a = var_40_to_fp16, b = var_1423_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_301_cast_fp16")];
+            tensor<bool, []> x_143_transpose_x_0 = const()[name = tensor<string, []>("x_143_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_143_transpose_y_0 = const()[name = tensor<string, []>("x_143_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 73, 64]> value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_11_cast_fp16)[name = tensor<string, []>("transpose_193")];
+            tensor<fp16, [1, 8, 3, 64]> x_143_cast_fp16 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301_cast_fp16, y = value_13_cast_fp16)[name = tensor<string, []>("x_143_cast_fp16")];
+            tensor<int32, [4]> var_1427_perm_0 = const()[name = tensor<string, []>("op_1427_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1428 = const()[name = tensor<string, []>("op_1428"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 3, 8, 64]> var_1427_cast_fp16 = transpose(perm = var_1427_perm_0, x = x_143_cast_fp16)[name = tensor<string, []>("transpose_189")];
+            tensor<fp16, [1, 3, 512]> input_303_cast_fp16 = reshape(shape = var_1428, x = var_1427_cast_fp16)[name = tensor<string, []>("input_303_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_5_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71826496)))];
+            tensor<fp16, [1, 3, 512]> linear_52_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16, x = input_303_cast_fp16)[name = tensor<string, []>("linear_52_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = tensor<string, []>("input_307_cast_fp16")];
+            tensor<int32, [1]> x_147_axes_0 = const()[name = tensor<string, []>("x_147_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72350848)))];
+            tensor<fp16, [512]> encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72351936)))];
+            tensor<fp16, [1, 3, 512]> x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = tensor<string, []>("x_147_cast_fp16")];
+            tensor<int32, [3]> input_309_perm_0 = const()[name = tensor<string, []>("input_309_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_311_pad_type_0 = const()[name = tensor<string, []>("input_311_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_311_strides_0 = const()[name = tensor<string, []>("input_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_311_pad_0 = const()[name = tensor<string, []>("input_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_311_dilations_0 = const()[name = tensor<string, []>("input_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_311_groups_0 = const()[name = tensor<string, []>("input_311_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_5_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72353024)))];
+            tensor<fp16, [1, 512, 3]> input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = tensor<string, []>("transpose_188")];
+            tensor<fp16, [1, 1024, 3]> input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16, x = input_309_cast_fp16)[name = tensor<string, []>("input_311_cast_fp16")];
+            tensor<int32, []> x_149_split_num_splits_0 = const()[name = tensor<string, []>("x_149_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_149_split_axis_0 = const()[name = tensor<string, []>("x_149_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 3]> x_149_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_149_split_cast_fp16_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311_cast_fp16)[name = tensor<string, []>("x_149_split_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = tensor<string, []>("x_149_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_149_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> input_313_cast_fp16 = select(a = var_40_to_fp16, b = x_149_cast_fp16, cond = var_418)[name = tensor<string, []>("input_313_cast_fp16")];
+            tensor<bool, []> new_x_23_interleave_0 = const()[name = tensor<string, []>("new_x_23_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 11]> new_x_23_cast_fp16 = concat(axis = var_62, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = tensor<string, []>("new_x_23_cast_fp16")];
+            tensor<int32, [3]> next_cache_11_begin_0 = const()[name = tensor<string, []>("next_cache_11_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_11_end_0 = const()[name = tensor<string, []>("next_cache_11_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> next_cache_11_end_mask_0 = const()[name = tensor<string, []>("next_cache_11_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 9]> next_cache_11_cast_fp16 = slice_by_index(begin = next_cache_11_begin_0, end = next_cache_11_end_0, end_mask = next_cache_11_end_mask_0, x = new_x_23_cast_fp16)[name = tensor<string, []>("next_cache_11_cast_fp16")];
+            tensor<int32, [3]> var_1469_begin_0 = const()[name = tensor<string, []>("op_1469_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> var_1469_end_0 = const()[name = tensor<string, []>("op_1469_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> var_1469_end_mask_0 = const()[name = tensor<string, []>("op_1469_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_1469_cast_fp16 = slice_by_index(begin = var_1469_begin_0, end = var_1469_end_0, end_mask = var_1469_end_mask_0, x = next_cache_11_cast_fp16)[name = tensor<string, []>("op_1469_cast_fp16")];
+            tensor<string, []> x_151_pad_type_0 = const()[name = tensor<string, []>("x_151_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_151_groups_0 = const()[name = tensor<string, []>("x_151_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_151_strides_0 = const()[name = tensor<string, []>("x_151_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_151_pad_0 = const()[name = tensor<string, []>("x_151_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_151_dilations_0 = const()[name = tensor<string, []>("x_151_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_5_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73401664)))];
+            tensor<fp16, [1, 512, 3]> x_151_cast_fp16 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16, x = new_x_23_cast_fp16)[name = tensor<string, []>("x_151_cast_fp16")];
+            tensor<int32, [3]> input_315_perm_0 = const()[name = tensor<string, []>("input_315_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_153_axes_0 = const()[name = tensor<string, []>("x_153_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73410944)))];
+            tensor<fp16, [512]> encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73412032)))];
+            tensor<fp16, [1, 3, 512]> input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = tensor<string, []>("transpose_187")];
+            tensor<fp16, [1, 3, 512]> x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = tensor<string, []>("x_153_cast_fp16")];
+            tensor<int32, [3]> input_317_perm_0 = const()[name = tensor<string, []>("input_317_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 3]> input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = tensor<string, []>("transpose_186")];
+            tensor<fp16, [1, 512, 3]> input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = tensor<string, []>("input_319_cast_fp16")];
+            tensor<string, []> x_155_pad_type_0 = const()[name = tensor<string, []>("x_155_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_155_strides_0 = const()[name = tensor<string, []>("x_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_155_pad_0 = const()[name = tensor<string, []>("x_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_155_dilations_0 = const()[name = tensor<string, []>("x_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_155_groups_0 = const()[name = tensor<string, []>("x_155_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_5_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73413120)))];
+            tensor<fp16, [1, 512, 3]> x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16, x = input_319_cast_fp16)[name = tensor<string, []>("x_155_cast_fp16")];
+            tensor<int32, [3]> input_321_perm_0 = const()[name = tensor<string, []>("input_321_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 3, 512]> input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = tensor<string, []>("transpose_185")];
+            tensor<fp16, [1, 3, 512]> input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = tensor<string, []>("input_323_cast_fp16")];
+            tensor<int32, [1]> input_325_axes_0 = const()[name = tensor<string, []>("input_325_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73937472)))];
+            tensor<fp16, [512]> encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73938560)))];
+            tensor<fp16, [1, 3, 512]> input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = tensor<string, []>("input_325_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_5_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73939648)))];
+            tensor<fp16, [1, 3, 2048]> linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16, x = input_325_cast_fp16)[name = tensor<string, []>("linear_53_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = tensor<string, []>("input_329_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_5_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76036864)))];
+            tensor<fp16, [1, 3, 512]> linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16, x = input_329_cast_fp16)[name = tensor<string, []>("linear_54_cast_fp16")];
+            tensor<fp16, []> var_1510_to_fp16 = const()[name = tensor<string, []>("op_1510_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_1511_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1510_to_fp16)[name = tensor<string, []>("op_1511_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1511_cast_fp16)[name = tensor<string, []>("input_335_cast_fp16")];
+            tensor<int32, [1]> input_337_axes_0 = const()[name = tensor<string, []>("input_337_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_5_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78134080)))];
+            tensor<fp16, [512]> encoder_layers_5_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78135168)))];
+            tensor<fp16, [1, 3, 512]> input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = tensor<string, []>("input_337_cast_fp16")];
+            tensor<int32, [4]> cache_25_begin_0 = const()[name = tensor<string, []>("cache_25_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_25_end_0 = const()[name = tensor<string, []>("cache_25_end_0"), val = tensor<int32, [4]>([7, 1, 70, 512])];
+            tensor<bool, [4]> cache_25_end_mask_0 = const()[name = tensor<string, []>("cache_25_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_25_squeeze_mask_0 = const()[name = tensor<string, []>("cache_25_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_25_cast_fp16 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_25_cast_fp16")];
+            tensor<int32, [4]> cache_27_begin_0 = const()[name = tensor<string, []>("cache_27_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_27_end_0 = const()[name = tensor<string, []>("cache_27_end_0"), val = tensor<int32, [4]>([7, 1, 512, 8])];
+            tensor<bool, [4]> cache_27_end_mask_0 = const()[name = tensor<string, []>("cache_27_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_27_squeeze_mask_0 = const()[name = tensor<string, []>("cache_27_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_27_cast_fp16 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_27_cast_fp16")];
+            tensor<int32, [1]> input_339_axes_0 = const()[name = tensor<string, []>("input_339_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78136256)))];
+            tensor<fp16, [512]> encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78137344)))];
+            tensor<fp16, [1, 3, 512]> input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = tensor<string, []>("input_339_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_6_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78138432)))];
+            tensor<fp16, [1, 3, 2048]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16, x = input_339_cast_fp16)[name = tensor<string, []>("linear_55_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = tensor<string, []>("input_343_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_6_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80235648)))];
+            tensor<fp16, [1, 3, 512]> linear_56_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16, x = input_343_cast_fp16)[name = tensor<string, []>("linear_56_cast_fp16")];
+            tensor<fp16, []> var_1545_to_fp16 = const()[name = tensor<string, []>("op_1545_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_1546_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1545_to_fp16)[name = tensor<string, []>("op_1546_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1546_cast_fp16)[name = tensor<string, []>("input_349_cast_fp16")];
+            tensor<int32, [1]> key_13_axes_0 = const()[name = tensor<string, []>("key_13_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82332864)))];
+            tensor<fp16, [512]> encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82333952)))];
+            tensor<fp16, [1, 3, 512]> key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")];
+            tensor<bool, []> input_351_interleave_0 = const()[name = tensor<string, []>("input_351_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 73, 512]> input_351_cast_fp16 = concat(axis = var_64, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = tensor<string, []>("input_351_cast_fp16")];
+            tensor<int32, [3]> var_1568_begin_0 = const()[name = tensor<string, []>("op_1568_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_1568_end_0 = const()[name = tensor<string, []>("op_1568_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_1568_end_mask_0 = const()[name = tensor<string, []>("op_1568_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 69, 512]> var_1568_cast_fp16 = slice_by_index(begin = var_1568_begin_0, end = var_1568_end_0, end_mask = var_1568_end_mask_0, x = cache_25_cast_fp16)[name = tensor<string, []>("op_1568_cast_fp16")];
+            tensor<int32, [3]> var_1571_begin_0 = const()[name = tensor<string, []>("op_1571_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_1571_end_0 = const()[name = tensor<string, []>("op_1571_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<bool, [3]> var_1571_end_mask_0 = const()[name = tensor<string, []>("op_1571_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 1, 512]> var_1571_cast_fp16 = slice_by_index(begin = var_1571_begin_0, end = var_1571_end_0, end_mask = var_1571_end_mask_0, x = key_13_cast_fp16)[name = tensor<string, []>("op_1571_cast_fp16")];
+            tensor<bool, []> var_1574_interleave_0 = const()[name = tensor<string, []>("op_1574_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_1574_cast_fp16 = concat(axis = var_64, interleave = var_1574_interleave_0, values = (var_1568_cast_fp16, var_1571_cast_fp16))[name = tensor<string, []>("op_1574_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_6_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82335040)))];
+            tensor<fp16, [1, 3, 512]> linear_57_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16, x = key_13_cast_fp16)[name = tensor<string, []>("linear_57_cast_fp16")];
+            tensor<int32, [4]> var_1578 = const()[name = tensor<string, []>("op_1578"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 3, 8, 64]> q_37_cast_fp16 = reshape(shape = var_1578, x = linear_57_cast_fp16)[name = tensor<string, []>("q_37_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_6_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82859392)))];
+            tensor<fp16, [1, 73, 512]> linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16, x = input_351_cast_fp16)[name = tensor<string, []>("linear_58_cast_fp16")];
+            tensor<int32, [4]> var_1582 = const()[name = tensor<string, []>("op_1582"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> k_25_cast_fp16 = reshape(shape = var_1582, x = linear_58_cast_fp16)[name = tensor<string, []>("k_25_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_6_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83383744)))];
+            tensor<fp16, [1, 73, 512]> linear_59_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16, x = input_351_cast_fp16)[name = tensor<string, []>("linear_59_cast_fp16")];
+            tensor<int32, [4]> var_1586 = const()[name = tensor<string, []>("op_1586"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> v_13_cast_fp16 = reshape(shape = var_1586, x = linear_59_cast_fp16)[name = tensor<string, []>("v_13_cast_fp16")];
+            tensor<int32, [4]> value_15_perm_0 = const()[name = tensor<string, []>("value_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83908096)))];
+            tensor<fp16, [1, 3, 8, 64]> var_1598_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1598_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83909184)))];
+            tensor<fp16, [1, 3, 8, 64]> var_1600_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1600_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_13_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_163_transpose_x_0 = const()[name = tensor<string, []>("x_163_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_163_transpose_y_0 = const()[name = tensor<string, []>("x_163_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 145]> var_1602_to_fp16 = const()[name = tensor<string, []>("op_1602_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83910272)))];
+            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1600_cast_fp16)[name = tensor<string, []>("transpose_183")];
+            tensor<fp16, [1, 8, 3, 145]> x_163_cast_fp16 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = var_1602_to_fp16)[name = tensor<string, []>("x_163_cast_fp16")];
+            tensor<int32, [8]> x_165_pad_0 = const()[name = tensor<string, []>("x_165_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_165_mode_0 = const()[name = tensor<string, []>("x_165_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_101_to_fp16 = const()[name = tensor<string, []>("const_101_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 3, 146]> x_165_cast_fp16 = pad(constant_val = const_101_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = tensor<string, []>("x_165_cast_fp16")];
+            tensor<int32, [4]> var_1610 = const()[name = tensor<string, []>("op_1610"), val = tensor<int32, [4]>([1, 8, -1, 3])];
+            tensor<fp16, [1, 8, 146, 3]> x_167_cast_fp16 = reshape(shape = var_1610, x = x_165_cast_fp16)[name = tensor<string, []>("x_167_cast_fp16")];
+            tensor<int32, [4]> var_1614_begin_0 = const()[name = tensor<string, []>("op_1614_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1614_end_0 = const()[name = tensor<string, []>("op_1614_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
+            tensor<bool, [4]> var_1614_end_mask_0 = const()[name = tensor<string, []>("op_1614_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 145, 3]> var_1614_cast_fp16 = slice_by_index(begin = var_1614_begin_0, end = var_1614_end_0, end_mask = var_1614_end_mask_0, x = x_167_cast_fp16)[name = tensor<string, []>("op_1614_cast_fp16")];
+            tensor<int32, [4]> var_1615 = const()[name = tensor<string, []>("op_1615"), val = tensor<int32, [4]>([1, 8, 3, 145])];
+            tensor<fp16, [1, 8, 3, 145]> matrix_bd_25_cast_fp16 = reshape(shape = var_1615, x = var_1614_cast_fp16)[name = tensor<string, []>("matrix_bd_25_cast_fp16")];
+            tensor<bool, []> matrix_ac_13_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_13_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_63_perm_0 = const()[name = tensor<string, []>("transpose_63_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_64_perm_0 = const()[name = tensor<string, []>("transpose_64_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 73]> transpose_64 = transpose(perm = transpose_64_perm_0, x = k_25_cast_fp16)[name = tensor<string, []>("transpose_181")];
+            tensor<fp16, [1, 8, 3, 64]> transpose_63 = transpose(perm = transpose_63_perm_0, x = var_1598_cast_fp16)[name = tensor<string, []>("transpose_182")];
+            tensor<fp16, [1, 8, 3, 73]> matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_63, y = transpose_64)[name = tensor<string, []>("matrix_ac_13_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_27_begin_0 = const()[name = tensor<string, []>("matrix_bd_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_27_end_0 = const()[name = tensor<string, []>("matrix_bd_27_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
+            tensor<bool, [4]> matrix_bd_27_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 3, 73]> matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25_cast_fp16)[name = tensor<string, []>("matrix_bd_27_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_1624_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = tensor<string, []>("op_1624_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_25_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_25_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_25_cast_fp16 = mul(x = var_1624_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_25_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> scores_27_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_27_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_1630_cast_fp16 = softmax(axis = var_62, x = scores_27_cast_fp16)[name = tensor<string, []>("op_1630_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> input_353_cast_fp16 = select(a = var_40_to_fp16, b = var_1630_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_353_cast_fp16")];
+            tensor<bool, []> x_169_transpose_x_0 = const()[name = tensor<string, []>("x_169_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_169_transpose_y_0 = const()[name = tensor<string, []>("x_169_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 73, 64]> value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_13_cast_fp16)[name = tensor<string, []>("transpose_184")];
+            tensor<fp16, [1, 8, 3, 64]> x_169_cast_fp16 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353_cast_fp16, y = value_15_cast_fp16)[name = tensor<string, []>("x_169_cast_fp16")];
+            tensor<int32, [4]> var_1634_perm_0 = const()[name = tensor<string, []>("op_1634_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1635 = const()[name = tensor<string, []>("op_1635"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 3, 8, 64]> var_1634_cast_fp16 = transpose(perm = var_1634_perm_0, x = x_169_cast_fp16)[name = tensor<string, []>("transpose_180")];
+            tensor<fp16, [1, 3, 512]> input_355_cast_fp16 = reshape(shape = var_1635, x = var_1634_cast_fp16)[name = tensor<string, []>("input_355_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_6_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84058816)))];
+            tensor<fp16, [1, 3, 512]> linear_61_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16, x = input_355_cast_fp16)[name = tensor<string, []>("linear_61_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = tensor<string, []>("input_359_cast_fp16")];
+            tensor<int32, [1]> x_173_axes_0 = const()[name = tensor<string, []>("x_173_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84583168)))];
+            tensor<fp16, [512]> encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84584256)))];
+            tensor<fp16, [1, 3, 512]> x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = tensor<string, []>("x_173_cast_fp16")];
+            tensor<int32, [3]> input_361_perm_0 = const()[name = tensor<string, []>("input_361_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_363_pad_type_0 = const()[name = tensor<string, []>("input_363_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_363_strides_0 = const()[name = tensor<string, []>("input_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_363_pad_0 = const()[name = tensor<string, []>("input_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_363_dilations_0 = const()[name = tensor<string, []>("input_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_363_groups_0 = const()[name = tensor<string, []>("input_363_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_6_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84585344)))];
+            tensor<fp16, [1, 512, 3]> input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = tensor<string, []>("transpose_179")];
+            tensor<fp16, [1, 1024, 3]> input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16, x = input_361_cast_fp16)[name = tensor<string, []>("input_363_cast_fp16")];
+            tensor<int32, []> x_175_split_num_splits_0 = const()[name = tensor<string, []>("x_175_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_175_split_axis_0 = const()[name = tensor<string, []>("x_175_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 3]> x_175_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_175_split_cast_fp16_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363_cast_fp16)[name = tensor<string, []>("x_175_split_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = tensor<string, []>("x_175_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_175_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> input_365_cast_fp16 = select(a = var_40_to_fp16, b = x_175_cast_fp16, cond = var_418)[name = tensor<string, []>("input_365_cast_fp16")];
+            tensor<bool, []> new_x_27_interleave_0 = const()[name = tensor<string, []>("new_x_27_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 11]> new_x_27_cast_fp16 = concat(axis = var_62, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = tensor<string, []>("new_x_27_cast_fp16")];
+            tensor<int32, [3]> next_cache_13_begin_0 = const()[name = tensor<string, []>("next_cache_13_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_13_end_0 = const()[name = tensor<string, []>("next_cache_13_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> next_cache_13_end_mask_0 = const()[name = tensor<string, []>("next_cache_13_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 9]> next_cache_13_cast_fp16 = slice_by_index(begin = next_cache_13_begin_0, end = next_cache_13_end_0, end_mask = next_cache_13_end_mask_0, x = new_x_27_cast_fp16)[name = tensor<string, []>("next_cache_13_cast_fp16")];
+            tensor<int32, [3]> var_1676_begin_0 = const()[name = tensor<string, []>("op_1676_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> var_1676_end_0 = const()[name = tensor<string, []>("op_1676_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> var_1676_end_mask_0 = const()[name = tensor<string, []>("op_1676_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_1676_cast_fp16 = slice_by_index(begin = var_1676_begin_0, end = var_1676_end_0, end_mask = var_1676_end_mask_0, x = next_cache_13_cast_fp16)[name = tensor<string, []>("op_1676_cast_fp16")];
+            tensor<string, []> x_177_pad_type_0 = const()[name = tensor<string, []>("x_177_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_177_groups_0 = const()[name = tensor<string, []>("x_177_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_177_strides_0 = const()[name = tensor<string, []>("x_177_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_177_pad_0 = const()[name = tensor<string, []>("x_177_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_177_dilations_0 = const()[name = tensor<string, []>("x_177_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_6_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85633984)))];
+            tensor<fp16, [1, 512, 3]> x_177_cast_fp16 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16, x = new_x_27_cast_fp16)[name = tensor<string, []>("x_177_cast_fp16")];
+            tensor<int32, [3]> input_367_perm_0 = const()[name = tensor<string, []>("input_367_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_179_axes_0 = const()[name = tensor<string, []>("x_179_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85643264)))];
+            tensor<fp16, [512]> encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85644352)))];
+            tensor<fp16, [1, 3, 512]> input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = tensor<string, []>("transpose_178")];
+            tensor<fp16, [1, 3, 512]> x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = tensor<string, []>("x_179_cast_fp16")];
+            tensor<int32, [3]> input_369_perm_0 = const()[name = tensor<string, []>("input_369_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 3]> input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = tensor<string, []>("transpose_177")];
+            tensor<fp16, [1, 512, 3]> input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = tensor<string, []>("input_371_cast_fp16")];
+            tensor<string, []> x_181_pad_type_0 = const()[name = tensor<string, []>("x_181_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_181_strides_0 = const()[name = tensor<string, []>("x_181_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_181_pad_0 = const()[name = tensor<string, []>("x_181_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_181_dilations_0 = const()[name = tensor<string, []>("x_181_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_181_groups_0 = const()[name = tensor<string, []>("x_181_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_6_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85645440)))];
+            tensor<fp16, [1, 512, 3]> x_181_cast_fp16 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16, x = input_371_cast_fp16)[name = tensor<string, []>("x_181_cast_fp16")];
+            tensor<int32, [3]> input_373_perm_0 = const()[name = tensor<string, []>("input_373_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 3, 512]> input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = tensor<string, []>("transpose_176")];
+            tensor<fp16, [1, 3, 512]> input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = tensor<string, []>("input_375_cast_fp16")];
+            tensor<int32, [1]> input_377_axes_0 = const()[name = tensor<string, []>("input_377_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86169792)))];
+            tensor<fp16, [512]> encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86170880)))];
+            tensor<fp16, [1, 3, 512]> input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = tensor<string, []>("input_377_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_6_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86171968)))];
+            tensor<fp16, [1, 3, 2048]> linear_62_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16, x = input_377_cast_fp16)[name = tensor<string, []>("linear_62_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = tensor<string, []>("input_381_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_6_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88269184)))];
+            tensor<fp16, [1, 3, 512]> linear_63_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16, x = input_381_cast_fp16)[name = tensor<string, []>("linear_63_cast_fp16")];
+            tensor<fp16, []> var_1717_to_fp16 = const()[name = tensor<string, []>("op_1717_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_1718_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1717_to_fp16)[name = tensor<string, []>("op_1718_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1718_cast_fp16)[name = tensor<string, []>("input_387_cast_fp16")];
+            tensor<int32, [1]> input_389_axes_0 = const()[name = tensor<string, []>("input_389_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_6_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90366400)))];
+            tensor<fp16, [512]> encoder_layers_6_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90367488)))];
+            tensor<fp16, [1, 3, 512]> input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = tensor<string, []>("input_389_cast_fp16")];
+            tensor<int32, [4]> cache_29_begin_0 = const()[name = tensor<string, []>("cache_29_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_29_end_0 = const()[name = tensor<string, []>("cache_29_end_0"), val = tensor<int32, [4]>([8, 1, 70, 512])];
+            tensor<bool, [4]> cache_29_end_mask_0 = const()[name = tensor<string, []>("cache_29_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_29_squeeze_mask_0 = const()[name = tensor<string, []>("cache_29_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_29_cast_fp16 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_29_cast_fp16")];
+            tensor<int32, [4]> cache_31_begin_0 = const()[name = tensor<string, []>("cache_31_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_31_end_0 = const()[name = tensor<string, []>("cache_31_end_0"), val = tensor<int32, [4]>([8, 1, 512, 8])];
+            tensor<bool, [4]> cache_31_end_mask_0 = const()[name = tensor<string, []>("cache_31_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_31_squeeze_mask_0 = const()[name = tensor<string, []>("cache_31_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_31_cast_fp16 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_31_cast_fp16")];
+            tensor<int32, [1]> input_391_axes_0 = const()[name = tensor<string, []>("input_391_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90368576)))];
+            tensor<fp16, [512]> encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90369664)))];
+            tensor<fp16, [1, 3, 512]> input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = tensor<string, []>("input_391_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_7_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90370752)))];
+            tensor<fp16, [1, 3, 2048]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16, x = input_391_cast_fp16)[name = tensor<string, []>("linear_64_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = tensor<string, []>("input_395_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_7_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92467968)))];
+            tensor<fp16, [1, 3, 512]> linear_65_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16, x = input_395_cast_fp16)[name = tensor<string, []>("linear_65_cast_fp16")];
+            tensor<fp16, []> var_1752_to_fp16 = const()[name = tensor<string, []>("op_1752_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_1753_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1752_to_fp16)[name = tensor<string, []>("op_1753_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1753_cast_fp16)[name = tensor<string, []>("input_401_cast_fp16")];
+            tensor<int32, [1]> key_15_axes_0 = const()[name = tensor<string, []>("key_15_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94565184)))];
+            tensor<fp16, [512]> encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94566272)))];
+            tensor<fp16, [1, 3, 512]> key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = tensor<string, []>("key_15_cast_fp16")];
+            tensor<bool, []> input_403_interleave_0 = const()[name = tensor<string, []>("input_403_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 73, 512]> input_403_cast_fp16 = concat(axis = var_64, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = tensor<string, []>("input_403_cast_fp16")];
+            tensor<int32, [3]> var_1775_begin_0 = const()[name = tensor<string, []>("op_1775_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_1775_end_0 = const()[name = tensor<string, []>("op_1775_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_1775_end_mask_0 = const()[name = tensor<string, []>("op_1775_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 69, 512]> var_1775_cast_fp16 = slice_by_index(begin = var_1775_begin_0, end = var_1775_end_0, end_mask = var_1775_end_mask_0, x = cache_29_cast_fp16)[name = tensor<string, []>("op_1775_cast_fp16")];
+            tensor<int32, [3]> var_1778_begin_0 = const()[name = tensor<string, []>("op_1778_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_1778_end_0 = const()[name = tensor<string, []>("op_1778_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<bool, [3]> var_1778_end_mask_0 = const()[name = tensor<string, []>("op_1778_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 1, 512]> var_1778_cast_fp16 = slice_by_index(begin = var_1778_begin_0, end = var_1778_end_0, end_mask = var_1778_end_mask_0, x = key_15_cast_fp16)[name = tensor<string, []>("op_1778_cast_fp16")];
+            tensor<bool, []> var_1781_interleave_0 = const()[name = tensor<string, []>("op_1781_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_1781_cast_fp16 = concat(axis = var_64, interleave = var_1781_interleave_0, values = (var_1775_cast_fp16, var_1778_cast_fp16))[name = tensor<string, []>("op_1781_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_7_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94567360)))];
+            tensor<fp16, [1, 3, 512]> linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16, x = key_15_cast_fp16)[name = tensor<string, []>("linear_66_cast_fp16")];
+            tensor<int32, [4]> var_1785 = const()[name = tensor<string, []>("op_1785"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 3, 8, 64]> q_43_cast_fp16 = reshape(shape = var_1785, x = linear_66_cast_fp16)[name = tensor<string, []>("q_43_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_7_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(95091712)))];
+            tensor<fp16, [1, 73, 512]> linear_67_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16, x = input_403_cast_fp16)[name = tensor<string, []>("linear_67_cast_fp16")];
+            tensor<int32, [4]> var_1789 = const()[name = tensor<string, []>("op_1789"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> k_29_cast_fp16 = reshape(shape = var_1789, x = linear_67_cast_fp16)[name = tensor<string, []>("k_29_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_7_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(95616064)))];
+            tensor<fp16, [1, 73, 512]> linear_68_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16, x = input_403_cast_fp16)[name = tensor<string, []>("linear_68_cast_fp16")];
+            tensor<int32, [4]> var_1793 = const()[name = tensor<string, []>("op_1793"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> v_15_cast_fp16 = reshape(shape = var_1793, x = linear_68_cast_fp16)[name = tensor<string, []>("v_15_cast_fp16")];
+            tensor<int32, [4]> value_17_perm_0 = const()[name = tensor<string, []>("value_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96140416)))];
+            tensor<fp16, [1, 3, 8, 64]> var_1805_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1805_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96141504)))];
+            tensor<fp16, [1, 3, 8, 64]> var_1807_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1807_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_15_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_189_transpose_x_0 = const()[name = tensor<string, []>("x_189_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_189_transpose_y_0 = const()[name = tensor<string, []>("x_189_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 145]> var_1809_to_fp16 = const()[name = tensor<string, []>("op_1809_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96142592)))];
+            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_1807_cast_fp16)[name = tensor<string, []>("transpose_174")];
+            tensor<fp16, [1, 8, 3, 145]> x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = var_1809_to_fp16)[name = tensor<string, []>("x_189_cast_fp16")];
+            tensor<int32, [8]> x_191_pad_0 = const()[name = tensor<string, []>("x_191_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_191_mode_0 = const()[name = tensor<string, []>("x_191_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_114_to_fp16 = const()[name = tensor<string, []>("const_114_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 3, 146]> x_191_cast_fp16 = pad(constant_val = const_114_to_fp16, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189_cast_fp16)[name = tensor<string, []>("x_191_cast_fp16")];
+            tensor<int32, [4]> var_1817 = const()[name = tensor<string, []>("op_1817"), val = tensor<int32, [4]>([1, 8, -1, 3])];
+            tensor<fp16, [1, 8, 146, 3]> x_193_cast_fp16 = reshape(shape = var_1817, x = x_191_cast_fp16)[name = tensor<string, []>("x_193_cast_fp16")];
+            tensor<int32, [4]> var_1821_begin_0 = const()[name = tensor<string, []>("op_1821_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1821_end_0 = const()[name = tensor<string, []>("op_1821_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
+            tensor<bool, [4]> var_1821_end_mask_0 = const()[name = tensor<string, []>("op_1821_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 145, 3]> var_1821_cast_fp16 = slice_by_index(begin = var_1821_begin_0, end = var_1821_end_0, end_mask = var_1821_end_mask_0, x = x_193_cast_fp16)[name = tensor<string, []>("op_1821_cast_fp16")];
+            tensor<int32, [4]> var_1822 = const()[name = tensor<string, []>("op_1822"), val = tensor<int32, [4]>([1, 8, 3, 145])];
+            tensor<fp16, [1, 8, 3, 145]> matrix_bd_29_cast_fp16 = reshape(shape = var_1822, x = var_1821_cast_fp16)[name = tensor<string, []>("matrix_bd_29_cast_fp16")];
+            tensor<bool, []> matrix_ac_15_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_15_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_15_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_15_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_65_perm_0 = const()[name = tensor<string, []>("transpose_65_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_66_perm_0 = const()[name = tensor<string, []>("transpose_66_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 73]> transpose_66 = transpose(perm = transpose_66_perm_0, x = k_29_cast_fp16)[name = tensor<string, []>("transpose_172")];
+            tensor<fp16, [1, 8, 3, 64]> transpose_65 = transpose(perm = transpose_65_perm_0, x = var_1805_cast_fp16)[name = tensor<string, []>("transpose_173")];
+            tensor<fp16, [1, 8, 3, 73]> matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_65, y = transpose_66)[name = tensor<string, []>("matrix_ac_15_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_31_begin_0 = const()[name = tensor<string, []>("matrix_bd_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_31_end_0 = const()[name = tensor<string, []>("matrix_bd_31_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
+            tensor<bool, [4]> matrix_bd_31_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 3, 73]> matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29_cast_fp16)[name = tensor<string, []>("matrix_bd_31_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_1831_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = tensor<string, []>("op_1831_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_29_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_29_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_29_cast_fp16 = mul(x = var_1831_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_29_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> scores_31_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_31_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_1837_cast_fp16 = softmax(axis = var_62, x = scores_31_cast_fp16)[name = tensor<string, []>("op_1837_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> input_405_cast_fp16 = select(a = var_40_to_fp16, b = var_1837_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_405_cast_fp16")];
+            tensor<bool, []> x_195_transpose_x_0 = const()[name = tensor<string, []>("x_195_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_195_transpose_y_0 = const()[name = tensor<string, []>("x_195_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 73, 64]> value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_15_cast_fp16)[name = tensor<string, []>("transpose_175")];
+            tensor<fp16, [1, 8, 3, 64]> x_195_cast_fp16 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405_cast_fp16, y = value_17_cast_fp16)[name = tensor<string, []>("x_195_cast_fp16")];
+            tensor<int32, [4]> var_1841_perm_0 = const()[name = tensor<string, []>("op_1841_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1842 = const()[name = tensor<string, []>("op_1842"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 3, 8, 64]> var_1841_cast_fp16 = transpose(perm = var_1841_perm_0, x = x_195_cast_fp16)[name = tensor<string, []>("transpose_171")];
+            tensor<fp16, [1, 3, 512]> input_407_cast_fp16 = reshape(shape = var_1842, x = var_1841_cast_fp16)[name = tensor<string, []>("input_407_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_7_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96291136)))];
+            tensor<fp16, [1, 3, 512]> linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16, x = input_407_cast_fp16)[name = tensor<string, []>("linear_70_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = tensor<string, []>("input_411_cast_fp16")];
+            tensor<int32, [1]> x_199_axes_0 = const()[name = tensor<string, []>("x_199_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96815488)))];
+            tensor<fp16, [512]> encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96816576)))];
+            tensor<fp16, [1, 3, 512]> x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = tensor<string, []>("x_199_cast_fp16")];
+            tensor<int32, [3]> input_413_perm_0 = const()[name = tensor<string, []>("input_413_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_415_pad_type_0 = const()[name = tensor<string, []>("input_415_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_415_strides_0 = const()[name = tensor<string, []>("input_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_415_pad_0 = const()[name = tensor<string, []>("input_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_415_dilations_0 = const()[name = tensor<string, []>("input_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_415_groups_0 = const()[name = tensor<string, []>("input_415_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_7_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96817664)))];
+            tensor<fp16, [1, 512, 3]> input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = tensor<string, []>("transpose_170")];
+            tensor<fp16, [1, 1024, 3]> input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16, x = input_413_cast_fp16)[name = tensor<string, []>("input_415_cast_fp16")];
+            tensor<int32, []> x_201_split_num_splits_0 = const()[name = tensor<string, []>("x_201_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_201_split_axis_0 = const()[name = tensor<string, []>("x_201_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 3]> x_201_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_201_split_cast_fp16_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415_cast_fp16)[name = tensor<string, []>("x_201_split_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = tensor<string, []>("x_201_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_201_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> input_417_cast_fp16 = select(a = var_40_to_fp16, b = x_201_cast_fp16, cond = var_418)[name = tensor<string, []>("input_417_cast_fp16")];
+            tensor<bool, []> new_x_31_interleave_0 = const()[name = tensor<string, []>("new_x_31_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 11]> new_x_31_cast_fp16 = concat(axis = var_62, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = tensor<string, []>("new_x_31_cast_fp16")];
+            tensor<int32, [3]> next_cache_15_begin_0 = const()[name = tensor<string, []>("next_cache_15_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_15_end_0 = const()[name = tensor<string, []>("next_cache_15_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> next_cache_15_end_mask_0 = const()[name = tensor<string, []>("next_cache_15_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 9]> next_cache_15_cast_fp16 = slice_by_index(begin = next_cache_15_begin_0, end = next_cache_15_end_0, end_mask = next_cache_15_end_mask_0, x = new_x_31_cast_fp16)[name = tensor<string, []>("next_cache_15_cast_fp16")];
+            tensor<int32, [3]> var_1883_begin_0 = const()[name = tensor<string, []>("op_1883_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> var_1883_end_0 = const()[name = tensor<string, []>("op_1883_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> var_1883_end_mask_0 = const()[name = tensor<string, []>("op_1883_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_1883_cast_fp16 = slice_by_index(begin = var_1883_begin_0, end = var_1883_end_0, end_mask = var_1883_end_mask_0, x = next_cache_15_cast_fp16)[name = tensor<string, []>("op_1883_cast_fp16")];
+            tensor<string, []> x_203_pad_type_0 = const()[name = tensor<string, []>("x_203_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_203_groups_0 = const()[name = tensor<string, []>("x_203_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_203_strides_0 = const()[name = tensor<string, []>("x_203_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_203_pad_0 = const()[name = tensor<string, []>("x_203_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_203_dilations_0 = const()[name = tensor<string, []>("x_203_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_7_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97866304)))];
+            tensor<fp16, [1, 512, 3]> x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16, x = new_x_31_cast_fp16)[name = tensor<string, []>("x_203_cast_fp16")];
+            tensor<int32, [3]> input_419_perm_0 = const()[name = tensor<string, []>("input_419_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_205_axes_0 = const()[name = tensor<string, []>("x_205_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97875584)))];
+            tensor<fp16, [512]> encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97876672)))];
+            tensor<fp16, [1, 3, 512]> input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = tensor<string, []>("transpose_169")];
+            tensor<fp16, [1, 3, 512]> x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = tensor<string, []>("x_205_cast_fp16")];
+            tensor<int32, [3]> input_421_perm_0 = const()[name = tensor<string, []>("input_421_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 3]> input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = tensor<string, []>("transpose_168")];
+            tensor<fp16, [1, 512, 3]> input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor<string, []>("input_423_cast_fp16")];
+            tensor<string, []> x_207_pad_type_0 = const()[name = tensor<string, []>("x_207_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_207_strides_0 = const()[name = tensor<string, []>("x_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_207_pad_0 = const()[name = tensor<string, []>("x_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_207_dilations_0 = const()[name = tensor<string, []>("x_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_207_groups_0 = const()[name = tensor<string, []>("x_207_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_7_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97877760)))];
+            tensor<fp16, [1, 512, 3]> x_207_cast_fp16 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16, x = input_423_cast_fp16)[name = tensor<string, []>("x_207_cast_fp16")];
+            tensor<int32, [3]> input_425_perm_0 = const()[name = tensor<string, []>("input_425_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 3, 512]> input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = tensor<string, []>("transpose_167")];
+            tensor<fp16, [1, 3, 512]> input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = tensor<string, []>("input_427_cast_fp16")];
+            tensor<int32, [1]> input_429_axes_0 = const()[name = tensor<string, []>("input_429_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98402112)))];
+            tensor<fp16, [512]> encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98403200)))];
+            tensor<fp16, [1, 3, 512]> input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = tensor<string, []>("input_429_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_7_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98404288)))];
+            tensor<fp16, [1, 3, 2048]> linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16, x = input_429_cast_fp16)[name = tensor<string, []>("linear_71_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = tensor<string, []>("input_433_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_7_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100501504)))];
+            tensor<fp16, [1, 3, 512]> linear_72_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16, x = input_433_cast_fp16)[name = tensor<string, []>("linear_72_cast_fp16")];
+            tensor<fp16, []> var_1924_to_fp16 = const()[name = tensor<string, []>("op_1924_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_1925_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_1924_to_fp16)[name = tensor<string, []>("op_1925_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_1925_cast_fp16)[name = tensor<string, []>("input_439_cast_fp16")];
+            tensor<int32, [1]> input_441_axes_0 = const()[name = tensor<string, []>("input_441_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_7_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102598720)))];
+            tensor<fp16, [512]> encoder_layers_7_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102599808)))];
+            tensor<fp16, [1, 3, 512]> input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = tensor<string, []>("input_441_cast_fp16")];
+            tensor<int32, [4]> cache_33_begin_0 = const()[name = tensor<string, []>("cache_33_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_33_end_0 = const()[name = tensor<string, []>("cache_33_end_0"), val = tensor<int32, [4]>([9, 1, 70, 512])];
+            tensor<bool, [4]> cache_33_end_mask_0 = const()[name = tensor<string, []>("cache_33_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_33_squeeze_mask_0 = const()[name = tensor<string, []>("cache_33_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_33_cast_fp16 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_33_cast_fp16")];
+            tensor<int32, [4]> cache_35_begin_0 = const()[name = tensor<string, []>("cache_35_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_35_end_0 = const()[name = tensor<string, []>("cache_35_end_0"), val = tensor<int32, [4]>([9, 1, 512, 8])];
+            tensor<bool, [4]> cache_35_end_mask_0 = const()[name = tensor<string, []>("cache_35_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_35_squeeze_mask_0 = const()[name = tensor<string, []>("cache_35_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_35_cast_fp16 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_35_cast_fp16")];
+            tensor<int32, [1]> input_443_axes_0 = const()[name = tensor<string, []>("input_443_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102600896)))];
+            tensor<fp16, [512]> encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102601984)))];
+            tensor<fp16, [1, 3, 512]> input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = tensor<string, []>("input_443_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_8_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102603072)))];
+            tensor<fp16, [1, 3, 2048]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16, x = input_443_cast_fp16)[name = tensor<string, []>("linear_73_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = tensor<string, []>("input_447_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_8_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104700288)))];
+            tensor<fp16, [1, 3, 512]> linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16, x = input_447_cast_fp16)[name = tensor<string, []>("linear_74_cast_fp16")];
+            tensor<fp16, []> var_1959_to_fp16 = const()[name = tensor<string, []>("op_1959_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_1960_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_1959_to_fp16)[name = tensor<string, []>("op_1960_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_1960_cast_fp16)[name = tensor<string, []>("input_453_cast_fp16")];
+            tensor<int32, [1]> key_17_axes_0 = const()[name = tensor<string, []>("key_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106797504)))];
+            tensor<fp16, [512]> encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106798592)))];
+            tensor<fp16, [1, 3, 512]> key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = tensor<string, []>("key_17_cast_fp16")];
+            tensor<bool, []> input_455_interleave_0 = const()[name = tensor<string, []>("input_455_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 73, 512]> input_455_cast_fp16 = concat(axis = var_64, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = tensor<string, []>("input_455_cast_fp16")];
+            tensor<int32, [3]> var_1982_begin_0 = const()[name = tensor<string, []>("op_1982_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_1982_end_0 = const()[name = tensor<string, []>("op_1982_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_1982_end_mask_0 = const()[name = tensor<string, []>("op_1982_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 69, 512]> var_1982_cast_fp16 = slice_by_index(begin = var_1982_begin_0, end = var_1982_end_0, end_mask = var_1982_end_mask_0, x = cache_33_cast_fp16)[name = tensor<string, []>("op_1982_cast_fp16")];
+            tensor<int32, [3]> var_1985_begin_0 = const()[name = tensor<string, []>("op_1985_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_1985_end_0 = const()[name = tensor<string, []>("op_1985_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<bool, [3]> var_1985_end_mask_0 = const()[name = tensor<string, []>("op_1985_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 1, 512]> var_1985_cast_fp16 = slice_by_index(begin = var_1985_begin_0, end = var_1985_end_0, end_mask = var_1985_end_mask_0, x = key_17_cast_fp16)[name = tensor<string, []>("op_1985_cast_fp16")];
+            tensor<bool, []> var_1988_interleave_0 = const()[name = tensor<string, []>("op_1988_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_1988_cast_fp16 = concat(axis = var_64, interleave = var_1988_interleave_0, values = (var_1982_cast_fp16, var_1985_cast_fp16))[name = tensor<string, []>("op_1988_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_8_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106799680)))];
+            tensor<fp16, [1, 3, 512]> linear_75_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16, x = key_17_cast_fp16)[name = tensor<string, []>("linear_75_cast_fp16")];
+            tensor<int32, [4]> var_1992 = const()[name = tensor<string, []>("op_1992"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 3, 8, 64]> q_49_cast_fp16 = reshape(shape = var_1992, x = linear_75_cast_fp16)[name = tensor<string, []>("q_49_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_8_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107324032)))];
+            tensor<fp16, [1, 73, 512]> linear_76_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16, x = input_455_cast_fp16)[name = tensor<string, []>("linear_76_cast_fp16")];
+            tensor<int32, [4]> var_1996 = const()[name = tensor<string, []>("op_1996"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> k_33_cast_fp16 = reshape(shape = var_1996, x = linear_76_cast_fp16)[name = tensor<string, []>("k_33_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_8_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107848384)))];
+            tensor<fp16, [1, 73, 512]> linear_77_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16, x = input_455_cast_fp16)[name = tensor<string, []>("linear_77_cast_fp16")];
+            tensor<int32, [4]> var_2000 = const()[name = tensor<string, []>("op_2000"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> v_17_cast_fp16 = reshape(shape = var_2000, x = linear_77_cast_fp16)[name = tensor<string, []>("v_17_cast_fp16")];
+            tensor<int32, [4]> value_19_perm_0 = const()[name = tensor<string, []>("value_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108372736)))];
+            tensor<fp16, [1, 3, 8, 64]> var_2012_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2012_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108373824)))];
+            tensor<fp16, [1, 3, 8, 64]> var_2014_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2014_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_17_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_215_transpose_x_0 = const()[name = tensor<string, []>("x_215_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_215_transpose_y_0 = const()[name = tensor<string, []>("x_215_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 145]> var_2016_to_fp16 = const()[name = tensor<string, []>("op_2016_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108374912)))];
+            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2014_cast_fp16)[name = tensor<string, []>("transpose_165")];
+            tensor<fp16, [1, 8, 3, 145]> x_215_cast_fp16 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = var_2016_to_fp16)[name = tensor<string, []>("x_215_cast_fp16")];
+            tensor<int32, [8]> x_217_pad_0 = const()[name = tensor<string, []>("x_217_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_217_mode_0 = const()[name = tensor<string, []>("x_217_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_127_to_fp16 = const()[name = tensor<string, []>("const_127_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 3, 146]> x_217_cast_fp16 = pad(constant_val = const_127_to_fp16, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215_cast_fp16)[name = tensor<string, []>("x_217_cast_fp16")];
+            tensor<int32, [4]> var_2024 = const()[name = tensor<string, []>("op_2024"), val = tensor<int32, [4]>([1, 8, -1, 3])];
+            tensor<fp16, [1, 8, 146, 3]> x_219_cast_fp16 = reshape(shape = var_2024, x = x_217_cast_fp16)[name = tensor<string, []>("x_219_cast_fp16")];
+            tensor<int32, [4]> var_2028_begin_0 = const()[name = tensor<string, []>("op_2028_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2028_end_0 = const()[name = tensor<string, []>("op_2028_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
+            tensor<bool, [4]> var_2028_end_mask_0 = const()[name = tensor<string, []>("op_2028_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 145, 3]> var_2028_cast_fp16 = slice_by_index(begin = var_2028_begin_0, end = var_2028_end_0, end_mask = var_2028_end_mask_0, x = x_219_cast_fp16)[name = tensor<string, []>("op_2028_cast_fp16")];
+            tensor<int32, [4]> var_2029 = const()[name = tensor<string, []>("op_2029"), val = tensor<int32, [4]>([1, 8, 3, 145])];
+            tensor<fp16, [1, 8, 3, 145]> matrix_bd_33_cast_fp16 = reshape(shape = var_2029, x = var_2028_cast_fp16)[name = tensor<string, []>("matrix_bd_33_cast_fp16")];
+            tensor<bool, []> matrix_ac_17_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_17_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_17_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_17_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_67_perm_0 = const()[name = tensor<string, []>("transpose_67_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_68_perm_0 = const()[name = tensor<string, []>("transpose_68_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 73]> transpose_68 = transpose(perm = transpose_68_perm_0, x = k_33_cast_fp16)[name = tensor<string, []>("transpose_163")];
+            tensor<fp16, [1, 8, 3, 64]> transpose_67 = transpose(perm = transpose_67_perm_0, x = var_2012_cast_fp16)[name = tensor<string, []>("transpose_164")];
+            tensor<fp16, [1, 8, 3, 73]> matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_67, y = transpose_68)[name = tensor<string, []>("matrix_ac_17_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_35_begin_0 = const()[name = tensor<string, []>("matrix_bd_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_35_end_0 = const()[name = tensor<string, []>("matrix_bd_35_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
+            tensor<bool, [4]> matrix_bd_35_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 3, 73]> matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33_cast_fp16)[name = tensor<string, []>("matrix_bd_35_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_2038_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = tensor<string, []>("op_2038_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_33_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_33_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_33_cast_fp16 = mul(x = var_2038_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_33_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> scores_35_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_35_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_2044_cast_fp16 = softmax(axis = var_62, x = scores_35_cast_fp16)[name = tensor<string, []>("op_2044_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> input_457_cast_fp16 = select(a = var_40_to_fp16, b = var_2044_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_457_cast_fp16")];
+            tensor<bool, []> x_221_transpose_x_0 = const()[name = tensor<string, []>("x_221_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_221_transpose_y_0 = const()[name = tensor<string, []>("x_221_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 73, 64]> value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_17_cast_fp16)[name = tensor<string, []>("transpose_166")];
+            tensor<fp16, [1, 8, 3, 64]> x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457_cast_fp16, y = value_19_cast_fp16)[name = tensor<string, []>("x_221_cast_fp16")];
+            tensor<int32, [4]> var_2048_perm_0 = const()[name = tensor<string, []>("op_2048_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2049 = const()[name = tensor<string, []>("op_2049"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 3, 8, 64]> var_2048_cast_fp16 = transpose(perm = var_2048_perm_0, x = x_221_cast_fp16)[name = tensor<string, []>("transpose_162")];
+            tensor<fp16, [1, 3, 512]> input_459_cast_fp16 = reshape(shape = var_2049, x = var_2048_cast_fp16)[name = tensor<string, []>("input_459_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_8_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108523456)))];
+            tensor<fp16, [1, 3, 512]> linear_79_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16, x = input_459_cast_fp16)[name = tensor<string, []>("linear_79_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = tensor<string, []>("input_463_cast_fp16")];
+            tensor<int32, [1]> x_225_axes_0 = const()[name = tensor<string, []>("x_225_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109047808)))];
+            tensor<fp16, [512]> encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109048896)))];
+            tensor<fp16, [1, 3, 512]> x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = tensor<string, []>("x_225_cast_fp16")];
+            tensor<int32, [3]> input_465_perm_0 = const()[name = tensor<string, []>("input_465_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_467_pad_type_0 = const()[name = tensor<string, []>("input_467_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_467_strides_0 = const()[name = tensor<string, []>("input_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_467_pad_0 = const()[name = tensor<string, []>("input_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_467_dilations_0 = const()[name = tensor<string, []>("input_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_467_groups_0 = const()[name = tensor<string, []>("input_467_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_8_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109049984)))];
+            tensor<fp16, [1, 512, 3]> input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = tensor<string, []>("transpose_161")];
+            tensor<fp16, [1, 1024, 3]> input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16, x = input_465_cast_fp16)[name = tensor<string, []>("input_467_cast_fp16")];
+            tensor<int32, []> x_227_split_num_splits_0 = const()[name = tensor<string, []>("x_227_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_227_split_axis_0 = const()[name = tensor<string, []>("x_227_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 3]> x_227_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_227_split_cast_fp16_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467_cast_fp16)[name = tensor<string, []>("x_227_split_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = tensor<string, []>("x_227_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_227_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> input_469_cast_fp16 = select(a = var_40_to_fp16, b = x_227_cast_fp16, cond = var_418)[name = tensor<string, []>("input_469_cast_fp16")];
+            tensor<bool, []> new_x_35_interleave_0 = const()[name = tensor<string, []>("new_x_35_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 11]> new_x_35_cast_fp16 = concat(axis = var_62, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = tensor<string, []>("new_x_35_cast_fp16")];
+            tensor<int32, [3]> next_cache_17_begin_0 = const()[name = tensor<string, []>("next_cache_17_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_17_end_0 = const()[name = tensor<string, []>("next_cache_17_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> next_cache_17_end_mask_0 = const()[name = tensor<string, []>("next_cache_17_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 9]> next_cache_17_cast_fp16 = slice_by_index(begin = next_cache_17_begin_0, end = next_cache_17_end_0, end_mask = next_cache_17_end_mask_0, x = new_x_35_cast_fp16)[name = tensor<string, []>("next_cache_17_cast_fp16")];
+            tensor<int32, [3]> var_2090_begin_0 = const()[name = tensor<string, []>("op_2090_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> var_2090_end_0 = const()[name = tensor<string, []>("op_2090_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> var_2090_end_mask_0 = const()[name = tensor<string, []>("op_2090_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_2090_cast_fp16 = slice_by_index(begin = var_2090_begin_0, end = var_2090_end_0, end_mask = var_2090_end_mask_0, x = next_cache_17_cast_fp16)[name = tensor<string, []>("op_2090_cast_fp16")];
+            tensor<string, []> x_229_pad_type_0 = const()[name = tensor<string, []>("x_229_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_229_groups_0 = const()[name = tensor<string, []>("x_229_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_229_strides_0 = const()[name = tensor<string, []>("x_229_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_229_pad_0 = const()[name = tensor<string, []>("x_229_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_229_dilations_0 = const()[name = tensor<string, []>("x_229_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_8_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110098624)))];
+            tensor<fp16, [1, 512, 3]> x_229_cast_fp16 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16, x = new_x_35_cast_fp16)[name = tensor<string, []>("x_229_cast_fp16")];
+            tensor<int32, [3]> input_471_perm_0 = const()[name = tensor<string, []>("input_471_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_231_axes_0 = const()[name = tensor<string, []>("x_231_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110107904)))];
+            tensor<fp16, [512]> encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110108992)))];
+            tensor<fp16, [1, 3, 512]> input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = tensor<string, []>("transpose_160")];
+            tensor<fp16, [1, 3, 512]> x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = tensor<string, []>("x_231_cast_fp16")];
+            tensor<int32, [3]> input_473_perm_0 = const()[name = tensor<string, []>("input_473_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 3]> input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = tensor<string, []>("transpose_159")];
+            tensor<fp16, [1, 512, 3]> input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = tensor<string, []>("input_475_cast_fp16")];
+            tensor<string, []> x_233_pad_type_0 = const()[name = tensor<string, []>("x_233_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_233_strides_0 = const()[name = tensor<string, []>("x_233_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_233_pad_0 = const()[name = tensor<string, []>("x_233_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_233_dilations_0 = const()[name = tensor<string, []>("x_233_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_233_groups_0 = const()[name = tensor<string, []>("x_233_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_8_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110110080)))];
+            tensor<fp16, [1, 512, 3]> x_233_cast_fp16 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16, x = input_475_cast_fp16)[name = tensor<string, []>("x_233_cast_fp16")];
+            tensor<int32, [3]> input_477_perm_0 = const()[name = tensor<string, []>("input_477_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 3, 512]> input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = tensor<string, []>("transpose_158")];
+            tensor<fp16, [1, 3, 512]> input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = tensor<string, []>("input_479_cast_fp16")];
+            tensor<int32, [1]> input_481_axes_0 = const()[name = tensor<string, []>("input_481_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110634432)))];
+            tensor<fp16, [512]> encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110635520)))];
+            tensor<fp16, [1, 3, 512]> input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = tensor<string, []>("input_481_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_8_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110636608)))];
+            tensor<fp16, [1, 3, 2048]> linear_80_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16, x = input_481_cast_fp16)[name = tensor<string, []>("linear_80_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = tensor<string, []>("input_485_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_8_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112733824)))];
+            tensor<fp16, [1, 3, 512]> linear_81_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16, x = input_485_cast_fp16)[name = tensor<string, []>("linear_81_cast_fp16")];
+            tensor<fp16, []> var_2131_to_fp16 = const()[name = tensor<string, []>("op_2131_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_2132_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2131_to_fp16)[name = tensor<string, []>("op_2132_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2132_cast_fp16)[name = tensor<string, []>("input_491_cast_fp16")];
+            tensor<int32, [1]> input_493_axes_0 = const()[name = tensor<string, []>("input_493_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_8_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114831040)))];
+            tensor<fp16, [512]> encoder_layers_8_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114832128)))];
+            tensor<fp16, [1, 3, 512]> input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = tensor<string, []>("input_493_cast_fp16")];
+            tensor<int32, [4]> cache_37_begin_0 = const()[name = tensor<string, []>("cache_37_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_37_end_0 = const()[name = tensor<string, []>("cache_37_end_0"), val = tensor<int32, [4]>([10, 1, 70, 512])];
+            tensor<bool, [4]> cache_37_end_mask_0 = const()[name = tensor<string, []>("cache_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_37_squeeze_mask_0 = const()[name = tensor<string, []>("cache_37_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_37_cast_fp16 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_37_cast_fp16")];
+            tensor<int32, [4]> cache_39_begin_0 = const()[name = tensor<string, []>("cache_39_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_39_end_0 = const()[name = tensor<string, []>("cache_39_end_0"), val = tensor<int32, [4]>([10, 1, 512, 8])];
+            tensor<bool, [4]> cache_39_end_mask_0 = const()[name = tensor<string, []>("cache_39_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_39_squeeze_mask_0 = const()[name = tensor<string, []>("cache_39_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_39_cast_fp16 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_39_cast_fp16")];
+            tensor<int32, [1]> input_495_axes_0 = const()[name = tensor<string, []>("input_495_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114833216)))];
+            tensor<fp16, [512]> encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114834304)))];
+            tensor<fp16, [1, 3, 512]> input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = tensor<string, []>("input_495_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_9_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114835392)))];
+            tensor<fp16, [1, 3, 2048]> linear_82_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16, x = input_495_cast_fp16)[name = tensor<string, []>("linear_82_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = tensor<string, []>("input_499_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_9_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116932608)))];
+            tensor<fp16, [1, 3, 512]> linear_83_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16, x = input_499_cast_fp16)[name = tensor<string, []>("linear_83_cast_fp16")];
+            tensor<fp16, []> var_2166_to_fp16 = const()[name = tensor<string, []>("op_2166_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_2167_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2166_to_fp16)[name = tensor<string, []>("op_2167_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2167_cast_fp16)[name = tensor<string, []>("input_505_cast_fp16")];
+            tensor<int32, [1]> key_19_axes_0 = const()[name = tensor<string, []>("key_19_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119029824)))];
+            tensor<fp16, [512]> encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119030912)))];
+            tensor<fp16, [1, 3, 512]> key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = tensor<string, []>("key_19_cast_fp16")];
+            tensor<bool, []> input_507_interleave_0 = const()[name = tensor<string, []>("input_507_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 73, 512]> input_507_cast_fp16 = concat(axis = var_64, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = tensor<string, []>("input_507_cast_fp16")];
+            tensor<int32, [3]> var_2189_begin_0 = const()[name = tensor<string, []>("op_2189_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_2189_end_0 = const()[name = tensor<string, []>("op_2189_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_2189_end_mask_0 = const()[name = tensor<string, []>("op_2189_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 69, 512]> var_2189_cast_fp16 = slice_by_index(begin = var_2189_begin_0, end = var_2189_end_0, end_mask = var_2189_end_mask_0, x = cache_37_cast_fp16)[name = tensor<string, []>("op_2189_cast_fp16")];
+            tensor<int32, [3]> var_2192_begin_0 = const()[name = tensor<string, []>("op_2192_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_2192_end_0 = const()[name = tensor<string, []>("op_2192_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<bool, [3]> var_2192_end_mask_0 = const()[name = tensor<string, []>("op_2192_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 1, 512]> var_2192_cast_fp16 = slice_by_index(begin = var_2192_begin_0, end = var_2192_end_0, end_mask = var_2192_end_mask_0, x = key_19_cast_fp16)[name = tensor<string, []>("op_2192_cast_fp16")];
+            tensor<bool, []> var_2195_interleave_0 = const()[name = tensor<string, []>("op_2195_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_2195_cast_fp16 = concat(axis = var_64, interleave = var_2195_interleave_0, values = (var_2189_cast_fp16, var_2192_cast_fp16))[name = tensor<string, []>("op_2195_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_9_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119032000)))];
+            tensor<fp16, [1, 3, 512]> linear_84_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16, x = key_19_cast_fp16)[name = tensor<string, []>("linear_84_cast_fp16")];
+            tensor<int32, [4]> var_2199 = const()[name = tensor<string, []>("op_2199"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 3, 8, 64]> q_55_cast_fp16 = reshape(shape = var_2199, x = linear_84_cast_fp16)[name = tensor<string, []>("q_55_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_9_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119556352)))];
+            tensor<fp16, [1, 73, 512]> linear_85_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16, x = input_507_cast_fp16)[name = tensor<string, []>("linear_85_cast_fp16")];
+            tensor<int32, [4]> var_2203 = const()[name = tensor<string, []>("op_2203"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> k_37_cast_fp16 = reshape(shape = var_2203, x = linear_85_cast_fp16)[name = tensor<string, []>("k_37_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_9_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120080704)))];
+            tensor<fp16, [1, 73, 512]> linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16, x = input_507_cast_fp16)[name = tensor<string, []>("linear_86_cast_fp16")];
+            tensor<int32, [4]> var_2207 = const()[name = tensor<string, []>("op_2207"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> v_19_cast_fp16 = reshape(shape = var_2207, x = linear_86_cast_fp16)[name = tensor<string, []>("v_19_cast_fp16")];
+            tensor<int32, [4]> value_21_perm_0 = const()[name = tensor<string, []>("value_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120605056)))];
+            tensor<fp16, [1, 3, 8, 64]> var_2219_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2219_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120606144)))];
+            tensor<fp16, [1, 3, 8, 64]> var_2221_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2221_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_19_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_241_transpose_x_0 = const()[name = tensor<string, []>("x_241_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_241_transpose_y_0 = const()[name = tensor<string, []>("x_241_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 145]> var_2223_to_fp16 = const()[name = tensor<string, []>("op_2223_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120607232)))];
+            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2221_cast_fp16)[name = tensor<string, []>("transpose_156")];
+            tensor<fp16, [1, 8, 3, 145]> x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = var_2223_to_fp16)[name = tensor<string, []>("x_241_cast_fp16")];
+            tensor<int32, [8]> x_243_pad_0 = const()[name = tensor<string, []>("x_243_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_243_mode_0 = const()[name = tensor<string, []>("x_243_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_140_to_fp16 = const()[name = tensor<string, []>("const_140_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 3, 146]> x_243_cast_fp16 = pad(constant_val = const_140_to_fp16, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241_cast_fp16)[name = tensor<string, []>("x_243_cast_fp16")];
+            tensor<int32, [4]> var_2231 = const()[name = tensor<string, []>("op_2231"), val = tensor<int32, [4]>([1, 8, -1, 3])];
+            tensor<fp16, [1, 8, 146, 3]> x_245_cast_fp16 = reshape(shape = var_2231, x = x_243_cast_fp16)[name = tensor<string, []>("x_245_cast_fp16")];
+            tensor<int32, [4]> var_2235_begin_0 = const()[name = tensor<string, []>("op_2235_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2235_end_0 = const()[name = tensor<string, []>("op_2235_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
+            tensor<bool, [4]> var_2235_end_mask_0 = const()[name = tensor<string, []>("op_2235_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 145, 3]> var_2235_cast_fp16 = slice_by_index(begin = var_2235_begin_0, end = var_2235_end_0, end_mask = var_2235_end_mask_0, x = x_245_cast_fp16)[name = tensor<string, []>("op_2235_cast_fp16")];
+            tensor<int32, [4]> var_2236 = const()[name = tensor<string, []>("op_2236"), val = tensor<int32, [4]>([1, 8, 3, 145])];
+            tensor<fp16, [1, 8, 3, 145]> matrix_bd_37_cast_fp16 = reshape(shape = var_2236, x = var_2235_cast_fp16)[name = tensor<string, []>("matrix_bd_37_cast_fp16")];
+            tensor<bool, []> matrix_ac_19_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_19_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_19_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_19_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_69_perm_0 = const()[name = tensor<string, []>("transpose_69_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_70_perm_0 = const()[name = tensor<string, []>("transpose_70_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 73]> transpose_70 = transpose(perm = transpose_70_perm_0, x = k_37_cast_fp16)[name = tensor<string, []>("transpose_154")];
+            tensor<fp16, [1, 8, 3, 64]> transpose_69 = transpose(perm = transpose_69_perm_0, x = var_2219_cast_fp16)[name = tensor<string, []>("transpose_155")];
+            tensor<fp16, [1, 8, 3, 73]> matrix_ac_19_cast_fp16 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_69, y = transpose_70)[name = tensor<string, []>("matrix_ac_19_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_39_begin_0 = const()[name = tensor<string, []>("matrix_bd_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_39_end_0 = const()[name = tensor<string, []>("matrix_bd_39_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
+            tensor<bool, [4]> matrix_bd_39_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 3, 73]> matrix_bd_39_cast_fp16 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37_cast_fp16)[name = tensor<string, []>("matrix_bd_39_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_2245_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = tensor<string, []>("op_2245_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_37_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_37_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_37_cast_fp16 = mul(x = var_2245_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_37_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> scores_39_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_39_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_2251_cast_fp16 = softmax(axis = var_62, x = scores_39_cast_fp16)[name = tensor<string, []>("op_2251_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> input_509_cast_fp16 = select(a = var_40_to_fp16, b = var_2251_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_509_cast_fp16")];
+            tensor<bool, []> x_247_transpose_x_0 = const()[name = tensor<string, []>("x_247_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_247_transpose_y_0 = const()[name = tensor<string, []>("x_247_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 73, 64]> value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_19_cast_fp16)[name = tensor<string, []>("transpose_157")];
+            tensor<fp16, [1, 8, 3, 64]> x_247_cast_fp16 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509_cast_fp16, y = value_21_cast_fp16)[name = tensor<string, []>("x_247_cast_fp16")];
+            tensor<int32, [4]> var_2255_perm_0 = const()[name = tensor<string, []>("op_2255_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2256 = const()[name = tensor<string, []>("op_2256"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 3, 8, 64]> var_2255_cast_fp16 = transpose(perm = var_2255_perm_0, x = x_247_cast_fp16)[name = tensor<string, []>("transpose_153")];
+            tensor<fp16, [1, 3, 512]> input_511_cast_fp16 = reshape(shape = var_2256, x = var_2255_cast_fp16)[name = tensor<string, []>("input_511_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_9_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120755776)))];
+            tensor<fp16, [1, 3, 512]> linear_88_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16, x = input_511_cast_fp16)[name = tensor<string, []>("linear_88_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = tensor<string, []>("input_515_cast_fp16")];
+            tensor<int32, [1]> x_251_axes_0 = const()[name = tensor<string, []>("x_251_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121280128)))];
+            tensor<fp16, [512]> encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121281216)))];
+            tensor<fp16, [1, 3, 512]> x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = tensor<string, []>("x_251_cast_fp16")];
+            tensor<int32, [3]> input_517_perm_0 = const()[name = tensor<string, []>("input_517_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_519_pad_type_0 = const()[name = tensor<string, []>("input_519_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_519_strides_0 = const()[name = tensor<string, []>("input_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_519_pad_0 = const()[name = tensor<string, []>("input_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_519_dilations_0 = const()[name = tensor<string, []>("input_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_519_groups_0 = const()[name = tensor<string, []>("input_519_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_9_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121282304)))];
+            tensor<fp16, [1, 512, 3]> input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = tensor<string, []>("transpose_152")];
+            tensor<fp16, [1, 1024, 3]> input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16, x = input_517_cast_fp16)[name = tensor<string, []>("input_519_cast_fp16")];
+            tensor<int32, []> x_253_split_num_splits_0 = const()[name = tensor<string, []>("x_253_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_253_split_axis_0 = const()[name = tensor<string, []>("x_253_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 3]> x_253_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_253_split_cast_fp16_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519_cast_fp16)[name = tensor<string, []>("x_253_split_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = tensor<string, []>("x_253_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_253_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> input_521_cast_fp16 = select(a = var_40_to_fp16, b = x_253_cast_fp16, cond = var_418)[name = tensor<string, []>("input_521_cast_fp16")];
+            tensor<bool, []> new_x_39_interleave_0 = const()[name = tensor<string, []>("new_x_39_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 11]> new_x_39_cast_fp16 = concat(axis = var_62, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = tensor<string, []>("new_x_39_cast_fp16")];
+            tensor<int32, [3]> next_cache_19_begin_0 = const()[name = tensor<string, []>("next_cache_19_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_19_end_0 = const()[name = tensor<string, []>("next_cache_19_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> next_cache_19_end_mask_0 = const()[name = tensor<string, []>("next_cache_19_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 9]> next_cache_19_cast_fp16 = slice_by_index(begin = next_cache_19_begin_0, end = next_cache_19_end_0, end_mask = next_cache_19_end_mask_0, x = new_x_39_cast_fp16)[name = tensor<string, []>("next_cache_19_cast_fp16")];
+            tensor<int32, [3]> var_2297_begin_0 = const()[name = tensor<string, []>("op_2297_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> var_2297_end_0 = const()[name = tensor<string, []>("op_2297_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> var_2297_end_mask_0 = const()[name = tensor<string, []>("op_2297_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_2297_cast_fp16 = slice_by_index(begin = var_2297_begin_0, end = var_2297_end_0, end_mask = var_2297_end_mask_0, x = next_cache_19_cast_fp16)[name = tensor<string, []>("op_2297_cast_fp16")];
+            tensor<string, []> x_255_pad_type_0 = const()[name = tensor<string, []>("x_255_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_255_groups_0 = const()[name = tensor<string, []>("x_255_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_255_strides_0 = const()[name = tensor<string, []>("x_255_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_255_pad_0 = const()[name = tensor<string, []>("x_255_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_255_dilations_0 = const()[name = tensor<string, []>("x_255_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_9_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122330944)))];
+            tensor<fp16, [1, 512, 3]> x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16, x = new_x_39_cast_fp16)[name = tensor<string, []>("x_255_cast_fp16")];
+            tensor<int32, [3]> input_523_perm_0 = const()[name = tensor<string, []>("input_523_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_257_axes_0 = const()[name = tensor<string, []>("x_257_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122340224)))];
+            tensor<fp16, [512]> encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122341312)))];
+            tensor<fp16, [1, 3, 512]> input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = tensor<string, []>("transpose_151")];
+            tensor<fp16, [1, 3, 512]> x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = tensor<string, []>("x_257_cast_fp16")];
+            tensor<int32, [3]> input_525_perm_0 = const()[name = tensor<string, []>("input_525_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 3]> input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = tensor<string, []>("transpose_150")];
+            tensor<fp16, [1, 512, 3]> input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = tensor<string, []>("input_527_cast_fp16")];
+            tensor<string, []> x_259_pad_type_0 = const()[name = tensor<string, []>("x_259_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_259_strides_0 = const()[name = tensor<string, []>("x_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_259_pad_0 = const()[name = tensor<string, []>("x_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_259_dilations_0 = const()[name = tensor<string, []>("x_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_259_groups_0 = const()[name = tensor<string, []>("x_259_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_9_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122342400)))];
+            tensor<fp16, [1, 512, 3]> x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16, x = input_527_cast_fp16)[name = tensor<string, []>("x_259_cast_fp16")];
+            tensor<int32, [3]> input_529_perm_0 = const()[name = tensor<string, []>("input_529_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 3, 512]> input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = tensor<string, []>("transpose_149")];
+            tensor<fp16, [1, 3, 512]> input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = tensor<string, []>("input_531_cast_fp16")];
+            tensor<int32, [1]> input_533_axes_0 = const()[name = tensor<string, []>("input_533_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122866752)))];
+            tensor<fp16, [512]> encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122867840)))];
+            tensor<fp16, [1, 3, 512]> input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = tensor<string, []>("input_533_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_9_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122868928)))];
+            tensor<fp16, [1, 3, 2048]> linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16, x = input_533_cast_fp16)[name = tensor<string, []>("linear_89_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = tensor<string, []>("input_537_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_9_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124966144)))];
+            tensor<fp16, [1, 3, 512]> linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16, x = input_537_cast_fp16)[name = tensor<string, []>("linear_90_cast_fp16")];
+            tensor<fp16, []> var_2338_to_fp16 = const()[name = tensor<string, []>("op_2338_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_2339_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2338_to_fp16)[name = tensor<string, []>("op_2339_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2339_cast_fp16)[name = tensor<string, []>("input_543_cast_fp16")];
+            tensor<int32, [1]> input_545_axes_0 = const()[name = tensor<string, []>("input_545_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_9_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127063360)))];
+            tensor<fp16, [512]> encoder_layers_9_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127064448)))];
+            tensor<fp16, [1, 3, 512]> input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = tensor<string, []>("input_545_cast_fp16")];
+            tensor<int32, [4]> cache_41_begin_0 = const()[name = tensor<string, []>("cache_41_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_41_end_0 = const()[name = tensor<string, []>("cache_41_end_0"), val = tensor<int32, [4]>([11, 1, 70, 512])];
+            tensor<bool, [4]> cache_41_end_mask_0 = const()[name = tensor<string, []>("cache_41_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_41_squeeze_mask_0 = const()[name = tensor<string, []>("cache_41_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_41_cast_fp16 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_41_cast_fp16")];
+            tensor<int32, [4]> cache_43_begin_0 = const()[name = tensor<string, []>("cache_43_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_43_end_0 = const()[name = tensor<string, []>("cache_43_end_0"), val = tensor<int32, [4]>([11, 1, 512, 8])];
+            tensor<bool, [4]> cache_43_end_mask_0 = const()[name = tensor<string, []>("cache_43_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_43_squeeze_mask_0 = const()[name = tensor<string, []>("cache_43_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_43_cast_fp16 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_43_cast_fp16")];
+            tensor<int32, [1]> input_547_axes_0 = const()[name = tensor<string, []>("input_547_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127065536)))];
+            tensor<fp16, [512]> encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127066624)))];
+            tensor<fp16, [1, 3, 512]> input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = tensor<string, []>("input_547_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_10_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127067712)))];
+            tensor<fp16, [1, 3, 2048]> linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16, x = input_547_cast_fp16)[name = tensor<string, []>("linear_91_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = tensor<string, []>("input_551_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_10_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129164928)))];
+            tensor<fp16, [1, 3, 512]> linear_92_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16, x = input_551_cast_fp16)[name = tensor<string, []>("linear_92_cast_fp16")];
+            tensor<fp16, []> var_2373_to_fp16 = const()[name = tensor<string, []>("op_2373_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_2374_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2373_to_fp16)[name = tensor<string, []>("op_2374_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2374_cast_fp16)[name = tensor<string, []>("input_557_cast_fp16")];
+            tensor<int32, [1]> key_21_axes_0 = const()[name = tensor<string, []>("key_21_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131262144)))];
+            tensor<fp16, [512]> encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131263232)))];
+            tensor<fp16, [1, 3, 512]> key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = tensor<string, []>("key_21_cast_fp16")];
+            tensor<bool, []> input_559_interleave_0 = const()[name = tensor<string, []>("input_559_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 73, 512]> input_559_cast_fp16 = concat(axis = var_64, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = tensor<string, []>("input_559_cast_fp16")];
+            tensor<int32, [3]> var_2396_begin_0 = const()[name = tensor<string, []>("op_2396_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_2396_end_0 = const()[name = tensor<string, []>("op_2396_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_2396_end_mask_0 = const()[name = tensor<string, []>("op_2396_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 69, 512]> var_2396_cast_fp16 = slice_by_index(begin = var_2396_begin_0, end = var_2396_end_0, end_mask = var_2396_end_mask_0, x = cache_41_cast_fp16)[name = tensor<string, []>("op_2396_cast_fp16")];
+            tensor<int32, [3]> var_2399_begin_0 = const()[name = tensor<string, []>("op_2399_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_2399_end_0 = const()[name = tensor<string, []>("op_2399_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<bool, [3]> var_2399_end_mask_0 = const()[name = tensor<string, []>("op_2399_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 1, 512]> var_2399_cast_fp16 = slice_by_index(begin = var_2399_begin_0, end = var_2399_end_0, end_mask = var_2399_end_mask_0, x = key_21_cast_fp16)[name = tensor<string, []>("op_2399_cast_fp16")];
+            tensor<bool, []> var_2402_interleave_0 = const()[name = tensor<string, []>("op_2402_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_2402_cast_fp16 = concat(axis = var_64, interleave = var_2402_interleave_0, values = (var_2396_cast_fp16, var_2399_cast_fp16))[name = tensor<string, []>("op_2402_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_10_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131264320)))];
+            tensor<fp16, [1, 3, 512]> linear_93_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16, x = key_21_cast_fp16)[name = tensor<string, []>("linear_93_cast_fp16")];
+            tensor<int32, [4]> var_2406 = const()[name = tensor<string, []>("op_2406"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 3, 8, 64]> q_61_cast_fp16 = reshape(shape = var_2406, x = linear_93_cast_fp16)[name = tensor<string, []>("q_61_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_10_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131788672)))];
+            tensor<fp16, [1, 73, 512]> linear_94_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16, x = input_559_cast_fp16)[name = tensor<string, []>("linear_94_cast_fp16")];
+            tensor<int32, [4]> var_2410 = const()[name = tensor<string, []>("op_2410"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> k_41_cast_fp16 = reshape(shape = var_2410, x = linear_94_cast_fp16)[name = tensor<string, []>("k_41_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_10_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132313024)))];
+            tensor<fp16, [1, 73, 512]> linear_95_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16, x = input_559_cast_fp16)[name = tensor<string, []>("linear_95_cast_fp16")];
+            tensor<int32, [4]> var_2414 = const()[name = tensor<string, []>("op_2414"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> v_21_cast_fp16 = reshape(shape = var_2414, x = linear_95_cast_fp16)[name = tensor<string, []>("v_21_cast_fp16")];
+            tensor<int32, [4]> value_23_perm_0 = const()[name = tensor<string, []>("value_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132837376)))];
+            tensor<fp16, [1, 3, 8, 64]> var_2426_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2426_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132838464)))];
+            tensor<fp16, [1, 3, 8, 64]> var_2428_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2428_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_21_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_267_transpose_x_0 = const()[name = tensor<string, []>("x_267_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_267_transpose_y_0 = const()[name = tensor<string, []>("x_267_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 145]> var_2430_to_fp16 = const()[name = tensor<string, []>("op_2430_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132839552)))];
+            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2428_cast_fp16)[name = tensor<string, []>("transpose_147")];
+            tensor<fp16, [1, 8, 3, 145]> x_267_cast_fp16 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = var_2430_to_fp16)[name = tensor<string, []>("x_267_cast_fp16")];
+            tensor<int32, [8]> x_269_pad_0 = const()[name = tensor<string, []>("x_269_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_269_mode_0 = const()[name = tensor<string, []>("x_269_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_153_to_fp16 = const()[name = tensor<string, []>("const_153_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 3, 146]> x_269_cast_fp16 = pad(constant_val = const_153_to_fp16, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267_cast_fp16)[name = tensor<string, []>("x_269_cast_fp16")];
+            tensor<int32, [4]> var_2438 = const()[name = tensor<string, []>("op_2438"), val = tensor<int32, [4]>([1, 8, -1, 3])];
+            tensor<fp16, [1, 8, 146, 3]> x_271_cast_fp16 = reshape(shape = var_2438, x = x_269_cast_fp16)[name = tensor<string, []>("x_271_cast_fp16")];
+            tensor<int32, [4]> var_2442_begin_0 = const()[name = tensor<string, []>("op_2442_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2442_end_0 = const()[name = tensor<string, []>("op_2442_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
+            tensor<bool, [4]> var_2442_end_mask_0 = const()[name = tensor<string, []>("op_2442_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 145, 3]> var_2442_cast_fp16 = slice_by_index(begin = var_2442_begin_0, end = var_2442_end_0, end_mask = var_2442_end_mask_0, x = x_271_cast_fp16)[name = tensor<string, []>("op_2442_cast_fp16")];
+            tensor<int32, [4]> var_2443 = const()[name = tensor<string, []>("op_2443"), val = tensor<int32, [4]>([1, 8, 3, 145])];
+            tensor<fp16, [1, 8, 3, 145]> matrix_bd_41_cast_fp16 = reshape(shape = var_2443, x = var_2442_cast_fp16)[name = tensor<string, []>("matrix_bd_41_cast_fp16")];
+            tensor<bool, []> matrix_ac_21_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_21_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_21_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_21_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_71_perm_0 = const()[name = tensor<string, []>("transpose_71_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_72_perm_0 = const()[name = tensor<string, []>("transpose_72_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 73]> transpose_72 = transpose(perm = transpose_72_perm_0, x = k_41_cast_fp16)[name = tensor<string, []>("transpose_145")];
+            tensor<fp16, [1, 8, 3, 64]> transpose_71 = transpose(perm = transpose_71_perm_0, x = var_2426_cast_fp16)[name = tensor<string, []>("transpose_146")];
+            tensor<fp16, [1, 8, 3, 73]> matrix_ac_21_cast_fp16 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_71, y = transpose_72)[name = tensor<string, []>("matrix_ac_21_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_43_begin_0 = const()[name = tensor<string, []>("matrix_bd_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_43_end_0 = const()[name = tensor<string, []>("matrix_bd_43_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
+            tensor<bool, [4]> matrix_bd_43_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 3, 73]> matrix_bd_43_cast_fp16 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41_cast_fp16)[name = tensor<string, []>("matrix_bd_43_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_2452_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = tensor<string, []>("op_2452_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_41_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_41_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_41_cast_fp16 = mul(x = var_2452_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_41_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> scores_43_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_43_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_2458_cast_fp16 = softmax(axis = var_62, x = scores_43_cast_fp16)[name = tensor<string, []>("op_2458_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> input_561_cast_fp16 = select(a = var_40_to_fp16, b = var_2458_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_561_cast_fp16")];
+            tensor<bool, []> x_273_transpose_x_0 = const()[name = tensor<string, []>("x_273_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_273_transpose_y_0 = const()[name = tensor<string, []>("x_273_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 73, 64]> value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_21_cast_fp16)[name = tensor<string, []>("transpose_148")];
+            tensor<fp16, [1, 8, 3, 64]> x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561_cast_fp16, y = value_23_cast_fp16)[name = tensor<string, []>("x_273_cast_fp16")];
+            tensor<int32, [4]> var_2462_perm_0 = const()[name = tensor<string, []>("op_2462_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2463 = const()[name = tensor<string, []>("op_2463"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 3, 8, 64]> var_2462_cast_fp16 = transpose(perm = var_2462_perm_0, x = x_273_cast_fp16)[name = tensor<string, []>("transpose_144")];
+            tensor<fp16, [1, 3, 512]> input_563_cast_fp16 = reshape(shape = var_2463, x = var_2462_cast_fp16)[name = tensor<string, []>("input_563_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_10_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132988096)))];
+            tensor<fp16, [1, 3, 512]> linear_97_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16, x = input_563_cast_fp16)[name = tensor<string, []>("linear_97_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = tensor<string, []>("input_567_cast_fp16")];
+            tensor<int32, [1]> x_277_axes_0 = const()[name = tensor<string, []>("x_277_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133512448)))];
+            tensor<fp16, [512]> encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133513536)))];
+            tensor<fp16, [1, 3, 512]> x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = tensor<string, []>("x_277_cast_fp16")];
+            tensor<int32, [3]> input_569_perm_0 = const()[name = tensor<string, []>("input_569_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_571_pad_type_0 = const()[name = tensor<string, []>("input_571_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_571_strides_0 = const()[name = tensor<string, []>("input_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_571_pad_0 = const()[name = tensor<string, []>("input_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_571_dilations_0 = const()[name = tensor<string, []>("input_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_571_groups_0 = const()[name = tensor<string, []>("input_571_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_10_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133514624)))];
+            tensor<fp16, [1, 512, 3]> input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = tensor<string, []>("transpose_143")];
+            tensor<fp16, [1, 1024, 3]> input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16, x = input_569_cast_fp16)[name = tensor<string, []>("input_571_cast_fp16")];
+            tensor<int32, []> x_279_split_num_splits_0 = const()[name = tensor<string, []>("x_279_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_279_split_axis_0 = const()[name = tensor<string, []>("x_279_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 3]> x_279_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_279_split_cast_fp16_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571_cast_fp16)[name = tensor<string, []>("x_279_split_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = tensor<string, []>("x_279_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_279_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> input_573_cast_fp16 = select(a = var_40_to_fp16, b = x_279_cast_fp16, cond = var_418)[name = tensor<string, []>("input_573_cast_fp16")];
+            tensor<bool, []> new_x_43_interleave_0 = const()[name = tensor<string, []>("new_x_43_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 11]> new_x_43_cast_fp16 = concat(axis = var_62, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = tensor<string, []>("new_x_43_cast_fp16")];
+            tensor<int32, [3]> next_cache_21_begin_0 = const()[name = tensor<string, []>("next_cache_21_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_21_end_0 = const()[name = tensor<string, []>("next_cache_21_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> next_cache_21_end_mask_0 = const()[name = tensor<string, []>("next_cache_21_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 9]> next_cache_21_cast_fp16 = slice_by_index(begin = next_cache_21_begin_0, end = next_cache_21_end_0, end_mask = next_cache_21_end_mask_0, x = new_x_43_cast_fp16)[name = tensor<string, []>("next_cache_21_cast_fp16")];
+            tensor<int32, [3]> var_2504_begin_0 = const()[name = tensor<string, []>("op_2504_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> var_2504_end_0 = const()[name = tensor<string, []>("op_2504_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> var_2504_end_mask_0 = const()[name = tensor<string, []>("op_2504_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_2504_cast_fp16 = slice_by_index(begin = var_2504_begin_0, end = var_2504_end_0, end_mask = var_2504_end_mask_0, x = next_cache_21_cast_fp16)[name = tensor<string, []>("op_2504_cast_fp16")];
+            tensor<string, []> x_281_pad_type_0 = const()[name = tensor<string, []>("x_281_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_281_groups_0 = const()[name = tensor<string, []>("x_281_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_281_strides_0 = const()[name = tensor<string, []>("x_281_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_281_pad_0 = const()[name = tensor<string, []>("x_281_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_281_dilations_0 = const()[name = tensor<string, []>("x_281_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_10_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134563264)))];
+            tensor<fp16, [1, 512, 3]> x_281_cast_fp16 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16, x = new_x_43_cast_fp16)[name = tensor<string, []>("x_281_cast_fp16")];
+            tensor<int32, [3]> input_575_perm_0 = const()[name = tensor<string, []>("input_575_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_283_axes_0 = const()[name = tensor<string, []>("x_283_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134572544)))];
+            tensor<fp16, [512]> encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134573632)))];
+            tensor<fp16, [1, 3, 512]> input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = tensor<string, []>("transpose_142")];
+            tensor<fp16, [1, 3, 512]> x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = tensor<string, []>("x_283_cast_fp16")];
+            tensor<int32, [3]> input_577_perm_0 = const()[name = tensor<string, []>("input_577_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 3]> input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = tensor<string, []>("transpose_141")];
+            tensor<fp16, [1, 512, 3]> input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = tensor<string, []>("input_579_cast_fp16")];
+            tensor<string, []> x_285_pad_type_0 = const()[name = tensor<string, []>("x_285_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_285_strides_0 = const()[name = tensor<string, []>("x_285_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_285_pad_0 = const()[name = tensor<string, []>("x_285_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_285_dilations_0 = const()[name = tensor<string, []>("x_285_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_285_groups_0 = const()[name = tensor<string, []>("x_285_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_10_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134574720)))];
+            tensor<fp16, [1, 512, 3]> x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16, x = input_579_cast_fp16)[name = tensor<string, []>("x_285_cast_fp16")];
+            tensor<int32, [3]> input_581_perm_0 = const()[name = tensor<string, []>("input_581_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 3, 512]> input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = tensor<string, []>("transpose_140")];
+            tensor<fp16, [1, 3, 512]> input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = tensor<string, []>("input_583_cast_fp16")];
+            tensor<int32, [1]> input_585_axes_0 = const()[name = tensor<string, []>("input_585_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135099072)))];
+            tensor<fp16, [512]> encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135100160)))];
+            tensor<fp16, [1, 3, 512]> input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = tensor<string, []>("input_585_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_10_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135101248)))];
+            tensor<fp16, [1, 3, 2048]> linear_98_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16, x = input_585_cast_fp16)[name = tensor<string, []>("linear_98_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = tensor<string, []>("input_589_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_10_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137198464)))];
+            tensor<fp16, [1, 3, 512]> linear_99_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16, x = input_589_cast_fp16)[name = tensor<string, []>("linear_99_cast_fp16")];
+            tensor<fp16, []> var_2545_to_fp16 = const()[name = tensor<string, []>("op_2545_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_2546_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2545_to_fp16)[name = tensor<string, []>("op_2546_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2546_cast_fp16)[name = tensor<string, []>("input_595_cast_fp16")];
+            tensor<int32, [1]> input_597_axes_0 = const()[name = tensor<string, []>("input_597_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_10_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139295680)))];
+            tensor<fp16, [512]> encoder_layers_10_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139296768)))];
+            tensor<fp16, [1, 3, 512]> input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = tensor<string, []>("input_597_cast_fp16")];
+            tensor<int32, [4]> cache_45_begin_0 = const()[name = tensor<string, []>("cache_45_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_45_end_0 = const()[name = tensor<string, []>("cache_45_end_0"), val = tensor<int32, [4]>([12, 1, 70, 512])];
+            tensor<bool, [4]> cache_45_end_mask_0 = const()[name = tensor<string, []>("cache_45_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_45_squeeze_mask_0 = const()[name = tensor<string, []>("cache_45_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_45_cast_fp16 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_45_cast_fp16")];
+            tensor<int32, [4]> cache_47_begin_0 = const()[name = tensor<string, []>("cache_47_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_47_end_0 = const()[name = tensor<string, []>("cache_47_end_0"), val = tensor<int32, [4]>([12, 1, 512, 8])];
+            tensor<bool, [4]> cache_47_end_mask_0 = const()[name = tensor<string, []>("cache_47_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_47_squeeze_mask_0 = const()[name = tensor<string, []>("cache_47_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_47_cast_fp16 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_47_cast_fp16")];
+            tensor<int32, [1]> input_599_axes_0 = const()[name = tensor<string, []>("input_599_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139297856)))];
+            tensor<fp16, [512]> encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139298944)))];
+            tensor<fp16, [1, 3, 512]> input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = tensor<string, []>("input_599_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_11_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139300032)))];
+            tensor<fp16, [1, 3, 2048]> linear_100_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16, x = input_599_cast_fp16)[name = tensor<string, []>("linear_100_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = tensor<string, []>("input_603_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_11_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141397248)))];
+            tensor<fp16, [1, 3, 512]> linear_101_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16, x = input_603_cast_fp16)[name = tensor<string, []>("linear_101_cast_fp16")];
+            tensor<fp16, []> var_2580_to_fp16 = const()[name = tensor<string, []>("op_2580_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_2581_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2580_to_fp16)[name = tensor<string, []>("op_2581_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2581_cast_fp16)[name = tensor<string, []>("input_609_cast_fp16")];
+            tensor<int32, [1]> key_23_axes_0 = const()[name = tensor<string, []>("key_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143494464)))];
+            tensor<fp16, [512]> encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143495552)))];
+            tensor<fp16, [1, 3, 512]> key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = tensor<string, []>("key_23_cast_fp16")];
+            tensor<bool, []> input_611_interleave_0 = const()[name = tensor<string, []>("input_611_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 73, 512]> input_611_cast_fp16 = concat(axis = var_64, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = tensor<string, []>("input_611_cast_fp16")];
+            tensor<int32, [3]> var_2603_begin_0 = const()[name = tensor<string, []>("op_2603_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_2603_end_0 = const()[name = tensor<string, []>("op_2603_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_2603_end_mask_0 = const()[name = tensor<string, []>("op_2603_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 69, 512]> var_2603_cast_fp16 = slice_by_index(begin = var_2603_begin_0, end = var_2603_end_0, end_mask = var_2603_end_mask_0, x = cache_45_cast_fp16)[name = tensor<string, []>("op_2603_cast_fp16")];
+            tensor<int32, [3]> var_2606_begin_0 = const()[name = tensor<string, []>("op_2606_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_2606_end_0 = const()[name = tensor<string, []>("op_2606_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<bool, [3]> var_2606_end_mask_0 = const()[name = tensor<string, []>("op_2606_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 1, 512]> var_2606_cast_fp16 = slice_by_index(begin = var_2606_begin_0, end = var_2606_end_0, end_mask = var_2606_end_mask_0, x = key_23_cast_fp16)[name = tensor<string, []>("op_2606_cast_fp16")];
+            tensor<bool, []> var_2609_interleave_0 = const()[name = tensor<string, []>("op_2609_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_2609_cast_fp16 = concat(axis = var_64, interleave = var_2609_interleave_0, values = (var_2603_cast_fp16, var_2606_cast_fp16))[name = tensor<string, []>("op_2609_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_11_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143496640)))];
+            tensor<fp16, [1, 3, 512]> linear_102_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16, x = key_23_cast_fp16)[name = tensor<string, []>("linear_102_cast_fp16")];
+            tensor<int32, [4]> var_2613 = const()[name = tensor<string, []>("op_2613"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 3, 8, 64]> q_67_cast_fp16 = reshape(shape = var_2613, x = linear_102_cast_fp16)[name = tensor<string, []>("q_67_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_11_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144020992)))];
+            tensor<fp16, [1, 73, 512]> linear_103_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16, x = input_611_cast_fp16)[name = tensor<string, []>("linear_103_cast_fp16")];
+            tensor<int32, [4]> var_2617 = const()[name = tensor<string, []>("op_2617"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> k_45_cast_fp16 = reshape(shape = var_2617, x = linear_103_cast_fp16)[name = tensor<string, []>("k_45_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_11_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144545344)))];
+            tensor<fp16, [1, 73, 512]> linear_104_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16, x = input_611_cast_fp16)[name = tensor<string, []>("linear_104_cast_fp16")];
+            tensor<int32, [4]> var_2621 = const()[name = tensor<string, []>("op_2621"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> v_23_cast_fp16 = reshape(shape = var_2621, x = linear_104_cast_fp16)[name = tensor<string, []>("v_23_cast_fp16")];
+            tensor<int32, [4]> value_25_perm_0 = const()[name = tensor<string, []>("value_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145069696)))];
+            tensor<fp16, [1, 3, 8, 64]> var_2633_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2633_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145070784)))];
+            tensor<fp16, [1, 3, 8, 64]> var_2635_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2635_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_23_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_293_transpose_x_0 = const()[name = tensor<string, []>("x_293_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_293_transpose_y_0 = const()[name = tensor<string, []>("x_293_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 145]> var_2637_to_fp16 = const()[name = tensor<string, []>("op_2637_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145071872)))];
+            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2635_cast_fp16)[name = tensor<string, []>("transpose_138")];
+            tensor<fp16, [1, 8, 3, 145]> x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = var_2637_to_fp16)[name = tensor<string, []>("x_293_cast_fp16")];
+            tensor<int32, [8]> x_295_pad_0 = const()[name = tensor<string, []>("x_295_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_295_mode_0 = const()[name = tensor<string, []>("x_295_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_166_to_fp16 = const()[name = tensor<string, []>("const_166_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 3, 146]> x_295_cast_fp16 = pad(constant_val = const_166_to_fp16, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293_cast_fp16)[name = tensor<string, []>("x_295_cast_fp16")];
+            tensor<int32, [4]> var_2645 = const()[name = tensor<string, []>("op_2645"), val = tensor<int32, [4]>([1, 8, -1, 3])];
+            tensor<fp16, [1, 8, 146, 3]> x_297_cast_fp16 = reshape(shape = var_2645, x = x_295_cast_fp16)[name = tensor<string, []>("x_297_cast_fp16")];
+            tensor<int32, [4]> var_2649_begin_0 = const()[name = tensor<string, []>("op_2649_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2649_end_0 = const()[name = tensor<string, []>("op_2649_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
+            tensor<bool, [4]> var_2649_end_mask_0 = const()[name = tensor<string, []>("op_2649_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 145, 3]> var_2649_cast_fp16 = slice_by_index(begin = var_2649_begin_0, end = var_2649_end_0, end_mask = var_2649_end_mask_0, x = x_297_cast_fp16)[name = tensor<string, []>("op_2649_cast_fp16")];
+            tensor<int32, [4]> var_2650 = const()[name = tensor<string, []>("op_2650"), val = tensor<int32, [4]>([1, 8, 3, 145])];
+            tensor<fp16, [1, 8, 3, 145]> matrix_bd_45_cast_fp16 = reshape(shape = var_2650, x = var_2649_cast_fp16)[name = tensor<string, []>("matrix_bd_45_cast_fp16")];
+            tensor<bool, []> matrix_ac_23_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_23_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_23_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_23_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_73_perm_0 = const()[name = tensor<string, []>("transpose_73_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_74_perm_0 = const()[name = tensor<string, []>("transpose_74_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 73]> transpose_74 = transpose(perm = transpose_74_perm_0, x = k_45_cast_fp16)[name = tensor<string, []>("transpose_136")];
+            tensor<fp16, [1, 8, 3, 64]> transpose_73 = transpose(perm = transpose_73_perm_0, x = var_2633_cast_fp16)[name = tensor<string, []>("transpose_137")];
+            tensor<fp16, [1, 8, 3, 73]> matrix_ac_23_cast_fp16 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_73, y = transpose_74)[name = tensor<string, []>("matrix_ac_23_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_47_begin_0 = const()[name = tensor<string, []>("matrix_bd_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_47_end_0 = const()[name = tensor<string, []>("matrix_bd_47_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
+            tensor<bool, [4]> matrix_bd_47_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 3, 73]> matrix_bd_47_cast_fp16 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45_cast_fp16)[name = tensor<string, []>("matrix_bd_47_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_2659_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = tensor<string, []>("op_2659_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_45_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_45_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_45_cast_fp16 = mul(x = var_2659_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_45_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> scores_47_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_47_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_2665_cast_fp16 = softmax(axis = var_62, x = scores_47_cast_fp16)[name = tensor<string, []>("op_2665_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> input_613_cast_fp16 = select(a = var_40_to_fp16, b = var_2665_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_613_cast_fp16")];
+            tensor<bool, []> x_299_transpose_x_0 = const()[name = tensor<string, []>("x_299_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_299_transpose_y_0 = const()[name = tensor<string, []>("x_299_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 73, 64]> value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_23_cast_fp16)[name = tensor<string, []>("transpose_139")];
+            tensor<fp16, [1, 8, 3, 64]> x_299_cast_fp16 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613_cast_fp16, y = value_25_cast_fp16)[name = tensor<string, []>("x_299_cast_fp16")];
+            tensor<int32, [4]> var_2669_perm_0 = const()[name = tensor<string, []>("op_2669_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2670 = const()[name = tensor<string, []>("op_2670"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 3, 8, 64]> var_2669_cast_fp16 = transpose(perm = var_2669_perm_0, x = x_299_cast_fp16)[name = tensor<string, []>("transpose_135")];
+            tensor<fp16, [1, 3, 512]> input_615_cast_fp16 = reshape(shape = var_2670, x = var_2669_cast_fp16)[name = tensor<string, []>("input_615_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_11_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145220416)))];
+            tensor<fp16, [1, 3, 512]> linear_106_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16, x = input_615_cast_fp16)[name = tensor<string, []>("linear_106_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = tensor<string, []>("input_619_cast_fp16")];
+            tensor<int32, [1]> x_303_axes_0 = const()[name = tensor<string, []>("x_303_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145744768)))];
+            tensor<fp16, [512]> encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145745856)))];
+            tensor<fp16, [1, 3, 512]> x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = tensor<string, []>("x_303_cast_fp16")];
+            tensor<int32, [3]> input_621_perm_0 = const()[name = tensor<string, []>("input_621_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_623_pad_type_0 = const()[name = tensor<string, []>("input_623_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_623_strides_0 = const()[name = tensor<string, []>("input_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_623_pad_0 = const()[name = tensor<string, []>("input_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_623_dilations_0 = const()[name = tensor<string, []>("input_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_623_groups_0 = const()[name = tensor<string, []>("input_623_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_11_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145746944)))];
+            tensor<fp16, [1, 512, 3]> input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = tensor<string, []>("transpose_134")];
+            tensor<fp16, [1, 1024, 3]> input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16, x = input_621_cast_fp16)[name = tensor<string, []>("input_623_cast_fp16")];
+            tensor<int32, []> x_305_split_num_splits_0 = const()[name = tensor<string, []>("x_305_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_305_split_axis_0 = const()[name = tensor<string, []>("x_305_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 3]> x_305_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623_cast_fp16)[name = tensor<string, []>("x_305_split_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = tensor<string, []>("x_305_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_305_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> input_625_cast_fp16 = select(a = var_40_to_fp16, b = x_305_cast_fp16, cond = var_418)[name = tensor<string, []>("input_625_cast_fp16")];
+            tensor<bool, []> new_x_47_interleave_0 = const()[name = tensor<string, []>("new_x_47_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 11]> new_x_47_cast_fp16 = concat(axis = var_62, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = tensor<string, []>("new_x_47_cast_fp16")];
+            tensor<int32, [3]> next_cache_23_begin_0 = const()[name = tensor<string, []>("next_cache_23_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_23_end_0 = const()[name = tensor<string, []>("next_cache_23_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> next_cache_23_end_mask_0 = const()[name = tensor<string, []>("next_cache_23_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 9]> next_cache_23_cast_fp16 = slice_by_index(begin = next_cache_23_begin_0, end = next_cache_23_end_0, end_mask = next_cache_23_end_mask_0, x = new_x_47_cast_fp16)[name = tensor<string, []>("next_cache_23_cast_fp16")];
+            tensor<int32, [3]> var_2711_begin_0 = const()[name = tensor<string, []>("op_2711_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> var_2711_end_0 = const()[name = tensor<string, []>("op_2711_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> var_2711_end_mask_0 = const()[name = tensor<string, []>("op_2711_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_2711_cast_fp16 = slice_by_index(begin = var_2711_begin_0, end = var_2711_end_0, end_mask = var_2711_end_mask_0, x = next_cache_23_cast_fp16)[name = tensor<string, []>("op_2711_cast_fp16")];
+            tensor<string, []> x_307_pad_type_0 = const()[name = tensor<string, []>("x_307_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_307_groups_0 = const()[name = tensor<string, []>("x_307_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_307_strides_0 = const()[name = tensor<string, []>("x_307_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_307_pad_0 = const()[name = tensor<string, []>("x_307_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_307_dilations_0 = const()[name = tensor<string, []>("x_307_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_11_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146795584)))];
+            tensor<fp16, [1, 512, 3]> x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16, x = new_x_47_cast_fp16)[name = tensor<string, []>("x_307_cast_fp16")];
+            tensor<int32, [3]> input_627_perm_0 = const()[name = tensor<string, []>("input_627_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_309_axes_0 = const()[name = tensor<string, []>("x_309_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146804864)))];
+            tensor<fp16, [512]> encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146805952)))];
+            tensor<fp16, [1, 3, 512]> input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = tensor<string, []>("transpose_133")];
+            tensor<fp16, [1, 3, 512]> x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = tensor<string, []>("x_309_cast_fp16")];
+            tensor<int32, [3]> input_629_perm_0 = const()[name = tensor<string, []>("input_629_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 3]> input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = tensor<string, []>("transpose_132")];
+            tensor<fp16, [1, 512, 3]> input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = tensor<string, []>("input_631_cast_fp16")];
+            tensor<string, []> x_311_pad_type_0 = const()[name = tensor<string, []>("x_311_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_311_strides_0 = const()[name = tensor<string, []>("x_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_311_pad_0 = const()[name = tensor<string, []>("x_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_311_dilations_0 = const()[name = tensor<string, []>("x_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_311_groups_0 = const()[name = tensor<string, []>("x_311_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_11_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146807040)))];
+            tensor<fp16, [1, 512, 3]> x_311_cast_fp16 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16, x = input_631_cast_fp16)[name = tensor<string, []>("x_311_cast_fp16")];
+            tensor<int32, [3]> input_633_perm_0 = const()[name = tensor<string, []>("input_633_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 3, 512]> input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = tensor<string, []>("transpose_131")];
+            tensor<fp16, [1, 3, 512]> input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = tensor<string, []>("input_635_cast_fp16")];
+            tensor<int32, [1]> input_637_axes_0 = const()[name = tensor<string, []>("input_637_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147331392)))];
+            tensor<fp16, [512]> encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147332480)))];
+            tensor<fp16, [1, 3, 512]> input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = tensor<string, []>("input_637_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_11_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147333568)))];
+            tensor<fp16, [1, 3, 2048]> linear_107_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16, x = input_637_cast_fp16)[name = tensor<string, []>("linear_107_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = tensor<string, []>("input_641_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_11_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149430784)))];
+            tensor<fp16, [1, 3, 512]> linear_108_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16, x = input_641_cast_fp16)[name = tensor<string, []>("linear_108_cast_fp16")];
+            tensor<fp16, []> var_2752_to_fp16 = const()[name = tensor<string, []>("op_2752_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_2753_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2752_to_fp16)[name = tensor<string, []>("op_2753_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2753_cast_fp16)[name = tensor<string, []>("input_647_cast_fp16")];
+            tensor<int32, [1]> input_649_axes_0 = const()[name = tensor<string, []>("input_649_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_11_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151528000)))];
+            tensor<fp16, [512]> encoder_layers_11_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151529088)))];
+            tensor<fp16, [1, 3, 512]> input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = tensor<string, []>("input_649_cast_fp16")];
+            tensor<int32, [4]> cache_49_begin_0 = const()[name = tensor<string, []>("cache_49_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_49_end_0 = const()[name = tensor<string, []>("cache_49_end_0"), val = tensor<int32, [4]>([13, 1, 70, 512])];
+            tensor<bool, [4]> cache_49_end_mask_0 = const()[name = tensor<string, []>("cache_49_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_49_squeeze_mask_0 = const()[name = tensor<string, []>("cache_49_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_49_cast_fp16 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_49_cast_fp16")];
+            tensor<int32, [4]> cache_51_begin_0 = const()[name = tensor<string, []>("cache_51_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_51_end_0 = const()[name = tensor<string, []>("cache_51_end_0"), val = tensor<int32, [4]>([13, 1, 512, 8])];
+            tensor<bool, [4]> cache_51_end_mask_0 = const()[name = tensor<string, []>("cache_51_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_51_squeeze_mask_0 = const()[name = tensor<string, []>("cache_51_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_51_cast_fp16 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_51_cast_fp16")];
+            tensor<int32, [1]> input_651_axes_0 = const()[name = tensor<string, []>("input_651_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151530176)))];
+            tensor<fp16, [512]> encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151531264)))];
+            tensor<fp16, [1, 3, 512]> input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = tensor<string, []>("input_651_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_12_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151532352)))];
+            tensor<fp16, [1, 3, 2048]> linear_109_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16, x = input_651_cast_fp16)[name = tensor<string, []>("linear_109_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = tensor<string, []>("input_655_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_12_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153629568)))];
+            tensor<fp16, [1, 3, 512]> linear_110_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16, x = input_655_cast_fp16)[name = tensor<string, []>("linear_110_cast_fp16")];
+            tensor<fp16, []> var_2787_to_fp16 = const()[name = tensor<string, []>("op_2787_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_2788_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_2787_to_fp16)[name = tensor<string, []>("op_2788_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_2788_cast_fp16)[name = tensor<string, []>("input_661_cast_fp16")];
+            tensor<int32, [1]> key_25_axes_0 = const()[name = tensor<string, []>("key_25_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155726784)))];
+            tensor<fp16, [512]> encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155727872)))];
+            tensor<fp16, [1, 3, 512]> key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = tensor<string, []>("key_25_cast_fp16")];
+            tensor<bool, []> input_663_interleave_0 = const()[name = tensor<string, []>("input_663_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 73, 512]> input_663_cast_fp16 = concat(axis = var_64, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = tensor<string, []>("input_663_cast_fp16")];
+            tensor<int32, [3]> var_2810_begin_0 = const()[name = tensor<string, []>("op_2810_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_2810_end_0 = const()[name = tensor<string, []>("op_2810_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_2810_end_mask_0 = const()[name = tensor<string, []>("op_2810_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 69, 512]> var_2810_cast_fp16 = slice_by_index(begin = var_2810_begin_0, end = var_2810_end_0, end_mask = var_2810_end_mask_0, x = cache_49_cast_fp16)[name = tensor<string, []>("op_2810_cast_fp16")];
+            tensor<int32, [3]> var_2813_begin_0 = const()[name = tensor<string, []>("op_2813_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_2813_end_0 = const()[name = tensor<string, []>("op_2813_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<bool, [3]> var_2813_end_mask_0 = const()[name = tensor<string, []>("op_2813_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 1, 512]> var_2813_cast_fp16 = slice_by_index(begin = var_2813_begin_0, end = var_2813_end_0, end_mask = var_2813_end_mask_0, x = key_25_cast_fp16)[name = tensor<string, []>("op_2813_cast_fp16")];
+            tensor<bool, []> var_2816_interleave_0 = const()[name = tensor<string, []>("op_2816_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_2816_cast_fp16 = concat(axis = var_64, interleave = var_2816_interleave_0, values = (var_2810_cast_fp16, var_2813_cast_fp16))[name = tensor<string, []>("op_2816_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_12_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155728960)))];
+            tensor<fp16, [1, 3, 512]> linear_111_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16, x = key_25_cast_fp16)[name = tensor<string, []>("linear_111_cast_fp16")];
+            tensor<int32, [4]> var_2820 = const()[name = tensor<string, []>("op_2820"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 3, 8, 64]> q_73_cast_fp16 = reshape(shape = var_2820, x = linear_111_cast_fp16)[name = tensor<string, []>("q_73_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_12_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156253312)))];
+            tensor<fp16, [1, 73, 512]> linear_112_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16, x = input_663_cast_fp16)[name = tensor<string, []>("linear_112_cast_fp16")];
+            tensor<int32, [4]> var_2824 = const()[name = tensor<string, []>("op_2824"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> k_49_cast_fp16 = reshape(shape = var_2824, x = linear_112_cast_fp16)[name = tensor<string, []>("k_49_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_12_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156777664)))];
+            tensor<fp16, [1, 73, 512]> linear_113_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16, x = input_663_cast_fp16)[name = tensor<string, []>("linear_113_cast_fp16")];
+            tensor<int32, [4]> var_2828 = const()[name = tensor<string, []>("op_2828"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> v_25_cast_fp16 = reshape(shape = var_2828, x = linear_113_cast_fp16)[name = tensor<string, []>("v_25_cast_fp16")];
+            tensor<int32, [4]> value_27_perm_0 = const()[name = tensor<string, []>("value_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157302016)))];
+            tensor<fp16, [1, 3, 8, 64]> var_2840_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2840_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157303104)))];
+            tensor<fp16, [1, 3, 8, 64]> var_2842_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2842_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_25_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_319_transpose_x_0 = const()[name = tensor<string, []>("x_319_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_319_transpose_y_0 = const()[name = tensor<string, []>("x_319_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 145]> var_2844_to_fp16 = const()[name = tensor<string, []>("op_2844_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157304192)))];
+            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_2842_cast_fp16)[name = tensor<string, []>("transpose_129")];
+            tensor<fp16, [1, 8, 3, 145]> x_319_cast_fp16 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = var_2844_to_fp16)[name = tensor<string, []>("x_319_cast_fp16")];
+            tensor<int32, [8]> x_321_pad_0 = const()[name = tensor<string, []>("x_321_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_321_mode_0 = const()[name = tensor<string, []>("x_321_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_179_to_fp16 = const()[name = tensor<string, []>("const_179_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 3, 146]> x_321_cast_fp16 = pad(constant_val = const_179_to_fp16, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319_cast_fp16)[name = tensor<string, []>("x_321_cast_fp16")];
+            tensor<int32, [4]> var_2852 = const()[name = tensor<string, []>("op_2852"), val = tensor<int32, [4]>([1, 8, -1, 3])];
+            tensor<fp16, [1, 8, 146, 3]> x_323_cast_fp16 = reshape(shape = var_2852, x = x_321_cast_fp16)[name = tensor<string, []>("x_323_cast_fp16")];
+            tensor<int32, [4]> var_2856_begin_0 = const()[name = tensor<string, []>("op_2856_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2856_end_0 = const()[name = tensor<string, []>("op_2856_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
+            tensor<bool, [4]> var_2856_end_mask_0 = const()[name = tensor<string, []>("op_2856_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 145, 3]> var_2856_cast_fp16 = slice_by_index(begin = var_2856_begin_0, end = var_2856_end_0, end_mask = var_2856_end_mask_0, x = x_323_cast_fp16)[name = tensor<string, []>("op_2856_cast_fp16")];
+            tensor<int32, [4]> var_2857 = const()[name = tensor<string, []>("op_2857"), val = tensor<int32, [4]>([1, 8, 3, 145])];
+            tensor<fp16, [1, 8, 3, 145]> matrix_bd_49_cast_fp16 = reshape(shape = var_2857, x = var_2856_cast_fp16)[name = tensor<string, []>("matrix_bd_49_cast_fp16")];
+            tensor<bool, []> matrix_ac_25_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_25_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_25_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_25_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_75_perm_0 = const()[name = tensor<string, []>("transpose_75_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_76_perm_0 = const()[name = tensor<string, []>("transpose_76_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 73]> transpose_76 = transpose(perm = transpose_76_perm_0, x = k_49_cast_fp16)[name = tensor<string, []>("transpose_127")];
+            tensor<fp16, [1, 8, 3, 64]> transpose_75 = transpose(perm = transpose_75_perm_0, x = var_2840_cast_fp16)[name = tensor<string, []>("transpose_128")];
+            tensor<fp16, [1, 8, 3, 73]> matrix_ac_25_cast_fp16 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_75, y = transpose_76)[name = tensor<string, []>("matrix_ac_25_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_51_begin_0 = const()[name = tensor<string, []>("matrix_bd_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_51_end_0 = const()[name = tensor<string, []>("matrix_bd_51_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
+            tensor<bool, [4]> matrix_bd_51_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 3, 73]> matrix_bd_51_cast_fp16 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49_cast_fp16)[name = tensor<string, []>("matrix_bd_51_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_2866_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = tensor<string, []>("op_2866_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_49_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_49_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_49_cast_fp16 = mul(x = var_2866_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_49_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> scores_51_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_51_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_2872_cast_fp16 = softmax(axis = var_62, x = scores_51_cast_fp16)[name = tensor<string, []>("op_2872_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> input_665_cast_fp16 = select(a = var_40_to_fp16, b = var_2872_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_665_cast_fp16")];
+            tensor<bool, []> x_325_transpose_x_0 = const()[name = tensor<string, []>("x_325_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_325_transpose_y_0 = const()[name = tensor<string, []>("x_325_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 73, 64]> value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_25_cast_fp16)[name = tensor<string, []>("transpose_130")];
+            tensor<fp16, [1, 8, 3, 64]> x_325_cast_fp16 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665_cast_fp16, y = value_27_cast_fp16)[name = tensor<string, []>("x_325_cast_fp16")];
+            tensor<int32, [4]> var_2876_perm_0 = const()[name = tensor<string, []>("op_2876_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2877 = const()[name = tensor<string, []>("op_2877"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 3, 8, 64]> var_2876_cast_fp16 = transpose(perm = var_2876_perm_0, x = x_325_cast_fp16)[name = tensor<string, []>("transpose_126")];
+            tensor<fp16, [1, 3, 512]> input_667_cast_fp16 = reshape(shape = var_2877, x = var_2876_cast_fp16)[name = tensor<string, []>("input_667_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_12_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157452736)))];
+            tensor<fp16, [1, 3, 512]> linear_115_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16, x = input_667_cast_fp16)[name = tensor<string, []>("linear_115_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = tensor<string, []>("input_671_cast_fp16")];
+            tensor<int32, [1]> x_329_axes_0 = const()[name = tensor<string, []>("x_329_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157977088)))];
+            tensor<fp16, [512]> encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157978176)))];
+            tensor<fp16, [1, 3, 512]> x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = tensor<string, []>("x_329_cast_fp16")];
+            tensor<int32, [3]> input_673_perm_0 = const()[name = tensor<string, []>("input_673_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_675_pad_type_0 = const()[name = tensor<string, []>("input_675_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_675_strides_0 = const()[name = tensor<string, []>("input_675_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_675_pad_0 = const()[name = tensor<string, []>("input_675_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_675_dilations_0 = const()[name = tensor<string, []>("input_675_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_675_groups_0 = const()[name = tensor<string, []>("input_675_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_12_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157979264)))];
+            tensor<fp16, [1, 512, 3]> input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = tensor<string, []>("transpose_125")];
+            tensor<fp16, [1, 1024, 3]> input_675_cast_fp16 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16, x = input_673_cast_fp16)[name = tensor<string, []>("input_675_cast_fp16")];
+            tensor<int32, []> x_331_split_num_splits_0 = const()[name = tensor<string, []>("x_331_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_331_split_axis_0 = const()[name = tensor<string, []>("x_331_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 3]> x_331_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_331_split_cast_fp16_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675_cast_fp16)[name = tensor<string, []>("x_331_split_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = tensor<string, []>("x_331_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_331_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> input_677_cast_fp16 = select(a = var_40_to_fp16, b = x_331_cast_fp16, cond = var_418)[name = tensor<string, []>("input_677_cast_fp16")];
+            tensor<bool, []> new_x_51_interleave_0 = const()[name = tensor<string, []>("new_x_51_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 11]> new_x_51_cast_fp16 = concat(axis = var_62, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = tensor<string, []>("new_x_51_cast_fp16")];
+            tensor<int32, [3]> next_cache_25_begin_0 = const()[name = tensor<string, []>("next_cache_25_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_25_end_0 = const()[name = tensor<string, []>("next_cache_25_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> next_cache_25_end_mask_0 = const()[name = tensor<string, []>("next_cache_25_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 9]> next_cache_25_cast_fp16 = slice_by_index(begin = next_cache_25_begin_0, end = next_cache_25_end_0, end_mask = next_cache_25_end_mask_0, x = new_x_51_cast_fp16)[name = tensor<string, []>("next_cache_25_cast_fp16")];
+            tensor<int32, [3]> var_2918_begin_0 = const()[name = tensor<string, []>("op_2918_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> var_2918_end_0 = const()[name = tensor<string, []>("op_2918_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> var_2918_end_mask_0 = const()[name = tensor<string, []>("op_2918_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_2918_cast_fp16 = slice_by_index(begin = var_2918_begin_0, end = var_2918_end_0, end_mask = var_2918_end_mask_0, x = next_cache_25_cast_fp16)[name = tensor<string, []>("op_2918_cast_fp16")];
+            tensor<string, []> x_333_pad_type_0 = const()[name = tensor<string, []>("x_333_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_333_groups_0 = const()[name = tensor<string, []>("x_333_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_333_strides_0 = const()[name = tensor<string, []>("x_333_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_333_pad_0 = const()[name = tensor<string, []>("x_333_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_333_dilations_0 = const()[name = tensor<string, []>("x_333_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_12_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159027904)))];
+            tensor<fp16, [1, 512, 3]> x_333_cast_fp16 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16, x = new_x_51_cast_fp16)[name = tensor<string, []>("x_333_cast_fp16")];
+            tensor<int32, [3]> input_679_perm_0 = const()[name = tensor<string, []>("input_679_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_335_axes_0 = const()[name = tensor<string, []>("x_335_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159037184)))];
+            tensor<fp16, [512]> encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159038272)))];
+            tensor<fp16, [1, 3, 512]> input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = tensor<string, []>("transpose_124")];
+            tensor<fp16, [1, 3, 512]> x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = tensor<string, []>("x_335_cast_fp16")];
+            tensor<int32, [3]> input_681_perm_0 = const()[name = tensor<string, []>("input_681_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 3]> input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = tensor<string, []>("transpose_123")];
+            tensor<fp16, [1, 512, 3]> input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = tensor<string, []>("input_683_cast_fp16")];
+            tensor<string, []> x_337_pad_type_0 = const()[name = tensor<string, []>("x_337_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_337_strides_0 = const()[name = tensor<string, []>("x_337_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_337_pad_0 = const()[name = tensor<string, []>("x_337_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_337_dilations_0 = const()[name = tensor<string, []>("x_337_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_337_groups_0 = const()[name = tensor<string, []>("x_337_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_12_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159039360)))];
+            tensor<fp16, [1, 512, 3]> x_337_cast_fp16 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16, x = input_683_cast_fp16)[name = tensor<string, []>("x_337_cast_fp16")];
+            tensor<int32, [3]> input_685_perm_0 = const()[name = tensor<string, []>("input_685_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 3, 512]> input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = tensor<string, []>("transpose_122")];
+            tensor<fp16, [1, 3, 512]> input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = tensor<string, []>("input_687_cast_fp16")];
+            tensor<int32, [1]> input_689_axes_0 = const()[name = tensor<string, []>("input_689_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159563712)))];
+            tensor<fp16, [512]> encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159564800)))];
+            tensor<fp16, [1, 3, 512]> input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = tensor<string, []>("input_689_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_12_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159565888)))];
+            tensor<fp16, [1, 3, 2048]> linear_116_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16, x = input_689_cast_fp16)[name = tensor<string, []>("linear_116_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = tensor<string, []>("input_693_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_12_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161663104)))];
+            tensor<fp16, [1, 3, 512]> linear_117_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16, x = input_693_cast_fp16)[name = tensor<string, []>("linear_117_cast_fp16")];
+            tensor<fp16, []> var_2959_to_fp16 = const()[name = tensor<string, []>("op_2959_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_2960_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_2959_to_fp16)[name = tensor<string, []>("op_2960_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_2960_cast_fp16)[name = tensor<string, []>("input_699_cast_fp16")];
+            tensor<int32, [1]> input_701_axes_0 = const()[name = tensor<string, []>("input_701_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_12_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163760320)))];
+            tensor<fp16, [512]> encoder_layers_12_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163761408)))];
+            tensor<fp16, [1, 3, 512]> input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = tensor<string, []>("input_701_cast_fp16")];
+            tensor<int32, [4]> cache_53_begin_0 = const()[name = tensor<string, []>("cache_53_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_53_end_0 = const()[name = tensor<string, []>("cache_53_end_0"), val = tensor<int32, [4]>([14, 1, 70, 512])];
+            tensor<bool, [4]> cache_53_end_mask_0 = const()[name = tensor<string, []>("cache_53_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_53_squeeze_mask_0 = const()[name = tensor<string, []>("cache_53_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_53_cast_fp16 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_53_cast_fp16")];
+            tensor<int32, [4]> cache_55_begin_0 = const()[name = tensor<string, []>("cache_55_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_55_end_0 = const()[name = tensor<string, []>("cache_55_end_0"), val = tensor<int32, [4]>([14, 1, 512, 8])];
+            tensor<bool, [4]> cache_55_end_mask_0 = const()[name = tensor<string, []>("cache_55_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_55_squeeze_mask_0 = const()[name = tensor<string, []>("cache_55_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_55_cast_fp16 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_55_cast_fp16")];
+            tensor<int32, [1]> input_703_axes_0 = const()[name = tensor<string, []>("input_703_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163762496)))];
+            tensor<fp16, [512]> encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163763584)))];
+            tensor<fp16, [1, 3, 512]> input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = tensor<string, []>("input_703_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_13_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163764672)))];
+            tensor<fp16, [1, 3, 2048]> linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16, x = input_703_cast_fp16)[name = tensor<string, []>("linear_118_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = tensor<string, []>("input_707_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_13_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165861888)))];
+            tensor<fp16, [1, 3, 512]> linear_119_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16, x = input_707_cast_fp16)[name = tensor<string, []>("linear_119_cast_fp16")];
+            tensor<fp16, []> var_2994_to_fp16 = const()[name = tensor<string, []>("op_2994_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_2995_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_2994_to_fp16)[name = tensor<string, []>("op_2995_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_2995_cast_fp16)[name = tensor<string, []>("input_713_cast_fp16")];
+            tensor<int32, [1]> key_27_axes_0 = const()[name = tensor<string, []>("key_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167959104)))];
+            tensor<fp16, [512]> encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167960192)))];
+            tensor<fp16, [1, 3, 512]> key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = tensor<string, []>("key_27_cast_fp16")];
+            tensor<bool, []> input_715_interleave_0 = const()[name = tensor<string, []>("input_715_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 73, 512]> input_715_cast_fp16 = concat(axis = var_64, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = tensor<string, []>("input_715_cast_fp16")];
+            tensor<int32, [3]> var_3017_begin_0 = const()[name = tensor<string, []>("op_3017_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_3017_end_0 = const()[name = tensor<string, []>("op_3017_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_3017_end_mask_0 = const()[name = tensor<string, []>("op_3017_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 69, 512]> var_3017_cast_fp16 = slice_by_index(begin = var_3017_begin_0, end = var_3017_end_0, end_mask = var_3017_end_mask_0, x = cache_53_cast_fp16)[name = tensor<string, []>("op_3017_cast_fp16")];
+            tensor<int32, [3]> var_3020_begin_0 = const()[name = tensor<string, []>("op_3020_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_3020_end_0 = const()[name = tensor<string, []>("op_3020_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<bool, [3]> var_3020_end_mask_0 = const()[name = tensor<string, []>("op_3020_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 1, 512]> var_3020_cast_fp16 = slice_by_index(begin = var_3020_begin_0, end = var_3020_end_0, end_mask = var_3020_end_mask_0, x = key_27_cast_fp16)[name = tensor<string, []>("op_3020_cast_fp16")];
+            tensor<bool, []> var_3023_interleave_0 = const()[name = tensor<string, []>("op_3023_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_3023_cast_fp16 = concat(axis = var_64, interleave = var_3023_interleave_0, values = (var_3017_cast_fp16, var_3020_cast_fp16))[name = tensor<string, []>("op_3023_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_13_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167961280)))];
+            tensor<fp16, [1, 3, 512]> linear_120_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16, x = key_27_cast_fp16)[name = tensor<string, []>("linear_120_cast_fp16")];
+            tensor<int32, [4]> var_3027 = const()[name = tensor<string, []>("op_3027"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 3, 8, 64]> q_79_cast_fp16 = reshape(shape = var_3027, x = linear_120_cast_fp16)[name = tensor<string, []>("q_79_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_13_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168485632)))];
+            tensor<fp16, [1, 73, 512]> linear_121_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16, x = input_715_cast_fp16)[name = tensor<string, []>("linear_121_cast_fp16")];
+            tensor<int32, [4]> var_3031 = const()[name = tensor<string, []>("op_3031"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> k_53_cast_fp16 = reshape(shape = var_3031, x = linear_121_cast_fp16)[name = tensor<string, []>("k_53_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_13_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169009984)))];
+            tensor<fp16, [1, 73, 512]> linear_122_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16, x = input_715_cast_fp16)[name = tensor<string, []>("linear_122_cast_fp16")];
+            tensor<int32, [4]> var_3035 = const()[name = tensor<string, []>("op_3035"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> v_27_cast_fp16 = reshape(shape = var_3035, x = linear_122_cast_fp16)[name = tensor<string, []>("v_27_cast_fp16")];
+            tensor<int32, [4]> value_29_perm_0 = const()[name = tensor<string, []>("value_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169534336)))];
+            tensor<fp16, [1, 3, 8, 64]> var_3047_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3047_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169535424)))];
+            tensor<fp16, [1, 3, 8, 64]> var_3049_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3049_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_27_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_345_transpose_x_0 = const()[name = tensor<string, []>("x_345_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_345_transpose_y_0 = const()[name = tensor<string, []>("x_345_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 145]> var_3051_to_fp16 = const()[name = tensor<string, []>("op_3051_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169536512)))];
+            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3049_cast_fp16)[name = tensor<string, []>("transpose_120")];
+            tensor<fp16, [1, 8, 3, 145]> x_345_cast_fp16 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = var_3051_to_fp16)[name = tensor<string, []>("x_345_cast_fp16")];
+            tensor<int32, [8]> x_347_pad_0 = const()[name = tensor<string, []>("x_347_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_347_mode_0 = const()[name = tensor<string, []>("x_347_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_192_to_fp16 = const()[name = tensor<string, []>("const_192_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 3, 146]> x_347_cast_fp16 = pad(constant_val = const_192_to_fp16, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345_cast_fp16)[name = tensor<string, []>("x_347_cast_fp16")];
+            tensor<int32, [4]> var_3059 = const()[name = tensor<string, []>("op_3059"), val = tensor<int32, [4]>([1, 8, -1, 3])];
+            tensor<fp16, [1, 8, 146, 3]> x_349_cast_fp16 = reshape(shape = var_3059, x = x_347_cast_fp16)[name = tensor<string, []>("x_349_cast_fp16")];
+            tensor<int32, [4]> var_3063_begin_0 = const()[name = tensor<string, []>("op_3063_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3063_end_0 = const()[name = tensor<string, []>("op_3063_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
+            tensor<bool, [4]> var_3063_end_mask_0 = const()[name = tensor<string, []>("op_3063_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 145, 3]> var_3063_cast_fp16 = slice_by_index(begin = var_3063_begin_0, end = var_3063_end_0, end_mask = var_3063_end_mask_0, x = x_349_cast_fp16)[name = tensor<string, []>("op_3063_cast_fp16")];
+            tensor<int32, [4]> var_3064 = const()[name = tensor<string, []>("op_3064"), val = tensor<int32, [4]>([1, 8, 3, 145])];
+            tensor<fp16, [1, 8, 3, 145]> matrix_bd_53_cast_fp16 = reshape(shape = var_3064, x = var_3063_cast_fp16)[name = tensor<string, []>("matrix_bd_53_cast_fp16")];
+            tensor<bool, []> matrix_ac_27_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_27_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_27_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_27_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_77_perm_0 = const()[name = tensor<string, []>("transpose_77_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_78_perm_0 = const()[name = tensor<string, []>("transpose_78_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 73]> transpose_78 = transpose(perm = transpose_78_perm_0, x = k_53_cast_fp16)[name = tensor<string, []>("transpose_118")];
+            tensor<fp16, [1, 8, 3, 64]> transpose_77 = transpose(perm = transpose_77_perm_0, x = var_3047_cast_fp16)[name = tensor<string, []>("transpose_119")];
+            tensor<fp16, [1, 8, 3, 73]> matrix_ac_27_cast_fp16 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_77, y = transpose_78)[name = tensor<string, []>("matrix_ac_27_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_55_begin_0 = const()[name = tensor<string, []>("matrix_bd_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_55_end_0 = const()[name = tensor<string, []>("matrix_bd_55_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
+            tensor<bool, [4]> matrix_bd_55_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 3, 73]> matrix_bd_55_cast_fp16 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53_cast_fp16)[name = tensor<string, []>("matrix_bd_55_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_3073_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = tensor<string, []>("op_3073_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_53_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_53_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_53_cast_fp16 = mul(x = var_3073_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_53_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> scores_55_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_55_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_3079_cast_fp16 = softmax(axis = var_62, x = scores_55_cast_fp16)[name = tensor<string, []>("op_3079_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> input_717_cast_fp16 = select(a = var_40_to_fp16, b = var_3079_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_717_cast_fp16")];
+            tensor<bool, []> x_351_transpose_x_0 = const()[name = tensor<string, []>("x_351_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_351_transpose_y_0 = const()[name = tensor<string, []>("x_351_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 73, 64]> value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_27_cast_fp16)[name = tensor<string, []>("transpose_121")];
+            tensor<fp16, [1, 8, 3, 64]> x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717_cast_fp16, y = value_29_cast_fp16)[name = tensor<string, []>("x_351_cast_fp16")];
+            tensor<int32, [4]> var_3083_perm_0 = const()[name = tensor<string, []>("op_3083_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3084 = const()[name = tensor<string, []>("op_3084"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 3, 8, 64]> var_3083_cast_fp16 = transpose(perm = var_3083_perm_0, x = x_351_cast_fp16)[name = tensor<string, []>("transpose_117")];
+            tensor<fp16, [1, 3, 512]> input_719_cast_fp16 = reshape(shape = var_3084, x = var_3083_cast_fp16)[name = tensor<string, []>("input_719_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_13_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169685056)))];
+            tensor<fp16, [1, 3, 512]> linear_124_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16, x = input_719_cast_fp16)[name = tensor<string, []>("linear_124_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = tensor<string, []>("input_723_cast_fp16")];
+            tensor<int32, [1]> x_355_axes_0 = const()[name = tensor<string, []>("x_355_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170209408)))];
+            tensor<fp16, [512]> encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170210496)))];
+            tensor<fp16, [1, 3, 512]> x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = tensor<string, []>("x_355_cast_fp16")];
+            tensor<int32, [3]> input_725_perm_0 = const()[name = tensor<string, []>("input_725_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_727_pad_type_0 = const()[name = tensor<string, []>("input_727_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_727_strides_0 = const()[name = tensor<string, []>("input_727_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_727_pad_0 = const()[name = tensor<string, []>("input_727_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_727_dilations_0 = const()[name = tensor<string, []>("input_727_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_727_groups_0 = const()[name = tensor<string, []>("input_727_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_13_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170211584)))];
+            tensor<fp16, [1, 512, 3]> input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = tensor<string, []>("transpose_116")];
+            tensor<fp16, [1, 1024, 3]> input_727_cast_fp16 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16, x = input_725_cast_fp16)[name = tensor<string, []>("input_727_cast_fp16")];
+            tensor<int32, []> x_357_split_num_splits_0 = const()[name = tensor<string, []>("x_357_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_357_split_axis_0 = const()[name = tensor<string, []>("x_357_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 3]> x_357_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727_cast_fp16)[name = tensor<string, []>("x_357_split_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = tensor<string, []>("x_357_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_357_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> input_729_cast_fp16 = select(a = var_40_to_fp16, b = x_357_cast_fp16, cond = var_418)[name = tensor<string, []>("input_729_cast_fp16")];
+            tensor<bool, []> new_x_55_interleave_0 = const()[name = tensor<string, []>("new_x_55_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 11]> new_x_55_cast_fp16 = concat(axis = var_62, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = tensor<string, []>("new_x_55_cast_fp16")];
+            tensor<int32, [3]> next_cache_27_begin_0 = const()[name = tensor<string, []>("next_cache_27_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_27_end_0 = const()[name = tensor<string, []>("next_cache_27_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> next_cache_27_end_mask_0 = const()[name = tensor<string, []>("next_cache_27_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 9]> next_cache_27_cast_fp16 = slice_by_index(begin = next_cache_27_begin_0, end = next_cache_27_end_0, end_mask = next_cache_27_end_mask_0, x = new_x_55_cast_fp16)[name = tensor<string, []>("next_cache_27_cast_fp16")];
+            tensor<int32, [3]> var_3125_begin_0 = const()[name = tensor<string, []>("op_3125_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> var_3125_end_0 = const()[name = tensor<string, []>("op_3125_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> var_3125_end_mask_0 = const()[name = tensor<string, []>("op_3125_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_3125_cast_fp16 = slice_by_index(begin = var_3125_begin_0, end = var_3125_end_0, end_mask = var_3125_end_mask_0, x = next_cache_27_cast_fp16)[name = tensor<string, []>("op_3125_cast_fp16")];
+            tensor<string, []> x_359_pad_type_0 = const()[name = tensor<string, []>("x_359_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_359_groups_0 = const()[name = tensor<string, []>("x_359_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_359_strides_0 = const()[name = tensor<string, []>("x_359_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_359_pad_0 = const()[name = tensor<string, []>("x_359_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_359_dilations_0 = const()[name = tensor<string, []>("x_359_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_13_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171260224)))];
+            tensor<fp16, [1, 512, 3]> x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16, x = new_x_55_cast_fp16)[name = tensor<string, []>("x_359_cast_fp16")];
+            tensor<int32, [3]> input_731_perm_0 = const()[name = tensor<string, []>("input_731_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_361_axes_0 = const()[name = tensor<string, []>("x_361_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171269504)))];
+            tensor<fp16, [512]> encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171270592)))];
+            tensor<fp16, [1, 3, 512]> input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = tensor<string, []>("transpose_115")];
+            tensor<fp16, [1, 3, 512]> x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = tensor<string, []>("x_361_cast_fp16")];
+            tensor<int32, [3]> input_733_perm_0 = const()[name = tensor<string, []>("input_733_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 3]> input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = tensor<string, []>("transpose_114")];
+            tensor<fp16, [1, 512, 3]> input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = tensor<string, []>("input_735_cast_fp16")];
+            tensor<string, []> x_363_pad_type_0 = const()[name = tensor<string, []>("x_363_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_363_strides_0 = const()[name = tensor<string, []>("x_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_363_pad_0 = const()[name = tensor<string, []>("x_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_363_dilations_0 = const()[name = tensor<string, []>("x_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_363_groups_0 = const()[name = tensor<string, []>("x_363_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_13_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171271680)))];
+            tensor<fp16, [1, 512, 3]> x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16, x = input_735_cast_fp16)[name = tensor<string, []>("x_363_cast_fp16")];
+            tensor<int32, [3]> input_737_perm_0 = const()[name = tensor<string, []>("input_737_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 3, 512]> input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = tensor<string, []>("transpose_113")];
+            tensor<fp16, [1, 3, 512]> input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = tensor<string, []>("input_739_cast_fp16")];
+            tensor<int32, [1]> input_741_axes_0 = const()[name = tensor<string, []>("input_741_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171796032)))];
+            tensor<fp16, [512]> encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171797120)))];
+            tensor<fp16, [1, 3, 512]> input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = tensor<string, []>("input_741_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_13_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171798208)))];
+            tensor<fp16, [1, 3, 2048]> linear_125_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16, x = input_741_cast_fp16)[name = tensor<string, []>("linear_125_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = tensor<string, []>("input_745_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_13_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173895424)))];
+            tensor<fp16, [1, 3, 512]> linear_126_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16, x = input_745_cast_fp16)[name = tensor<string, []>("linear_126_cast_fp16")];
+            tensor<fp16, []> var_3166_to_fp16 = const()[name = tensor<string, []>("op_3166_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_3167_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3166_to_fp16)[name = tensor<string, []>("op_3167_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3167_cast_fp16)[name = tensor<string, []>("input_751_cast_fp16")];
+            tensor<int32, [1]> input_753_axes_0 = const()[name = tensor<string, []>("input_753_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_13_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175992640)))];
+            tensor<fp16, [512]> encoder_layers_13_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175993728)))];
+            tensor<fp16, [1, 3, 512]> input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = tensor<string, []>("input_753_cast_fp16")];
+            tensor<int32, [4]> cache_57_begin_0 = const()[name = tensor<string, []>("cache_57_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_57_end_0 = const()[name = tensor<string, []>("cache_57_end_0"), val = tensor<int32, [4]>([15, 1, 70, 512])];
+            tensor<bool, [4]> cache_57_end_mask_0 = const()[name = tensor<string, []>("cache_57_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_57_squeeze_mask_0 = const()[name = tensor<string, []>("cache_57_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_57_cast_fp16 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_57_cast_fp16")];
+            tensor<int32, [4]> cache_59_begin_0 = const()[name = tensor<string, []>("cache_59_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_59_end_0 = const()[name = tensor<string, []>("cache_59_end_0"), val = tensor<int32, [4]>([15, 1, 512, 8])];
+            tensor<bool, [4]> cache_59_end_mask_0 = const()[name = tensor<string, []>("cache_59_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_59_squeeze_mask_0 = const()[name = tensor<string, []>("cache_59_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_59_cast_fp16 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_59_cast_fp16")];
+            tensor<int32, [1]> input_755_axes_0 = const()[name = tensor<string, []>("input_755_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175994816)))];
+            tensor<fp16, [512]> encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175995904)))];
+            tensor<fp16, [1, 3, 512]> input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = tensor<string, []>("input_755_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_14_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175996992)))];
+            tensor<fp16, [1, 3, 2048]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16, x = input_755_cast_fp16)[name = tensor<string, []>("linear_127_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = tensor<string, []>("input_759_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_14_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178094208)))];
+            tensor<fp16, [1, 3, 512]> linear_128_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16, x = input_759_cast_fp16)[name = tensor<string, []>("linear_128_cast_fp16")];
+            tensor<fp16, []> var_3201_to_fp16 = const()[name = tensor<string, []>("op_3201_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_3202_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3201_to_fp16)[name = tensor<string, []>("op_3202_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3202_cast_fp16)[name = tensor<string, []>("input_765_cast_fp16")];
+            tensor<int32, [1]> key_29_axes_0 = const()[name = tensor<string, []>("key_29_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180191424)))];
+            tensor<fp16, [512]> encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180192512)))];
+            tensor<fp16, [1, 3, 512]> key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = tensor<string, []>("key_29_cast_fp16")];
+            tensor<bool, []> input_767_interleave_0 = const()[name = tensor<string, []>("input_767_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 73, 512]> input_767_cast_fp16 = concat(axis = var_64, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = tensor<string, []>("input_767_cast_fp16")];
+            tensor<int32, [3]> var_3224_begin_0 = const()[name = tensor<string, []>("op_3224_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_3224_end_0 = const()[name = tensor<string, []>("op_3224_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_3224_end_mask_0 = const()[name = tensor<string, []>("op_3224_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 69, 512]> var_3224_cast_fp16 = slice_by_index(begin = var_3224_begin_0, end = var_3224_end_0, end_mask = var_3224_end_mask_0, x = cache_57_cast_fp16)[name = tensor<string, []>("op_3224_cast_fp16")];
+            tensor<int32, [3]> var_3227_begin_0 = const()[name = tensor<string, []>("op_3227_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_3227_end_0 = const()[name = tensor<string, []>("op_3227_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<bool, [3]> var_3227_end_mask_0 = const()[name = tensor<string, []>("op_3227_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 1, 512]> var_3227_cast_fp16 = slice_by_index(begin = var_3227_begin_0, end = var_3227_end_0, end_mask = var_3227_end_mask_0, x = key_29_cast_fp16)[name = tensor<string, []>("op_3227_cast_fp16")];
+            tensor<bool, []> var_3230_interleave_0 = const()[name = tensor<string, []>("op_3230_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_3230_cast_fp16 = concat(axis = var_64, interleave = var_3230_interleave_0, values = (var_3224_cast_fp16, var_3227_cast_fp16))[name = tensor<string, []>("op_3230_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_14_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180193600)))];
+            tensor<fp16, [1, 3, 512]> linear_129_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16, x = key_29_cast_fp16)[name = tensor<string, []>("linear_129_cast_fp16")];
+            tensor<int32, [4]> var_3234 = const()[name = tensor<string, []>("op_3234"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 3, 8, 64]> q_85_cast_fp16 = reshape(shape = var_3234, x = linear_129_cast_fp16)[name = tensor<string, []>("q_85_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_14_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180717952)))];
+            tensor<fp16, [1, 73, 512]> linear_130_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16, x = input_767_cast_fp16)[name = tensor<string, []>("linear_130_cast_fp16")];
+            tensor<int32, [4]> var_3238 = const()[name = tensor<string, []>("op_3238"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> k_57_cast_fp16 = reshape(shape = var_3238, x = linear_130_cast_fp16)[name = tensor<string, []>("k_57_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_14_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181242304)))];
+            tensor<fp16, [1, 73, 512]> linear_131_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16, x = input_767_cast_fp16)[name = tensor<string, []>("linear_131_cast_fp16")];
+            tensor<int32, [4]> var_3242 = const()[name = tensor<string, []>("op_3242"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> v_29_cast_fp16 = reshape(shape = var_3242, x = linear_131_cast_fp16)[name = tensor<string, []>("v_29_cast_fp16")];
+            tensor<int32, [4]> value_31_perm_0 = const()[name = tensor<string, []>("value_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181766656)))];
+            tensor<fp16, [1, 3, 8, 64]> var_3254_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3254_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181767744)))];
+            tensor<fp16, [1, 3, 8, 64]> var_3256_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3256_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_29_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_371_transpose_x_0 = const()[name = tensor<string, []>("x_371_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_371_transpose_y_0 = const()[name = tensor<string, []>("x_371_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 145]> var_3258_to_fp16 = const()[name = tensor<string, []>("op_3258_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181768832)))];
+            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3256_cast_fp16)[name = tensor<string, []>("transpose_111")];
+            tensor<fp16, [1, 8, 3, 145]> x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = var_3258_to_fp16)[name = tensor<string, []>("x_371_cast_fp16")];
+            tensor<int32, [8]> x_373_pad_0 = const()[name = tensor<string, []>("x_373_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_373_mode_0 = const()[name = tensor<string, []>("x_373_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_205_to_fp16 = const()[name = tensor<string, []>("const_205_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 3, 146]> x_373_cast_fp16 = pad(constant_val = const_205_to_fp16, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371_cast_fp16)[name = tensor<string, []>("x_373_cast_fp16")];
+            tensor<int32, [4]> var_3266 = const()[name = tensor<string, []>("op_3266"), val = tensor<int32, [4]>([1, 8, -1, 3])];
+            tensor<fp16, [1, 8, 146, 3]> x_375_cast_fp16 = reshape(shape = var_3266, x = x_373_cast_fp16)[name = tensor<string, []>("x_375_cast_fp16")];
+            tensor<int32, [4]> var_3270_begin_0 = const()[name = tensor<string, []>("op_3270_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3270_end_0 = const()[name = tensor<string, []>("op_3270_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
+            tensor<bool, [4]> var_3270_end_mask_0 = const()[name = tensor<string, []>("op_3270_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 145, 3]> var_3270_cast_fp16 = slice_by_index(begin = var_3270_begin_0, end = var_3270_end_0, end_mask = var_3270_end_mask_0, x = x_375_cast_fp16)[name = tensor<string, []>("op_3270_cast_fp16")];
+            tensor<int32, [4]> var_3271 = const()[name = tensor<string, []>("op_3271"), val = tensor<int32, [4]>([1, 8, 3, 145])];
+            tensor<fp16, [1, 8, 3, 145]> matrix_bd_57_cast_fp16 = reshape(shape = var_3271, x = var_3270_cast_fp16)[name = tensor<string, []>("matrix_bd_57_cast_fp16")];
+            tensor<bool, []> matrix_ac_29_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_29_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_29_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_29_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_79_perm_0 = const()[name = tensor<string, []>("transpose_79_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_80_perm_0 = const()[name = tensor<string, []>("transpose_80_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 73]> transpose_80 = transpose(perm = transpose_80_perm_0, x = k_57_cast_fp16)[name = tensor<string, []>("transpose_109")];
+            tensor<fp16, [1, 8, 3, 64]> transpose_79 = transpose(perm = transpose_79_perm_0, x = var_3254_cast_fp16)[name = tensor<string, []>("transpose_110")];
+            tensor<fp16, [1, 8, 3, 73]> matrix_ac_29_cast_fp16 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_79, y = transpose_80)[name = tensor<string, []>("matrix_ac_29_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_59_begin_0 = const()[name = tensor<string, []>("matrix_bd_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_59_end_0 = const()[name = tensor<string, []>("matrix_bd_59_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
+            tensor<bool, [4]> matrix_bd_59_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 3, 73]> matrix_bd_59_cast_fp16 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57_cast_fp16)[name = tensor<string, []>("matrix_bd_59_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_3280_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = tensor<string, []>("op_3280_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_57_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_57_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_57_cast_fp16 = mul(x = var_3280_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_57_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> scores_59_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_59_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_3286_cast_fp16 = softmax(axis = var_62, x = scores_59_cast_fp16)[name = tensor<string, []>("op_3286_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> input_769_cast_fp16 = select(a = var_40_to_fp16, b = var_3286_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_769_cast_fp16")];
+            tensor<bool, []> x_377_transpose_x_0 = const()[name = tensor<string, []>("x_377_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_377_transpose_y_0 = const()[name = tensor<string, []>("x_377_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 73, 64]> value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_29_cast_fp16)[name = tensor<string, []>("transpose_112")];
+            tensor<fp16, [1, 8, 3, 64]> x_377_cast_fp16 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769_cast_fp16, y = value_31_cast_fp16)[name = tensor<string, []>("x_377_cast_fp16")];
+            tensor<int32, [4]> var_3290_perm_0 = const()[name = tensor<string, []>("op_3290_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3291 = const()[name = tensor<string, []>("op_3291"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 3, 8, 64]> var_3290_cast_fp16 = transpose(perm = var_3290_perm_0, x = x_377_cast_fp16)[name = tensor<string, []>("transpose_108")];
+            tensor<fp16, [1, 3, 512]> input_771_cast_fp16 = reshape(shape = var_3291, x = var_3290_cast_fp16)[name = tensor<string, []>("input_771_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_14_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181917376)))];
+            tensor<fp16, [1, 3, 512]> linear_133_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16, x = input_771_cast_fp16)[name = tensor<string, []>("linear_133_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = tensor<string, []>("input_775_cast_fp16")];
+            tensor<int32, [1]> x_381_axes_0 = const()[name = tensor<string, []>("x_381_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182441728)))];
+            tensor<fp16, [512]> encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182442816)))];
+            tensor<fp16, [1, 3, 512]> x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = tensor<string, []>("x_381_cast_fp16")];
+            tensor<int32, [3]> input_777_perm_0 = const()[name = tensor<string, []>("input_777_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_779_pad_type_0 = const()[name = tensor<string, []>("input_779_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_779_strides_0 = const()[name = tensor<string, []>("input_779_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_779_pad_0 = const()[name = tensor<string, []>("input_779_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_779_dilations_0 = const()[name = tensor<string, []>("input_779_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_779_groups_0 = const()[name = tensor<string, []>("input_779_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_14_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182443904)))];
+            tensor<fp16, [1, 512, 3]> input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = tensor<string, []>("transpose_107")];
+            tensor<fp16, [1, 1024, 3]> input_779_cast_fp16 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16, x = input_777_cast_fp16)[name = tensor<string, []>("input_779_cast_fp16")];
+            tensor<int32, []> x_383_split_num_splits_0 = const()[name = tensor<string, []>("x_383_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_383_split_axis_0 = const()[name = tensor<string, []>("x_383_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 3]> x_383_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_383_split_cast_fp16_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779_cast_fp16)[name = tensor<string, []>("x_383_split_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = tensor<string, []>("x_383_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_383_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> input_781_cast_fp16 = select(a = var_40_to_fp16, b = x_383_cast_fp16, cond = var_418)[name = tensor<string, []>("input_781_cast_fp16")];
+            tensor<bool, []> new_x_59_interleave_0 = const()[name = tensor<string, []>("new_x_59_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 11]> new_x_59_cast_fp16 = concat(axis = var_62, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = tensor<string, []>("new_x_59_cast_fp16")];
+            tensor<int32, [3]> next_cache_29_begin_0 = const()[name = tensor<string, []>("next_cache_29_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_29_end_0 = const()[name = tensor<string, []>("next_cache_29_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> next_cache_29_end_mask_0 = const()[name = tensor<string, []>("next_cache_29_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 9]> next_cache_29_cast_fp16 = slice_by_index(begin = next_cache_29_begin_0, end = next_cache_29_end_0, end_mask = next_cache_29_end_mask_0, x = new_x_59_cast_fp16)[name = tensor<string, []>("next_cache_29_cast_fp16")];
+            tensor<int32, [3]> var_3332_begin_0 = const()[name = tensor<string, []>("op_3332_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> var_3332_end_0 = const()[name = tensor<string, []>("op_3332_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> var_3332_end_mask_0 = const()[name = tensor<string, []>("op_3332_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_3332_cast_fp16 = slice_by_index(begin = var_3332_begin_0, end = var_3332_end_0, end_mask = var_3332_end_mask_0, x = next_cache_29_cast_fp16)[name = tensor<string, []>("op_3332_cast_fp16")];
+            tensor<string, []> x_385_pad_type_0 = const()[name = tensor<string, []>("x_385_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_385_groups_0 = const()[name = tensor<string, []>("x_385_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_385_strides_0 = const()[name = tensor<string, []>("x_385_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_385_pad_0 = const()[name = tensor<string, []>("x_385_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_385_dilations_0 = const()[name = tensor<string, []>("x_385_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_14_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183492544)))];
+            tensor<fp16, [1, 512, 3]> x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16, x = new_x_59_cast_fp16)[name = tensor<string, []>("x_385_cast_fp16")];
+            tensor<int32, [3]> input_783_perm_0 = const()[name = tensor<string, []>("input_783_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_387_axes_0 = const()[name = tensor<string, []>("x_387_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183501824)))];
+            tensor<fp16, [512]> encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183502912)))];
+            tensor<fp16, [1, 3, 512]> input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = tensor<string, []>("transpose_106")];
+            tensor<fp16, [1, 3, 512]> x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = tensor<string, []>("x_387_cast_fp16")];
+            tensor<int32, [3]> input_785_perm_0 = const()[name = tensor<string, []>("input_785_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 3]> input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = tensor<string, []>("transpose_105")];
+            tensor<fp16, [1, 512, 3]> input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = tensor<string, []>("input_787_cast_fp16")];
+            tensor<string, []> x_389_pad_type_0 = const()[name = tensor<string, []>("x_389_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_389_strides_0 = const()[name = tensor<string, []>("x_389_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_389_pad_0 = const()[name = tensor<string, []>("x_389_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_389_dilations_0 = const()[name = tensor<string, []>("x_389_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_389_groups_0 = const()[name = tensor<string, []>("x_389_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_14_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183504000)))];
+            tensor<fp16, [1, 512, 3]> x_389_cast_fp16 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16, x = input_787_cast_fp16)[name = tensor<string, []>("x_389_cast_fp16")];
+            tensor<int32, [3]> input_789_perm_0 = const()[name = tensor<string, []>("input_789_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 3, 512]> input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = tensor<string, []>("transpose_104")];
+            tensor<fp16, [1, 3, 512]> input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = tensor<string, []>("input_791_cast_fp16")];
+            tensor<int32, [1]> input_793_axes_0 = const()[name = tensor<string, []>("input_793_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184028352)))];
+            tensor<fp16, [512]> encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184029440)))];
+            tensor<fp16, [1, 3, 512]> input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = tensor<string, []>("input_793_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_14_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184030528)))];
+            tensor<fp16, [1, 3, 2048]> linear_134_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16, x = input_793_cast_fp16)[name = tensor<string, []>("linear_134_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = tensor<string, []>("input_797_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_14_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186127744)))];
+            tensor<fp16, [1, 3, 512]> linear_135_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16, x = input_797_cast_fp16)[name = tensor<string, []>("linear_135_cast_fp16")];
+            tensor<fp16, []> var_3373_to_fp16 = const()[name = tensor<string, []>("op_3373_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_3374_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3373_to_fp16)[name = tensor<string, []>("op_3374_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3374_cast_fp16)[name = tensor<string, []>("input_803_cast_fp16")];
+            tensor<int32, [1]> input_805_axes_0 = const()[name = tensor<string, []>("input_805_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_14_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188224960)))];
+            tensor<fp16, [512]> encoder_layers_14_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188226048)))];
+            tensor<fp16, [1, 3, 512]> input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = tensor<string, []>("input_805_cast_fp16")];
+            tensor<int32, [4]> cache_61_begin_0 = const()[name = tensor<string, []>("cache_61_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_61_end_0 = const()[name = tensor<string, []>("cache_61_end_0"), val = tensor<int32, [4]>([16, 1, 70, 512])];
+            tensor<bool, [4]> cache_61_end_mask_0 = const()[name = tensor<string, []>("cache_61_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_61_squeeze_mask_0 = const()[name = tensor<string, []>("cache_61_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_61_cast_fp16 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_61_cast_fp16")];
+            tensor<int32, [4]> cache_63_begin_0 = const()[name = tensor<string, []>("cache_63_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_63_end_0 = const()[name = tensor<string, []>("cache_63_end_0"), val = tensor<int32, [4]>([16, 1, 512, 8])];
+            tensor<bool, [4]> cache_63_end_mask_0 = const()[name = tensor<string, []>("cache_63_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_63_squeeze_mask_0 = const()[name = tensor<string, []>("cache_63_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_63_cast_fp16 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_63_cast_fp16")];
+            tensor<int32, [1]> input_807_axes_0 = const()[name = tensor<string, []>("input_807_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188227136)))];
+            tensor<fp16, [512]> encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188228224)))];
+            tensor<fp16, [1, 3, 512]> input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = tensor<string, []>("input_807_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_15_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188229312)))];
+            tensor<fp16, [1, 3, 2048]> linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16, x = input_807_cast_fp16)[name = tensor<string, []>("linear_136_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = tensor<string, []>("input_811_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_15_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190326528)))];
+            tensor<fp16, [1, 3, 512]> linear_137_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16, x = input_811_cast_fp16)[name = tensor<string, []>("linear_137_cast_fp16")];
+            tensor<fp16, []> var_3408_to_fp16 = const()[name = tensor<string, []>("op_3408_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_3409_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3408_to_fp16)[name = tensor<string, []>("op_3409_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3409_cast_fp16)[name = tensor<string, []>("input_817_cast_fp16")];
+            tensor<int32, [1]> key_31_axes_0 = const()[name = tensor<string, []>("key_31_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192423744)))];
+            tensor<fp16, [512]> encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192424832)))];
+            tensor<fp16, [1, 3, 512]> key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = tensor<string, []>("key_31_cast_fp16")];
+            tensor<bool, []> input_819_interleave_0 = const()[name = tensor<string, []>("input_819_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 73, 512]> input_819_cast_fp16 = concat(axis = var_64, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = tensor<string, []>("input_819_cast_fp16")];
+            tensor<int32, [3]> var_3431_begin_0 = const()[name = tensor<string, []>("op_3431_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_3431_end_0 = const()[name = tensor<string, []>("op_3431_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_3431_end_mask_0 = const()[name = tensor<string, []>("op_3431_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 69, 512]> var_3431_cast_fp16 = slice_by_index(begin = var_3431_begin_0, end = var_3431_end_0, end_mask = var_3431_end_mask_0, x = cache_61_cast_fp16)[name = tensor<string, []>("op_3431_cast_fp16")];
+            tensor<int32, [3]> var_3434_begin_0 = const()[name = tensor<string, []>("op_3434_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_3434_end_0 = const()[name = tensor<string, []>("op_3434_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<bool, [3]> var_3434_end_mask_0 = const()[name = tensor<string, []>("op_3434_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 1, 512]> var_3434_cast_fp16 = slice_by_index(begin = var_3434_begin_0, end = var_3434_end_0, end_mask = var_3434_end_mask_0, x = key_31_cast_fp16)[name = tensor<string, []>("op_3434_cast_fp16")];
+            tensor<bool, []> var_3437_interleave_0 = const()[name = tensor<string, []>("op_3437_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_3437_cast_fp16 = concat(axis = var_64, interleave = var_3437_interleave_0, values = (var_3431_cast_fp16, var_3434_cast_fp16))[name = tensor<string, []>("op_3437_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_15_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192425920)))];
+            tensor<fp16, [1, 3, 512]> linear_138_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16, x = key_31_cast_fp16)[name = tensor<string, []>("linear_138_cast_fp16")];
+            tensor<int32, [4]> var_3441 = const()[name = tensor<string, []>("op_3441"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 3, 8, 64]> q_91_cast_fp16 = reshape(shape = var_3441, x = linear_138_cast_fp16)[name = tensor<string, []>("q_91_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_15_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192950272)))];
+            tensor<fp16, [1, 73, 512]> linear_139_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16, x = input_819_cast_fp16)[name = tensor<string, []>("linear_139_cast_fp16")];
+            tensor<int32, [4]> var_3445 = const()[name = tensor<string, []>("op_3445"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> k_61_cast_fp16 = reshape(shape = var_3445, x = linear_139_cast_fp16)[name = tensor<string, []>("k_61_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_15_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193474624)))];
+            tensor<fp16, [1, 73, 512]> linear_140_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16, x = input_819_cast_fp16)[name = tensor<string, []>("linear_140_cast_fp16")];
+            tensor<int32, [4]> var_3449 = const()[name = tensor<string, []>("op_3449"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> v_31_cast_fp16 = reshape(shape = var_3449, x = linear_140_cast_fp16)[name = tensor<string, []>("v_31_cast_fp16")];
+            tensor<int32, [4]> value_33_perm_0 = const()[name = tensor<string, []>("value_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193998976)))];
+            tensor<fp16, [1, 3, 8, 64]> var_3461_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3461_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194000064)))];
+            tensor<fp16, [1, 3, 8, 64]> var_3463_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3463_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_31_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_397_transpose_x_0 = const()[name = tensor<string, []>("x_397_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_397_transpose_y_0 = const()[name = tensor<string, []>("x_397_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 145]> var_3465_to_fp16 = const()[name = tensor<string, []>("op_3465_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194001152)))];
+            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3463_cast_fp16)[name = tensor<string, []>("transpose_102")];
+            tensor<fp16, [1, 8, 3, 145]> x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = var_3465_to_fp16)[name = tensor<string, []>("x_397_cast_fp16")];
+            tensor<int32, [8]> x_399_pad_0 = const()[name = tensor<string, []>("x_399_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_399_mode_0 = const()[name = tensor<string, []>("x_399_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_218_to_fp16 = const()[name = tensor<string, []>("const_218_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 3, 146]> x_399_cast_fp16 = pad(constant_val = const_218_to_fp16, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397_cast_fp16)[name = tensor<string, []>("x_399_cast_fp16")];
+            tensor<int32, [4]> var_3473 = const()[name = tensor<string, []>("op_3473"), val = tensor<int32, [4]>([1, 8, -1, 3])];
+            tensor<fp16, [1, 8, 146, 3]> x_401_cast_fp16 = reshape(shape = var_3473, x = x_399_cast_fp16)[name = tensor<string, []>("x_401_cast_fp16")];
+            tensor<int32, [4]> var_3477_begin_0 = const()[name = tensor<string, []>("op_3477_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3477_end_0 = const()[name = tensor<string, []>("op_3477_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
+            tensor<bool, [4]> var_3477_end_mask_0 = const()[name = tensor<string, []>("op_3477_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 145, 3]> var_3477_cast_fp16 = slice_by_index(begin = var_3477_begin_0, end = var_3477_end_0, end_mask = var_3477_end_mask_0, x = x_401_cast_fp16)[name = tensor<string, []>("op_3477_cast_fp16")];
+            tensor<int32, [4]> var_3478 = const()[name = tensor<string, []>("op_3478"), val = tensor<int32, [4]>([1, 8, 3, 145])];
+            tensor<fp16, [1, 8, 3, 145]> matrix_bd_61_cast_fp16 = reshape(shape = var_3478, x = var_3477_cast_fp16)[name = tensor<string, []>("matrix_bd_61_cast_fp16")];
+            tensor<bool, []> matrix_ac_31_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_31_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_31_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_31_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_81_perm_0 = const()[name = tensor<string, []>("transpose_81_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_82_perm_0 = const()[name = tensor<string, []>("transpose_82_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 73]> transpose_82 = transpose(perm = transpose_82_perm_0, x = k_61_cast_fp16)[name = tensor<string, []>("transpose_100")];
+            tensor<fp16, [1, 8, 3, 64]> transpose_81 = transpose(perm = transpose_81_perm_0, x = var_3461_cast_fp16)[name = tensor<string, []>("transpose_101")];
+            tensor<fp16, [1, 8, 3, 73]> matrix_ac_31_cast_fp16 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_81, y = transpose_82)[name = tensor<string, []>("matrix_ac_31_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_63_begin_0 = const()[name = tensor<string, []>("matrix_bd_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_63_end_0 = const()[name = tensor<string, []>("matrix_bd_63_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
+            tensor<bool, [4]> matrix_bd_63_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 3, 73]> matrix_bd_63_cast_fp16 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61_cast_fp16)[name = tensor<string, []>("matrix_bd_63_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_3487_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = tensor<string, []>("op_3487_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_61_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_61_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_61_cast_fp16 = mul(x = var_3487_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_61_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> scores_63_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_63_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_3493_cast_fp16 = softmax(axis = var_62, x = scores_63_cast_fp16)[name = tensor<string, []>("op_3493_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> input_821_cast_fp16 = select(a = var_40_to_fp16, b = var_3493_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_821_cast_fp16")];
+            tensor<bool, []> x_403_transpose_x_0 = const()[name = tensor<string, []>("x_403_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_403_transpose_y_0 = const()[name = tensor<string, []>("x_403_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 73, 64]> value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_31_cast_fp16)[name = tensor<string, []>("transpose_103")];
+            tensor<fp16, [1, 8, 3, 64]> x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821_cast_fp16, y = value_33_cast_fp16)[name = tensor<string, []>("x_403_cast_fp16")];
+            tensor<int32, [4]> var_3497_perm_0 = const()[name = tensor<string, []>("op_3497_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3498 = const()[name = tensor<string, []>("op_3498"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 3, 8, 64]> var_3497_cast_fp16 = transpose(perm = var_3497_perm_0, x = x_403_cast_fp16)[name = tensor<string, []>("transpose_99")];
+            tensor<fp16, [1, 3, 512]> input_823_cast_fp16 = reshape(shape = var_3498, x = var_3497_cast_fp16)[name = tensor<string, []>("input_823_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_15_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194149696)))];
+            tensor<fp16, [1, 3, 512]> linear_142_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16, x = input_823_cast_fp16)[name = tensor<string, []>("linear_142_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = tensor<string, []>("input_827_cast_fp16")];
+            tensor<int32, [1]> x_407_axes_0 = const()[name = tensor<string, []>("x_407_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194674048)))];
+            tensor<fp16, [512]> encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194675136)))];
+            tensor<fp16, [1, 3, 512]> x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = tensor<string, []>("x_407_cast_fp16")];
+            tensor<int32, [3]> input_829_perm_0 = const()[name = tensor<string, []>("input_829_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_831_pad_type_0 = const()[name = tensor<string, []>("input_831_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_831_strides_0 = const()[name = tensor<string, []>("input_831_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_831_pad_0 = const()[name = tensor<string, []>("input_831_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_831_dilations_0 = const()[name = tensor<string, []>("input_831_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_831_groups_0 = const()[name = tensor<string, []>("input_831_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_15_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194676224)))];
+            tensor<fp16, [1, 512, 3]> input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = tensor<string, []>("transpose_98")];
+            tensor<fp16, [1, 1024, 3]> input_831_cast_fp16 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16, x = input_829_cast_fp16)[name = tensor<string, []>("input_831_cast_fp16")];
+            tensor<int32, []> x_409_split_num_splits_0 = const()[name = tensor<string, []>("x_409_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_409_split_axis_0 = const()[name = tensor<string, []>("x_409_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 3]> x_409_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_409_split_cast_fp16_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831_cast_fp16)[name = tensor<string, []>("x_409_split_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = tensor<string, []>("x_409_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_409_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> input_833_cast_fp16 = select(a = var_40_to_fp16, b = x_409_cast_fp16, cond = var_418)[name = tensor<string, []>("input_833_cast_fp16")];
+            tensor<bool, []> new_x_63_interleave_0 = const()[name = tensor<string, []>("new_x_63_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 11]> new_x_63_cast_fp16 = concat(axis = var_62, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = tensor<string, []>("new_x_63_cast_fp16")];
+            tensor<int32, [3]> next_cache_31_begin_0 = const()[name = tensor<string, []>("next_cache_31_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_31_end_0 = const()[name = tensor<string, []>("next_cache_31_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> next_cache_31_end_mask_0 = const()[name = tensor<string, []>("next_cache_31_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 9]> next_cache_31_cast_fp16 = slice_by_index(begin = next_cache_31_begin_0, end = next_cache_31_end_0, end_mask = next_cache_31_end_mask_0, x = new_x_63_cast_fp16)[name = tensor<string, []>("next_cache_31_cast_fp16")];
+            tensor<int32, [3]> var_3539_begin_0 = const()[name = tensor<string, []>("op_3539_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> var_3539_end_0 = const()[name = tensor<string, []>("op_3539_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> var_3539_end_mask_0 = const()[name = tensor<string, []>("op_3539_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_3539_cast_fp16 = slice_by_index(begin = var_3539_begin_0, end = var_3539_end_0, end_mask = var_3539_end_mask_0, x = next_cache_31_cast_fp16)[name = tensor<string, []>("op_3539_cast_fp16")];
+            tensor<string, []> x_411_pad_type_0 = const()[name = tensor<string, []>("x_411_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_411_groups_0 = const()[name = tensor<string, []>("x_411_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_411_strides_0 = const()[name = tensor<string, []>("x_411_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_411_pad_0 = const()[name = tensor<string, []>("x_411_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_411_dilations_0 = const()[name = tensor<string, []>("x_411_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_15_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195724864)))];
+            tensor<fp16, [1, 512, 3]> x_411_cast_fp16 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16, x = new_x_63_cast_fp16)[name = tensor<string, []>("x_411_cast_fp16")];
+            tensor<int32, [3]> input_835_perm_0 = const()[name = tensor<string, []>("input_835_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_413_axes_0 = const()[name = tensor<string, []>("x_413_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195734144)))];
+            tensor<fp16, [512]> encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195735232)))];
+            tensor<fp16, [1, 3, 512]> input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = tensor<string, []>("transpose_97")];
+            tensor<fp16, [1, 3, 512]> x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = tensor<string, []>("x_413_cast_fp16")];
+            tensor<int32, [3]> input_837_perm_0 = const()[name = tensor<string, []>("input_837_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 3]> input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = tensor<string, []>("transpose_96")];
+            tensor<fp16, [1, 512, 3]> input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = tensor<string, []>("input_839_cast_fp16")];
+            tensor<string, []> x_415_pad_type_0 = const()[name = tensor<string, []>("x_415_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_415_strides_0 = const()[name = tensor<string, []>("x_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_415_pad_0 = const()[name = tensor<string, []>("x_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_415_dilations_0 = const()[name = tensor<string, []>("x_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_415_groups_0 = const()[name = tensor<string, []>("x_415_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_15_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195736320)))];
+            tensor<fp16, [1, 512, 3]> x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16, x = input_839_cast_fp16)[name = tensor<string, []>("x_415_cast_fp16")];
+            tensor<int32, [3]> input_841_perm_0 = const()[name = tensor<string, []>("input_841_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 3, 512]> input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = tensor<string, []>("transpose_95")];
+            tensor<fp16, [1, 3, 512]> input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = tensor<string, []>("input_843_cast_fp16")];
+            tensor<int32, [1]> input_845_axes_0 = const()[name = tensor<string, []>("input_845_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196260672)))];
+            tensor<fp16, [512]> encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196261760)))];
+            tensor<fp16, [1, 3, 512]> input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = tensor<string, []>("input_845_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_15_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196262848)))];
+            tensor<fp16, [1, 3, 2048]> linear_143_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16, x = input_845_cast_fp16)[name = tensor<string, []>("linear_143_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = tensor<string, []>("input_849_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_15_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198360064)))];
+            tensor<fp16, [1, 3, 512]> linear_144_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16, x = input_849_cast_fp16)[name = tensor<string, []>("linear_144_cast_fp16")];
+            tensor<fp16, []> var_3580_to_fp16 = const()[name = tensor<string, []>("op_3580_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_3581_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3580_to_fp16)[name = tensor<string, []>("op_3581_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3581_cast_fp16)[name = tensor<string, []>("input_855_cast_fp16")];
+            tensor<int32, [1]> input_857_axes_0 = const()[name = tensor<string, []>("input_857_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_15_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200457280)))];
+            tensor<fp16, [512]> encoder_layers_15_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200458368)))];
+            tensor<fp16, [1, 3, 512]> input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = tensor<string, []>("input_857_cast_fp16")];
+            tensor<int32, [4]> cache_65_begin_0 = const()[name = tensor<string, []>("cache_65_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_65_end_0 = const()[name = tensor<string, []>("cache_65_end_0"), val = tensor<int32, [4]>([17, 1, 70, 512])];
+            tensor<bool, [4]> cache_65_end_mask_0 = const()[name = tensor<string, []>("cache_65_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_65_squeeze_mask_0 = const()[name = tensor<string, []>("cache_65_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_65_cast_fp16 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_65_cast_fp16")];
+            tensor<int32, [4]> cache_begin_0 = const()[name = tensor<string, []>("cache_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_end_0 = const()[name = tensor<string, []>("cache_end_0"), val = tensor<int32, [4]>([17, 1, 512, 8])];
+            tensor<bool, [4]> cache_end_mask_0 = const()[name = tensor<string, []>("cache_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_squeeze_mask_0 = const()[name = tensor<string, []>("cache_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_cast_fp16 = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_cast_fp16")];
+            tensor<int32, [1]> input_859_axes_0 = const()[name = tensor<string, []>("input_859_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200459456)))];
+            tensor<fp16, [512]> encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200460544)))];
+            tensor<fp16, [1, 3, 512]> input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = tensor<string, []>("input_859_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_16_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200461632)))];
+            tensor<fp16, [1, 3, 2048]> linear_145_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16, x = input_859_cast_fp16)[name = tensor<string, []>("linear_145_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = tensor<string, []>("input_863_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_16_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202558848)))];
+            tensor<fp16, [1, 3, 512]> linear_146_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16, x = input_863_cast_fp16)[name = tensor<string, []>("linear_146_cast_fp16")];
+            tensor<fp16, []> var_3615_to_fp16 = const()[name = tensor<string, []>("op_3615_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_3616_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3615_to_fp16)[name = tensor<string, []>("op_3616_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3616_cast_fp16)[name = tensor<string, []>("input_869_cast_fp16")];
+            tensor<int32, [1]> key_axes_0 = const()[name = tensor<string, []>("key_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(204656064)))];
+            tensor<fp16, [512]> encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(204657152)))];
+            tensor<fp16, [1, 3, 512]> key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = tensor<string, []>("key_cast_fp16")];
+            tensor<bool, []> input_871_interleave_0 = const()[name = tensor<string, []>("input_871_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 73, 512]> input_871_cast_fp16 = concat(axis = var_64, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_cast_fp16))[name = tensor<string, []>("input_871_cast_fp16")];
+            tensor<int32, [3]> var_3638_begin_0 = const()[name = tensor<string, []>("op_3638_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_3638_end_0 = const()[name = tensor<string, []>("op_3638_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_3638_end_mask_0 = const()[name = tensor<string, []>("op_3638_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 69, 512]> var_3638_cast_fp16 = slice_by_index(begin = var_3638_begin_0, end = var_3638_end_0, end_mask = var_3638_end_mask_0, x = cache_65_cast_fp16)[name = tensor<string, []>("op_3638_cast_fp16")];
+            tensor<int32, [3]> var_3641_begin_0 = const()[name = tensor<string, []>("op_3641_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_3641_end_0 = const()[name = tensor<string, []>("op_3641_end_0"), val = tensor<int32, [3]>([1, 1, 512])];
+            tensor<bool, [3]> var_3641_end_mask_0 = const()[name = tensor<string, []>("op_3641_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 1, 512]> var_3641_cast_fp16 = slice_by_index(begin = var_3641_begin_0, end = var_3641_end_0, end_mask = var_3641_end_mask_0, x = key_cast_fp16)[name = tensor<string, []>("op_3641_cast_fp16")];
+            tensor<bool, []> cache_last_channel_cur_interleave_0 = const()[name = tensor<string, []>("cache_last_channel_cur_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> cache_last_channel_cur_cast_fp16 = concat(axis = var_64, interleave = cache_last_channel_cur_interleave_0, values = (var_3638_cast_fp16, var_3641_cast_fp16))[name = tensor<string, []>("cache_last_channel_cur_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_16_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(204658240)))];
+            tensor<fp16, [1, 3, 512]> linear_147_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16, x = key_cast_fp16)[name = tensor<string, []>("linear_147_cast_fp16")];
+            tensor<int32, [4]> var_3648 = const()[name = tensor<string, []>("op_3648"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 3, 8, 64]> q_97_cast_fp16 = reshape(shape = var_3648, x = linear_147_cast_fp16)[name = tensor<string, []>("q_97_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_16_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205182592)))];
+            tensor<fp16, [1, 73, 512]> linear_148_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16, x = input_871_cast_fp16)[name = tensor<string, []>("linear_148_cast_fp16")];
+            tensor<int32, [4]> var_3652 = const()[name = tensor<string, []>("op_3652"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> k_65_cast_fp16 = reshape(shape = var_3652, x = linear_148_cast_fp16)[name = tensor<string, []>("k_65_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_16_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205706944)))];
+            tensor<fp16, [1, 73, 512]> linear_149_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16, x = input_871_cast_fp16)[name = tensor<string, []>("linear_149_cast_fp16")];
+            tensor<int32, [4]> var_3656 = const()[name = tensor<string, []>("op_3656"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 73, 8, 64]> v_cast_fp16 = reshape(shape = var_3656, x = linear_149_cast_fp16)[name = tensor<string, []>("v_cast_fp16")];
+            tensor<int32, [4]> value_perm_0 = const()[name = tensor<string, []>("value_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206231296)))];
+            tensor<fp16, [1, 3, 8, 64]> var_3668_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3668_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206232384)))];
+            tensor<fp16, [1, 3, 8, 64]> var_3670_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3670_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_423_transpose_x_0 = const()[name = tensor<string, []>("x_423_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_423_transpose_y_0 = const()[name = tensor<string, []>("x_423_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 145]> var_3672_to_fp16 = const()[name = tensor<string, []>("op_3672_to_fp16"), val = tensor<fp16, [1, 8, 64, 145]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206233472)))];
+            tensor<fp16, [1, 8, 3, 64]> q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_3670_cast_fp16)[name = tensor<string, []>("transpose_93")];
+            tensor<fp16, [1, 8, 3, 145]> x_423_cast_fp16 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_cast_fp16, y = var_3672_to_fp16)[name = tensor<string, []>("x_423_cast_fp16")];
+            tensor<int32, [8]> x_425_pad_0 = const()[name = tensor<string, []>("x_425_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_425_mode_0 = const()[name = tensor<string, []>("x_425_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_231_to_fp16 = const()[name = tensor<string, []>("const_231_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 3, 146]> x_425_cast_fp16 = pad(constant_val = const_231_to_fp16, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423_cast_fp16)[name = tensor<string, []>("x_425_cast_fp16")];
+            tensor<int32, [4]> var_3680 = const()[name = tensor<string, []>("op_3680"), val = tensor<int32, [4]>([1, 8, -1, 3])];
+            tensor<fp16, [1, 8, 146, 3]> x_427_cast_fp16 = reshape(shape = var_3680, x = x_425_cast_fp16)[name = tensor<string, []>("x_427_cast_fp16")];
+            tensor<int32, [4]> var_3684_begin_0 = const()[name = tensor<string, []>("op_3684_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3684_end_0 = const()[name = tensor<string, []>("op_3684_end_0"), val = tensor<int32, [4]>([1, 8, 146, 3])];
+            tensor<bool, [4]> var_3684_end_mask_0 = const()[name = tensor<string, []>("op_3684_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 145, 3]> var_3684_cast_fp16 = slice_by_index(begin = var_3684_begin_0, end = var_3684_end_0, end_mask = var_3684_end_mask_0, x = x_427_cast_fp16)[name = tensor<string, []>("op_3684_cast_fp16")];
+            tensor<int32, [4]> var_3685 = const()[name = tensor<string, []>("op_3685"), val = tensor<int32, [4]>([1, 8, 3, 145])];
+            tensor<fp16, [1, 8, 3, 145]> matrix_bd_65_cast_fp16 = reshape(shape = var_3685, x = var_3684_cast_fp16)[name = tensor<string, []>("matrix_bd_65_cast_fp16")];
+            tensor<bool, []> matrix_ac_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_83_perm_0 = const()[name = tensor<string, []>("transpose_83_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_84_perm_0 = const()[name = tensor<string, []>("transpose_84_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 73]> transpose_84 = transpose(perm = transpose_84_perm_0, x = k_65_cast_fp16)[name = tensor<string, []>("transpose_91")];
+            tensor<fp16, [1, 8, 3, 64]> transpose_83 = transpose(perm = transpose_83_perm_0, x = var_3668_cast_fp16)[name = tensor<string, []>("transpose_92")];
+            tensor<fp16, [1, 8, 3, 73]> matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_83, y = transpose_84)[name = tensor<string, []>("matrix_ac_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_begin_0 = const()[name = tensor<string, []>("matrix_bd_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_end_0 = const()[name = tensor<string, []>("matrix_bd_end_0"), val = tensor<int32, [4]>([1, 8, 3, 73])];
+            tensor<bool, [4]> matrix_bd_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 3, 73]> matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_65_cast_fp16)[name = tensor<string, []>("matrix_bd_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_3694_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = tensor<string, []>("op_3694_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_65_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_65_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 3, 73]> _inversed_scores_65_cast_fp16 = mul(x = var_3694_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_65_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> scores_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_3)[name = tensor<string, []>("scores_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> var_3700_cast_fp16 = softmax(axis = var_62, x = scores_cast_fp16)[name = tensor<string, []>("op_3700_cast_fp16")];
+            tensor<fp16, [1, 8, 3, 73]> input_873_cast_fp16 = select(a = var_40_to_fp16, b = var_3700_cast_fp16, cond = mask_3)[name = tensor<string, []>("input_873_cast_fp16")];
+            tensor<bool, []> x_429_transpose_x_0 = const()[name = tensor<string, []>("x_429_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_429_transpose_y_0 = const()[name = tensor<string, []>("x_429_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 73, 64]> value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = tensor<string, []>("transpose_94")];
+            tensor<fp16, [1, 8, 3, 64]> x_429_cast_fp16 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873_cast_fp16, y = value_cast_fp16)[name = tensor<string, []>("x_429_cast_fp16")];
+            tensor<int32, [4]> var_3704_perm_0 = const()[name = tensor<string, []>("op_3704_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3705 = const()[name = tensor<string, []>("op_3705"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 3, 8, 64]> var_3704_cast_fp16 = transpose(perm = var_3704_perm_0, x = x_429_cast_fp16)[name = tensor<string, []>("transpose_90")];
+            tensor<fp16, [1, 3, 512]> input_875_cast_fp16 = reshape(shape = var_3705, x = var_3704_cast_fp16)[name = tensor<string, []>("input_875_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_16_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206382016)))];
+            tensor<fp16, [1, 3, 512]> linear_151_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16, x = input_875_cast_fp16)[name = tensor<string, []>("linear_151_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = tensor<string, []>("input_879_cast_fp16")];
+            tensor<int32, [1]> x_433_axes_0 = const()[name = tensor<string, []>("x_433_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206906368)))];
+            tensor<fp16, [512]> encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206907456)))];
+            tensor<fp16, [1, 3, 512]> x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = tensor<string, []>("x_433_cast_fp16")];
+            tensor<int32, [3]> input_881_perm_0 = const()[name = tensor<string, []>("input_881_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_883_pad_type_0 = const()[name = tensor<string, []>("input_883_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_883_strides_0 = const()[name = tensor<string, []>("input_883_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_883_pad_0 = const()[name = tensor<string, []>("input_883_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_883_dilations_0 = const()[name = tensor<string, []>("input_883_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_883_groups_0 = const()[name = tensor<string, []>("input_883_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_16_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206908544)))];
+            tensor<fp16, [1, 512, 3]> input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = tensor<string, []>("transpose_89")];
+            tensor<fp16, [1, 1024, 3]> input_883_cast_fp16 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16, x = input_881_cast_fp16)[name = tensor<string, []>("input_883_cast_fp16")];
+            tensor<int32, []> x_435_split_num_splits_0 = const()[name = tensor<string, []>("x_435_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_435_split_axis_0 = const()[name = tensor<string, []>("x_435_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 3]> x_435_split_cast_fp16_0, tensor<fp16, [1, 512, 3]> x_435_split_cast_fp16_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883_cast_fp16)[name = tensor<string, []>("x_435_split_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = tensor<string, []>("x_435_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_435_cast_fp16")];
+            tensor<fp16, [1, 512, 3]> input_885_cast_fp16 = select(a = var_40_to_fp16, b = x_435_cast_fp16, cond = var_418)[name = tensor<string, []>("input_885_cast_fp16")];
+            tensor<bool, []> new_x_interleave_0 = const()[name = tensor<string, []>("new_x_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 11]> new_x_cast_fp16 = concat(axis = var_62, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_885_cast_fp16))[name = tensor<string, []>("new_x_cast_fp16")];
+            tensor<int32, [3]> next_cache_begin_0 = const()[name = tensor<string, []>("next_cache_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_end_0 = const()[name = tensor<string, []>("next_cache_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> next_cache_end_mask_0 = const()[name = tensor<string, []>("next_cache_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 9]> next_cache_cast_fp16 = slice_by_index(begin = next_cache_begin_0, end = next_cache_end_0, end_mask = next_cache_end_mask_0, x = new_x_cast_fp16)[name = tensor<string, []>("next_cache_cast_fp16")];
+            tensor<int32, [3]> cache_last_time_cur_begin_0 = const()[name = tensor<string, []>("cache_last_time_cur_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> cache_last_time_cur_end_0 = const()[name = tensor<string, []>("cache_last_time_cur_end_0"), val = tensor<int32, [3]>([1, 512, 9])];
+            tensor<bool, [3]> cache_last_time_cur_end_mask_0 = const()[name = tensor<string, []>("cache_last_time_cur_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> cache_last_time_cur_cast_fp16 = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = next_cache_cast_fp16)[name = tensor<string, []>("cache_last_time_cur_cast_fp16")];
+            tensor<string, []> x_437_pad_type_0 = const()[name = tensor<string, []>("x_437_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_437_groups_0 = const()[name = tensor<string, []>("x_437_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_437_strides_0 = const()[name = tensor<string, []>("x_437_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_437_pad_0 = const()[name = tensor<string, []>("x_437_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_437_dilations_0 = const()[name = tensor<string, []>("x_437_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_16_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207957184)))];
+            tensor<fp16, [1, 512, 3]> x_437_cast_fp16 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16, x = new_x_cast_fp16)[name = tensor<string, []>("x_437_cast_fp16")];
+            tensor<int32, [3]> input_887_perm_0 = const()[name = tensor<string, []>("input_887_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_439_axes_0 = const()[name = tensor<string, []>("x_439_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207966464)))];
+            tensor<fp16, [512]> encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207967552)))];
+            tensor<fp16, [1, 3, 512]> input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = tensor<string, []>("transpose_88")];
+            tensor<fp16, [1, 3, 512]> x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = tensor<string, []>("x_439_cast_fp16")];
+            tensor<int32, [3]> input_889_perm_0 = const()[name = tensor<string, []>("input_889_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 3]> input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = tensor<string, []>("transpose_87")];
+            tensor<fp16, [1, 512, 3]> input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = tensor<string, []>("input_891_cast_fp16")];
+            tensor<string, []> x_441_pad_type_0 = const()[name = tensor<string, []>("x_441_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_441_strides_0 = const()[name = tensor<string, []>("x_441_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_441_pad_0 = const()[name = tensor<string, []>("x_441_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_441_dilations_0 = const()[name = tensor<string, []>("x_441_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_441_groups_0 = const()[name = tensor<string, []>("x_441_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_16_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207968640)))];
+            tensor<fp16, [1, 512, 3]> x_441_cast_fp16 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16, x = input_891_cast_fp16)[name = tensor<string, []>("x_441_cast_fp16")];
+            tensor<int32, [3]> input_893_perm_0 = const()[name = tensor<string, []>("input_893_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 3, 512]> input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = tensor<string, []>("transpose_86")];
+            tensor<fp16, [1, 3, 512]> input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = tensor<string, []>("input_895_cast_fp16")];
+            tensor<int32, [1]> input_897_axes_0 = const()[name = tensor<string, []>("input_897_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208492992)))];
+            tensor<fp16, [512]> encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208494080)))];
+            tensor<fp16, [1, 3, 512]> input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = tensor<string, []>("input_897_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_16_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208495168)))];
+            tensor<fp16, [1, 3, 2048]> linear_152_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16, x = input_897_cast_fp16)[name = tensor<string, []>("linear_152_cast_fp16")];
+            tensor<fp16, [1, 3, 2048]> input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = tensor<string, []>("input_901_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_16_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210592384)))];
+            tensor<fp16, [1, 3, 512]> linear_153_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16, x = input_901_cast_fp16)[name = tensor<string, []>("linear_153_cast_fp16")];
+            tensor<fp16, []> var_3787_to_fp16 = const()[name = tensor<string, []>("op_3787_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 3, 512]> var_3788_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_3787_to_fp16)[name = tensor<string, []>("op_3788_cast_fp16")];
+            tensor<fp16, [1, 3, 512]> input_cast_fp16 = add(x = input_895_cast_fp16, y = var_3788_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [1]> audio_signal_axes_0 = const()[name = tensor<string, []>("audio_signal_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_16_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212689600)))];
+            tensor<fp16, [512]> encoder_layers_16_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212690688)))];
+            tensor<fp16, [1, 3, 512]> audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("audio_signal_cast_fp16")];
+            tensor<int32, [3]> obj_1_perm_0 = const()[name = tensor<string, []>("obj_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> cast_178_dtype_0 = const()[name = tensor<string, []>("cast_178_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, []> obj_5_axis_0 = const()[name = tensor<string, []>("obj_5_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [17, 1, 70, 512]> obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_332_cast_fp16, var_539_cast_fp16, var_746_cast_fp16, var_953_cast_fp16, var_1160_cast_fp16, var_1367_cast_fp16, var_1574_cast_fp16, var_1781_cast_fp16, var_1988_cast_fp16, var_2195_cast_fp16, var_2402_cast_fp16, var_2609_cast_fp16, var_2816_cast_fp16, var_3023_cast_fp16, var_3230_cast_fp16, var_3437_cast_fp16, cache_last_channel_cur_cast_fp16))[name = tensor<string, []>("obj_5_cast_fp16")];
+            tensor<int32, []> obj_7_axis_0 = const()[name = tensor<string, []>("obj_7_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [17, 1, 512, 8]> obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_434_cast_fp16, var_641_cast_fp16, var_848_cast_fp16, var_1055_cast_fp16, var_1262_cast_fp16, var_1469_cast_fp16, var_1676_cast_fp16, var_1883_cast_fp16, var_2090_cast_fp16, var_2297_cast_fp16, var_2504_cast_fp16, var_2711_cast_fp16, var_2918_cast_fp16, var_3125_cast_fp16, var_3332_cast_fp16, var_3539_cast_fp16, cache_last_time_cur_cast_fp16))[name = tensor<string, []>("obj_7_cast_fp16")];
+            tensor<string, []> obj_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_7_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> var_3804 = add(x = cache_last_channel_len, y = cache_keep_size)[name = tensor<string, []>("op_3804")];
+            tensor<string, []> var_3804_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3804_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> const_237_to_fp16 = const()[name = tensor<string, []>("const_237_to_fp16"), val = tensor<fp16, []>(-inf)];
+            tensor<fp16, []> var_45_promoted_to_fp16 = const()[name = tensor<string, []>("op_45_promoted_to_fp16"), val = tensor<fp16, []>(0x1.18p+6)];
+            tensor<fp16, [1]> var_3804_to_fp16 = cast(dtype = var_3804_promoted_to_fp16_dtype_0, x = var_3804)[name = tensor<string, []>("cast_186")];
+            tensor<fp16, [1]> clip_1_cast_fp16 = clip(alpha = const_237_to_fp16, beta = var_45_promoted_to_fp16, x = var_3804_to_fp16)[name = tensor<string, []>("clip_1_cast_fp16")];
+            tensor<string, []> var_3831_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3831_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [3]> var_3846_begin_0 = const()[name = tensor<string, []>("op_3846_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_3846_end_0 = const()[name = tensor<string, []>("op_3846_end_0"), val = tensor<int32, [3]>([1, 512, 2])];
+            tensor<bool, [3]> var_3846_end_mask_0 = const()[name = tensor<string, []>("op_3846_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 3]> obj_1_cast_fp16 = transpose(perm = obj_1_perm_0, x = audio_signal_cast_fp16)[name = tensor<string, []>("transpose_85")];
+            tensor<fp16, [1, 512, 2]> var_3846_cast_fp16 = slice_by_index(begin = var_3846_begin_0, end = var_3846_end_0, end_mask = var_3846_end_mask_0, x = obj_1_cast_fp16)[name = tensor<string, []>("op_3846_cast_fp16")];
+            tensor<string, []> var_3846_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3846_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<string, []> cast_178_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_178_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> const_238_to_fp16 = const()[name = tensor<string, []>("const_238_to_fp16"), val = tensor<fp16, []>(-inf)];
+            tensor<fp16, []> var_3848_promoted_to_fp16 = const()[name = tensor<string, []>("op_3848_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<int32, [1]> clip_0_cast_fp16_to_int32 = cast(dtype = cast_178_dtype_0, x = clip_0_cast_fp16)[name = tensor<string, []>("cast_188")];
+            tensor<fp16, [1]> clip_0_cast_fp16_to_int32_to_fp16 = cast(dtype = cast_178_promoted_to_fp16_dtype_0, x = clip_0_cast_fp16_to_int32)[name = tensor<string, []>("cast_183")];
+            tensor<fp16, [1]> clip_2_cast_fp16 = clip(alpha = const_238_to_fp16, beta = var_3848_promoted_to_fp16, x = clip_0_cast_fp16_to_int32_to_fp16)[name = tensor<string, []>("clip_2_cast_fp16")];
+            tensor<string, []> cast_179_dtype_0 = const()[name = tensor<string, []>("cast_179_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<string, []> cast_180_dtype_0 = const()[name = tensor<string, []>("cast_180_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1]> new_cache_last_channel_len = cast(dtype = cast_180_dtype_0, x = clip_1_cast_fp16)[name = tensor<string, []>("cast_181")];
+            tensor<int32, [1]> encoded_length = cast(dtype = cast_179_dtype_0, x = clip_2_cast_fp16)[name = tensor<string, []>("cast_182")];
+            tensor<fp32, [1, 512, 2]> encoded_output = cast(dtype = var_3846_cast_fp16_to_fp32_dtype_0, x = var_3846_cast_fp16)[name = tensor<string, []>("cast_184")];
+            tensor<fp32, [17, 1, 70, 512]> new_cache_last_channel = cast(dtype = var_3831_cast_fp16_to_fp32_dtype_0, x = obj_5_cast_fp16)[name = tensor<string, []>("cast_185")];
+            tensor<fp32, [17, 1, 512, 8]> new_cache_last_time = cast(dtype = obj_7_cast_fp16_to_fp32_dtype_0, x = obj_7_cast_fp16)[name = tensor<string, []>("cast_187")];
+            tensor<fp32, [1, 128, 16]> new_pre_cache = cast(dtype = var_28_cast_fp16_to_fp32_dtype_0, x = var_28_cast_fp16)[name = tensor<string, []>("cast_194")];
+        } -> (encoded_output, encoded_length, new_pre_cache, new_cache_last_channel, new_cache_last_time, new_cache_last_channel_len);
+}
\ No newline at end of file
diff --git a/160ms/streaming_encoder.mlmodelc/weights/weight.bin b/160ms/streaming_encoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..36f875726932df5278655f7d4ef4e024c7ef9370
--- /dev/null
+++ b/160ms/streaming_encoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:12cd781a4300b52b6687587b7d8e37e0ce5c8ccb1dbea036008275e6abf5070c
+size 212691776
diff --git a/160ms/streaming_encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/160ms/streaming_encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..70022ef0bc4f8becf4754afe70d501e12951987a
--- /dev/null
+++ b/160ms/streaming_encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ca562061f63b3e908e835f45da5ea7668b560678c57c5279d7abd7d1e150846
+size 539800
diff --git a/160ms/streaming_encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/160ms/streaming_encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..36f875726932df5278655f7d4ef4e024c7ef9370
--- /dev/null
+++ b/160ms/streaming_encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:12cd781a4300b52b6687587b7d8e37e0ce5c8ccb1dbea036008275e6abf5070c
+size 212691776
diff --git a/160ms/streaming_encoder.mlpackage/Manifest.json b/160ms/streaming_encoder.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..98d489310c78257a8108a9ffc3cefb329029e499
--- /dev/null
+++ b/160ms/streaming_encoder.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "58E1A989-930B-49FA-AF42-1F7D3F023CA4": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "8E215896-5930-4E70-B597-BEF2276F4D32": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "8E215896-5930-4E70-B597-BEF2276F4D32"
+}
diff --git a/160ms/vocab.json b/160ms/vocab.json
new file mode 100644
index 0000000000000000000000000000000000000000..462081190dbb331a0dde13e99d5cfa99b302f0a7
--- /dev/null
+++ b/160ms/vocab.json
@@ -0,0 +1,1028 @@
+{
+  "0": "<unk>",
+  "1": "▁t",
+  "2": "▁th",
+  "3": "▁a",
+  "4": "▁i",
+  "5": "▁the",
+  "6": "▁s",
+  "7": "re",
+  "8": "▁w",
+  "9": "▁o",
+  "10": "in",
+  "11": "at",
+  "12": "er",
+  "13": "nd",
+  "14": "ou",
+  "15": "▁c",
+  "16": "▁b",
+  "17": "▁h",
+  "18": "en",
+  "19": "on",
+  "20": "▁m",
+  "21": "▁f",
+  "22": "ing",
+  "23": "▁p",
+  "24": "▁to",
+  "25": "▁and",
+  "26": "▁d",
+  "27": "an",
+  "28": "or",
+  "29": "es",
+  "30": "▁y",
+  "31": "▁l",
+  "32": "▁of",
+  "33": "ll",
+  "34": "▁in",
+  "35": "ed",
+  "36": "it",
+  "37": "▁g",
+  "38": "is",
+  "39": "▁you",
+  "40": "▁n",
+  "41": "ar",
+  "42": "om",
+  "43": "as",
+  "44": "ve",
+  "45": "▁e",
+  "46": "ic",
+  "47": "▁it",
+  "48": "al",
+  "49": "us",
+  "50": "▁wh",
+  "51": "▁we",
+  "52": "▁be",
+  "53": "ion",
+  "54": "ow",
+  "55": "le",
+  "56": "▁is",
+  "57": "et",
+  "58": "ent",
+  "59": "ot",
+  "60": "ut",
+  "61": "▁re",
+  "62": "▁on",
+  "63": "ay",
+  "64": "▁ha",
+  "65": "ig",
+  "66": "▁so",
+  "67": "ct",
+  "68": "▁he",
+  "69": "▁for",
+  "70": "ver",
+  "71": "ke",
+  "72": "ro",
+  "73": "▁st",
+  "74": "id",
+  "75": "▁go",
+  "76": "all",
+  "77": "se",
+  "78": "ly",
+  "79": "▁u",
+  "80": "ch",
+  "81": "st",
+  "82": "ld",
+  "83": "▁k",
+  "84": "ce",
+  "85": "ur",
+  "86": "▁li",
+  "87": "am",
+  "88": "▁r",
+  "89": "ht",
+  "90": "▁j",
+  "91": "ith",
+  "92": "▁se",
+  "93": "ir",
+  "94": "▁as",
+  "95": "▁an",
+  "96": "im",
+  "97": "▁do",
+  "98": "ad",
+  "99": "▁was",
+  "100": "ight",
+  "101": "th",
+  "102": "▁are",
+  "103": "▁but",
+  "104": "▁sh",
+  "105": "ust",
+  "106": "ally",
+  "107": "▁not",
+  "108": "▁or",
+  "109": "▁com",
+  "110": "▁can",
+  "111": "▁me",
+  "112": "op",
+  "113": "▁mo",
+  "114": "▁at",
+  "115": "ill",
+  "116": "▁ch",
+  "117": "▁ne",
+  "118": "ant",
+  "119": "▁de",
+  "120": "▁kn",
+  "121": "▁one",
+  "122": "il",
+  "123": "ol",
+  "124": "▁con",
+  "125": "ter",
+  "126": "▁ab",
+  "127": "▁fr",
+  "128": "ere",
+  "129": "ck",
+  "130": "▁al",
+  "131": "▁all",
+  "132": "qu",
+  "133": "▁pro",
+  "134": "▁som",
+  "135": "ould",
+  "136": "▁tw",
+  "137": "ul",
+  "138": "ra",
+  "139": "od",
+  "140": "ers",
+  "141": "▁su",
+  "142": "ive",
+  "143": "▁v",
+  "144": "use",
+  "145": "ate",
+  "146": "ge",
+  "147": "if",
+  "148": "▁ex",
+  "149": "ess",
+  "150": "pp",
+  "151": "▁lo",
+  "152": "out",
+  "153": "▁if",
+  "154": "est",
+  "155": "ain",
+  "156": "ist",
+  "157": "and",
+  "158": "ea",
+  "159": "very",
+  "160": "art",
+  "161": "▁wor",
+  "162": "▁my",
+  "163": "ab",
+  "164": "ment",
+  "165": "▁bec",
+  "166": "un",
+  "167": "ity",
+  "168": "ri",
+  "169": "pe",
+  "170": "ions",
+  "171": "▁by",
+  "172": "ok",
+  "173": "our",
+  "174": "ort",
+  "175": "ind",
+  "176": "ink",
+  "177": "nt",
+  "178": "▁up",
+  "179": "um",
+  "180": "▁don",
+  "181": "▁get",
+  "182": "red",
+  "183": "▁out",
+  "184": "el",
+  "185": "ause",
+  "186": "res",
+  "187": "▁ma",
+  "188": "ich",
+  "189": "▁us",
+  "190": "rou",
+  "191": "▁int",
+  "192": "em",
+  "193": "os",
+  "194": "ies",
+  "195": "ie",
+  "196": "▁pl",
+  "197": "▁tr",
+  "198": "ven",
+  "199": "ous",
+  "200": "▁le",
+  "201": "▁two",
+  "202": "ard",
+  "203": "ine",
+  "204": "▁co",
+  "205": "een",
+  "206": "▁now",
+  "207": "ty",
+  "208": "her",
+  "209": "ack",
+  "210": "▁pe",
+  "211": "ame",
+  "212": "▁how",
+  "213": "▁who",
+  "214": "▁see",
+  "215": "▁tim",
+  "216": "ect",
+  "217": "ast",
+  "218": "▁our",
+  "219": "ci",
+  "220": "ree",
+  "221": "ople",
+  "222": "gh",
+  "223": "▁no",
+  "224": "▁had",
+  "225": "▁man",
+  "226": "▁qu",
+  "227": "▁en",
+  "228": "ide",
+  "229": "ure",
+  "230": "ud",
+  "231": "so",
+  "232": "▁his",
+  "233": "▁sa",
+  "234": "▁sp",
+  "235": "▁say",
+  "236": "ose",
+  "237": "ther",
+  "238": "▁act",
+  "239": "▁ta",
+  "240": "▁cl",
+  "241": "ings",
+  "242": "pt",
+  "243": "king",
+  "244": "▁any",
+  "245": "▁has",
+  "246": "▁un",
+  "247": "iv",
+  "248": "▁im",
+  "249": "▁ag",
+  "250": "▁te",
+  "251": "▁fe",
+  "252": "one",
+  "253": "per",
+  "254": "ong",
+  "255": "▁po",
+  "256": "▁ad",
+  "257": "ff",
+  "258": "ore",
+  "259": "itt",
+  "260": "ans",
+  "261": "iz",
+  "262": "eah",
+  "263": "reat",
+  "264": "act",
+  "265": "own",
+  "266": "hing",
+  "267": "enty",
+  "268": "age",
+  "269": "ber",
+  "270": "ice",
+  "271": "▁am",
+  "272": "ple",
+  "273": "are",
+  "274": "▁per",
+  "275": "und",
+  "276": "ite",
+  "277": "ix",
+  "278": "pl",
+  "279": "▁way",
+  "280": "▁did",
+  "281": "▁pr",
+  "282": "▁got",
+  "283": "ars",
+  "284": "▁she",
+  "285": "▁let",
+  "286": "ag",
+  "287": "▁ac",
+  "288": "int",
+  "289": "▁ar",
+  "290": "ry",
+  "291": "ign",
+  "292": "ish",
+  "293": "▁fir",
+  "294": "ace",
+  "295": "ble",
+  "296": "og",
+  "297": "ue",
+  "298": "▁ye",
+  "299": "ap",
+  "300": "iff",
+  "301": "▁ro",
+  "302": "▁her",
+  "303": "nder",
+  "304": "▁ok",
+  "305": "▁res",
+  "306": "▁gu",
+  "307": "ence",
+  "308": "▁may",
+  "309": "ated",
+  "310": "ip",
+  "311": "▁bo",
+  "312": "▁him",
+  "313": "way",
+  "314": "ac",
+  "315": "ical",
+  "316": "ass",
+  "317": "ase",
+  "318": "▁dis",
+  "319": "able",
+  "320": "ick",
+  "321": "▁app",
+  "322": "ance",
+  "323": "▁pre",
+  "324": "▁six",
+  "325": "▁off",
+  "326": "▁new",
+  "327": "ia",
+  "328": "orm",
+  "329": "ank",
+  "330": "▁lot",
+  "331": "ach",
+  "332": "▁fo",
+  "333": "inet",
+  "334": "ire",
+  "335": "ary",
+  "336": "ult",
+  "337": "▁tal",
+  "338": "▁mu",
+  "339": "▁bl",
+  "340": "ount",
+  "341": "sel",
+  "342": "vel",
+  "343": "▁br",
+  "344": "▁imp",
+  "345": "ep",
+  "346": "cess",
+  "347": "ord",
+  "348": "▁sc",
+  "349": "▁inc",
+  "350": "ound",
+  "351": "ang",
+  "352": "be",
+  "353": "ress",
+  "354": "uct",
+  "355": "▁ind",
+  "356": "▁af",
+  "357": "ving",
+  "358": "▁oh",
+  "359": "▁bet",
+  "360": "▁use",
+  "361": "ome",
+  "362": "ens",
+  "363": "ys",
+  "364": "▁bu",
+  "365": "co",
+  "366": "ory",
+  "367": "ater",
+  "368": "ild",
+  "369": "ght",
+  "370": "ial",
+  "371": "▁day",
+  "372": "ning",
+  "373": "na",
+  "374": "ile",
+  "375": "▁spe",
+  "376": "▁mar",
+  "377": "ody",
+  "378": "ough",
+  "379": "ade",
+  "380": "vers",
+  "381": "xt",
+  "382": "▁fl",
+  "383": "▁ke",
+  "384": "ian",
+  "385": "▁sy",
+  "386": "▁put",
+  "387": "fore",
+  "388": "ub",
+  "389": "▁ph",
+  "390": "fe",
+  "391": "▁em",
+  "392": "▁ser",
+  "393": "form",
+  "394": "ting",
+  "395": "te",
+  "396": "av",
+  "397": "ious",
+  "398": "▁rec",
+  "399": "ks",
+  "400": "▁gr",
+  "401": "ces",
+  "402": "wn",
+  "403": "ors",
+  "404": "▁jo",
+  "405": "ents",
+  "406": "▁des",
+  "407": "▁try",
+  "408": "▁equ",
+  "409": "▁z",
+  "410": "▁rem",
+  "411": "▁str",
+  "412": "self",
+  "413": "▁bit",
+  "414": "ph",
+  "415": "ved",
+  "416": "▁why",
+  "417": "▁bas",
+  "418": "▁hel",
+  "419": "▁rel",
+  "420": "ath",
+  "421": "ject",
+  "422": "ail",
+  "423": "▁la",
+  "424": "ual",
+  "425": "▁god",
+  "426": "▁nat",
+  "427": "erm",
+  "428": "day",
+  "429": "▁id",
+  "430": "ft",
+  "431": "▁wr",
+  "432": "▁min",
+  "433": "ates",
+  "434": "▁gen",
+  "435": "tain",
+  "436": "▁ob",
+  "437": "ull",
+  "438": "ict",
+  "439": "▁tra",
+  "440": "▁end",
+  "441": "▁hig",
+  "442": "▁fif",
+  "443": "oth",
+  "444": "tern",
+  "445": "▁its",
+  "446": "vent",
+  "447": "▁sm",
+  "448": "ons",
+  "449": "▁add",
+  "450": "iss",
+  "451": "▁bel",
+  "452": "ful",
+  "453": "get",
+  "454": "▁ele",
+  "455": "▁rep",
+  "456": "ak",
+  "457": "▁ho",
+  "458": "▁pos",
+  "459": "▁num",
+  "460": "ange",
+  "461": "ves",
+  "462": "ific",
+  "463": "urn",
+  "464": "ise",
+  "465": "▁cr",
+  "466": "▁um",
+  "467": "ward",
+  "468": "▁reg",
+  "469": "ady",
+  "470": "ower",
+  "471": "uc",
+  "472": "▁dec",
+  "473": "lic",
+  "474": "▁set",
+  "475": "▁gon",
+  "476": "▁op",
+  "477": "▁ear",
+  "478": "▁sub",
+  "479": "▁sl",
+  "480": "les",
+  "481": "stem",
+  "482": "cial",
+  "483": "olog",
+  "484": "atch",
+  "485": "ily",
+  "486": "body",
+  "487": "nds",
+  "488": "ular",
+  "489": "ren",
+  "490": "▁own",
+  "491": "▁too",
+  "492": "cent",
+  "493": "ible",
+  "494": "pect",
+  "495": "ered",
+  "496": "ways",
+  "497": "teen",
+  "498": "▁uh",
+  "499": "▁big",
+  "500": "▁mod",
+  "501": "▁att",
+  "502": "▁car",
+  "503": "gr",
+  "504": "▁acc",
+  "505": "ied",
+  "506": "mun",
+  "507": "ib",
+  "508": "▁mon",
+  "509": "▁sch",
+  "510": "▁pol",
+  "511": "▁dat",
+  "512": "▁fin",
+  "513": "▁sim",
+  "514": "▁inv",
+  "515": "▁def",
+  "516": "ked",
+  "517": "▁ent",
+  "518": "▁yes",
+  "519": "ows",
+  "520": "ics",
+  "521": "ited",
+  "522": "ute",
+  "523": "ism",
+  "524": "ps",
+  "525": "▁ed",
+  "526": "▁el",
+  "527": "ably",
+  "528": "ppen",
+  "529": "als",
+  "530": "▁ten",
+  "531": "ract",
+  "532": "ss",
+  "533": "▁ass",
+  "534": "▁met",
+  "535": "gan",
+  "536": "▁eng",
+  "537": "▁stu",
+  "538": "ween",
+  "539": "arch",
+  "540": "▁gl",
+  "541": "▁cor",
+  "542": "▁dr",
+  "543": "vern",
+  "544": "▁ty",
+  "545": "▁run",
+  "546": "hip",
+  "547": "cus",
+  "548": "cond",
+  "549": "▁ins",
+  "550": "irty",
+  "551": "▁pub",
+  "552": "lud",
+  "553": "llow",
+  "554": "▁cou",
+  "555": "ew",
+  "556": "iew",
+  "557": "▁sur",
+  "558": "ero",
+  "559": "ood",
+  "560": "ness",
+  "561": "▁fun",
+  "562": "▁eff",
+  "563": "cept",
+  "564": "▁ca",
+  "565": "▁exp",
+  "566": "duct",
+  "567": "▁sw",
+  "568": "ize",
+  "569": "ope",
+  "570": "▁par",
+  "571": "kes",
+  "572": "cy",
+  "573": "▁ev",
+  "574": "▁ref",
+  "575": "ell",
+  "576": "▁bus",
+  "577": "ug",
+  "578": "rib",
+  "579": "▁cur",
+  "580": "mo",
+  "581": "ock",
+  "582": "ures",
+  "583": "air",
+  "584": "▁war",
+  "585": "str",
+  "586": "▁med",
+  "587": "▁wa",
+  "588": "▁val",
+  "589": "▁sin",
+  "590": "blem",
+  "591": "▁fam",
+  "592": "li",
+  "593": "▁far",
+  "594": "▁cle",
+  "595": "▁col",
+  "596": "mon",
+  "597": "▁gra",
+  "598": "led",
+  "599": "ense",
+  "600": "tin",
+  "601": "ues",
+  "602": "its",
+  "603": "▁mem",
+  "604": "▁inf",
+  "605": "▁eas",
+  "606": "ideo",
+  "607": "▁top",
+  "608": "io",
+  "609": "pan",
+  "610": "▁hum",
+  "611": "▁old",
+  "612": "ead",
+  "613": "▁ord",
+  "614": "ric",
+  "615": "ants",
+  "616": "oy",
+  "617": "esn",
+  "618": "uck",
+  "619": "ason",
+  "620": "ced",
+  "621": "ool",
+  "622": "rat",
+  "623": "ouse",
+  "624": "▁lar",
+  "625": "▁art",
+  "626": "▁wee",
+  "627": "▁cer",
+  "628": "ized",
+  "629": "▁mat",
+  "630": "con",
+  "631": "erg",
+  "632": "land",
+  "633": "ines",
+  "634": "▁chr",
+  "635": "▁aut",
+  "636": "▁lea",
+  "637": "▁sou",
+  "638": "oney",
+  "639": "tty",
+  "640": "▁ple",
+  "641": "ulat",
+  "642": "oks",
+  "643": "▁few",
+  "644": "▁sol",
+  "645": "▁che",
+  "646": "chn",
+  "647": "ird",
+  "648": "▁bre",
+  "649": "▁dur",
+  "650": "▁wom",
+  "651": "me",
+  "652": "izat",
+  "653": "eric",
+  "654": "ote",
+  "655": "▁uni",
+  "656": "eren",
+  "657": "arn",
+  "658": "ross",
+  "659": "ices",
+  "660": "ten",
+  "661": "eral",
+  "662": "ever",
+  "663": "ieve",
+  "664": "lish",
+  "665": "ash",
+  "666": "▁opp",
+  "667": "alth",
+  "668": "ger",
+  "669": "▁sk",
+  "670": "▁red",
+  "671": "peri",
+  "672": "▁det",
+  "673": "▁ext",
+  "674": "ner",
+  "675": "ah",
+  "676": "▁var",
+  "677": "▁loc",
+  "678": "gram",
+  "679": "ists",
+  "680": "ives",
+  "681": "▁es",
+  "682": "▁nor",
+  "683": "tro",
+  "684": "ale",
+  "685": "▁iss",
+  "686": "▁pri",
+  "687": "gin",
+  "688": "az",
+  "689": "oc",
+  "690": "▁pop",
+  "691": "ern",
+  "692": "▁sit",
+  "693": "ket",
+  "694": "▁pa",
+  "695": "▁law",
+  "696": "ages",
+  "697": "br",
+  "698": "▁cam",
+  "699": "▁mom",
+  "700": "osed",
+  "701": "▁bro",
+  "702": "ne",
+  "703": "bs",
+  "704": "▁cre",
+  "705": "erat",
+  "706": "▁sec",
+  "707": "▁cap",
+  "708": "▁vis",
+  "709": "▁pat",
+  "710": "ield",
+  "711": "iet",
+  "712": "▁tri",
+  "713": "up",
+  "714": "▁bra",
+  "715": "ts",
+  "716": "▁mot",
+  "717": "▁unt",
+  "718": "put",
+  "719": "bo",
+  "720": "ork",
+  "721": "mer",
+  "722": "ital",
+  "723": "▁air",
+  "724": "ined",
+  "725": "▁beh",
+  "726": "▁adv",
+  "727": "▁ret",
+  "728": "imes",
+  "729": "▁tea",
+  "730": "ural",
+  "731": "sid",
+  "732": "ters",
+  "733": "▁pur",
+  "734": "▁sci",
+  "735": "bers",
+  "736": "ient",
+  "737": "ier",
+  "738": "cc",
+  "739": "sw",
+  "740": "▁av",
+  "741": "reen",
+  "742": "ode",
+  "743": "ont",
+  "744": "▁dra",
+  "745": "ann",
+  "746": "nect",
+  "747": "▁x",
+  "748": "▁eu",
+  "749": "ton",
+  "750": "inat",
+  "751": "ene",
+  "752": "ared",
+  "753": "els",
+  "754": "▁mor",
+  "755": "▁rat",
+  "756": "cri",
+  "757": "▁men",
+  "758": "▁ah",
+  "759": "ames",
+  "760": "▁arm",
+  "761": "eak",
+  "762": "▁pay",
+  "763": "▁hal",
+  "764": "ins",
+  "765": "ilit",
+  "766": "stit",
+  "767": "▁ra",
+  "768": "▁leg",
+  "769": "cl",
+  "770": "pr",
+  "771": "▁wal",
+  "772": "▁bad",
+  "773": "▁ge",
+  "774": "roup",
+  "775": "▁mus",
+  "776": "man",
+  "777": "▁gi",
+  "778": "eds",
+  "779": "▁aw",
+  "780": "po",
+  "781": "ark",
+  "782": "row",
+  "783": "▁dep",
+  "784": "ully",
+  "785": "ral",
+  "786": "lect",
+  "787": "pend",
+  "788": "▁sev",
+  "789": "ime",
+  "790": "gest",
+  "791": "here",
+  "792": "▁yet",
+  "793": "ted",
+  "794": "▁rev",
+  "795": "ds",
+  "796": "▁ask",
+  "797": "less",
+  "798": "▁di",
+  "799": "ets",
+  "800": "line",
+  "801": "▁aff",
+  "802": "ired",
+  "803": "▁est",
+  "804": "ken",
+  "805": "vid",
+  "806": "most",
+  "807": "ivid",
+  "808": "unch",
+  "809": "par",
+  "810": "med",
+  "811": "rop",
+  "812": "ased",
+  "813": "eone",
+  "814": "▁ve",
+  "815": "▁abs",
+  "816": "ergy",
+  "817": "ret",
+  "818": "▁saw",
+  "819": "▁ey",
+  "820": "▁cal",
+  "821": "uat",
+  "822": "▁mid",
+  "823": "vat",
+  "824": "ream",
+  "825": "vice",
+  "826": "ians",
+  "827": "rent",
+  "828": "ctor",
+  "829": "err",
+  "830": "ush",
+  "831": "ases",
+  "832": "▁suc",
+  "833": "erms",
+  "834": "ave",
+  "835": "angu",
+  "836": "ries",
+  "837": "▁wo",
+  "838": "arts",
+  "839": "▁fil",
+  "840": "▁fat",
+  "841": "▁cho",
+  "842": "orts",
+  "843": "▁fre",
+  "844": "ee",
+  "845": "ught",
+  "846": "eng",
+  "847": "ump",
+  "848": "▁bar",
+  "849": "ying",
+  "850": "ane",
+  "851": "▁tem",
+  "852": "anks",
+  "853": "ury",
+  "854": "iat",
+  "855": "mit",
+  "856": "trol",
+  "857": "▁net",
+  "858": "▁maj",
+  "859": "▁cra",
+  "860": "ling",
+  "861": "▁fig",
+  "862": "orn",
+  "863": "icat",
+  "864": "pany",
+  "865": "▁occ",
+  "866": "ott",
+  "867": "ands",
+  "868": "▁exc",
+  "869": "▁mr",
+  "870": "ency",
+  "871": "rope",
+  "872": "itch",
+  "873": "▁lit",
+  "874": "abil",
+  "875": "not",
+  "876": "ma",
+  "877": "▁typ",
+  "878": "▁opt",
+  "879": "ob",
+  "880": "ser",
+  "881": "ety",
+  "882": "ms",
+  "883": "peci",
+  "884": "aces",
+  "885": "aut",
+  "886": "▁hon",
+  "887": "cuss",
+  "888": "▁sal",
+  "889": "▁sor",
+  "890": "att",
+  "891": "▁lab",
+  "892": "▁har",
+  "893": "urch",
+  "894": "nded",
+  "895": "uce",
+  "896": "ids",
+  "897": "▁hy",
+  "898": "▁fut",
+  "899": "▁ste",
+  "900": "ours",
+  "901": "ems",
+  "902": "utes",
+  "903": "ng",
+  "904": "ta",
+  "905": "▁won",
+  "906": "▁fa",
+  "907": "▁env",
+  "908": "ards",
+  "909": "▁job",
+  "910": "ium",
+  "911": "▁dot",
+  "912": "▁obv",
+  "913": "ina",
+  "914": "side",
+  "915": "elve",
+  "916": "cu",
+  "917": "▁jes",
+  "918": "▁pot",
+  "919": "▁pie",
+  "920": "▁tre",
+  "921": "▁hey",
+  "922": "▁mag",
+  "923": "ron",
+  "924": "▁key",
+  "925": "swer",
+  "926": "▁win",
+  "927": "ucat",
+  "928": "work",
+  "929": "ides",
+  "930": "▁low",
+  "931": "▁vol",
+  "932": "▁oth",
+  "933": "atic",
+  "934": "lf",
+  "935": "ads",
+  "936": "inds",
+  "937": "com",
+  "938": "ths",
+  "939": "▁ver",
+  "940": "ised",
+  "941": "lo",
+  "942": "▁squ",
+  "943": "▁cut",
+  "944": "oked",
+  "945": "irit",
+  "946": "ateg",
+  "947": "ppy",
+  "948": "mitt",
+  "949": "come",
+  "950": "hn",
+  "951": "igin",
+  "952": "mand",
+  "953": "▁dam",
+  "954": "ho",
+  "955": "▁da",
+  "956": "▁fur",
+  "957": "iron",
+  "958": "ilar",
+  "959": "▁fac",
+  "960": "▁neg",
+  "961": "▁ago",
+  "962": "ged",
+  "963": "miss",
+  "964": "enth",
+  "965": "▁dou",
+  "966": "▁hit",
+  "967": "▁guy",
+  "968": "▁bi",
+  "969": "ove",
+  "970": "fess",
+  "971": "ples",
+  "972": "owed",
+  "973": "ured",
+  "974": "▁ris",
+  "975": "ints",
+  "976": "rew",
+  "977": "▁sum",
+  "978": "▁hu",
+  "979": "ploy",
+  "980": "ude",
+  "981": "ried",
+  "982": "▁cir",
+  "983": "▁dev",
+  "984": "ear",
+  "985": "▁tot",
+  "986": "▁ann",
+  "987": "duc",
+  "988": "ik",
+  "989": "pon",
+  "990": "sted",
+  "991": "▁ide",
+  "992": "▁'",
+  "993": "ipp",
+  "994": "▁eat",
+  "995": "▁dom",
+  "996": "▁",
+  "997": "e",
+  "998": "t",
+  "999": "o",
+  "1000": "a",
+  "1001": "i",
+  "1002": "n",
+  "1003": "s",
+  "1004": "r",
+  "1005": "h",
+  "1006": "l",
+  "1007": "d",
+  "1008": "u",
+  "1009": "c",
+  "1010": "m",
+  "1011": "y",
+  "1012": "g",
+  "1013": "w",
+  "1014": "f",
+  "1015": "p",
+  "1016": "b",
+  "1017": "v",
+  "1018": "k",
+  "1019": "'",
+  "1020": "j",
+  "1021": "x",
+  "1022": "q",
+  "1023": "z",
+  "1024": "<EOU>",
+  "1025": "<EOB>"
+}
\ No newline at end of file
diff --git a/320ms/.DS_Store b/320ms/.DS_Store
new file mode 100644
index 0000000000000000000000000000000000000000..8e1aae8afb05f089f83507db6d67efd90fb51604
Binary files /dev/null and b/320ms/.DS_Store differ
diff --git a/320ms/convert_parakeet_eou.py b/320ms/convert_parakeet_eou.py
new file mode 100644
index 0000000000000000000000000000000000000000..9c865f90c8132840b5755dcd62fcfac15e3f0352
--- /dev/null
+++ b/320ms/convert_parakeet_eou.py
@@ -0,0 +1,740 @@
+#!/usr/bin/env python3
+"""CLI for exporting Parakeet Realtime EOU 120M components to CoreML.
+
+This model is a cache-aware streaming FastConformer-RNNT model optimized for
+low-latency speech recognition with end-of-utterance detection.
+
+Key differences from Parakeet TDT v3:
+- Smaller model (120M vs 600M params)
+- No duration outputs (standard RNNT, not TDT)
+- Cache-aware streaming encoder (17 layers, attention context [70,1])
+- Special <EOU> token for end-of-utterance detection
+- Optimized for 80-160ms latency
+
+Reference: https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1
+"""
+from __future__ import annotations
+
+import json
+from dataclasses import asdict
+from pathlib import Path
+from typing import Dict, Optional, Tuple
+
+import coremltools as ct
+import numpy as np
+import soundfile as sf
+import torch
+import typer
+
+import nemo.collections.asr as nemo_asr
+
+from individual_components import (
+    DecoderWrapper,
+    EncoderWrapper,
+    ExportSettings,
+    JointWrapper,
+    JointDecisionWrapper,
+    JointDecisionSingleStep,
+    PreprocessorWrapper,
+    MelEncoderWrapper,
+    _coreml_convert,
+)
+
+def apply_stft_patch():
+    # Monkey patch coremltools.stft to handle extra arguments from newer torch versions
+    try:
+        import coremltools.converters.mil.frontend.torch.ops as torch_ops
+        _original_stft = torch_ops.stft
+
+        def patched_stft(context, node):
+            if len(node.inputs) > 8:
+                node.inputs = node.inputs[:8]
+            return _original_stft(context, node)
+
+        torch_ops.stft = patched_stft
+        if "stft" in torch_ops._TORCH_OPS_REGISTRY:
+            torch_ops._TORCH_OPS_REGISTRY["stft"] = patched_stft
+        print("Monkey patched coremltools.stft for compatibility.")
+    except Exception as e:
+        print(f"Warning: Could not monkey patch stft: {e}")
+
+DEFAULT_MODEL_ID = "nvidia/parakeet_realtime_eou_120m-v1"
+AUTHOR = "Fluid Inference"
+
+
+def _compute_length(seconds: float, sample_rate: int) -> int:
+    return int(round(seconds * sample_rate))
+
+
+def _prepare_audio(
+    validation_audio: Optional[Path],
+    sample_rate: int,
+    max_samples: int,
+    seed: Optional[int],
+) -> torch.Tensor:
+    if validation_audio is None:
+        if seed is not None:
+            torch.manual_seed(seed)
+        audio = torch.randn(1, max_samples, dtype=torch.float32)
+        return audio
+
+    data, sr = sf.read(str(validation_audio), dtype="float32")
+    if sr != sample_rate:
+        raise typer.BadParameter(
+            f"Validation audio sample rate {sr} does not match model rate {sample_rate}"
+        )
+
+    if data.ndim > 1:
+        data = data[:, 0]
+
+    if data.size == 0:
+        raise typer.BadParameter("Validation audio is empty")
+
+    if data.size < max_samples:
+        pad_width = max_samples - data.size
+        data = np.pad(data, (0, pad_width))
+    elif data.size > max_samples:
+        data = data[:max_samples]
+
+    audio = torch.from_numpy(data).unsqueeze(0).to(dtype=torch.float32)
+    return audio
+
+
+def _save_mlpackage(model: ct.models.MLModel, path: Path, description: str) -> None:
+    try:
+        model.minimum_deployment_target = ct.target.iOS17
+    except Exception:
+        pass
+    model.short_description = description
+    model.author = AUTHOR
+    path.parent.mkdir(parents=True, exist_ok=True)
+    model.save(str(path))
+
+
+def _tensor_shape(tensor: torch.Tensor) -> Tuple[int, ...]:
+    return tuple(int(dim) for dim in tensor.shape)
+
+
+def _parse_compute_units(name: str) -> ct.ComputeUnit:
+    """Parse a human-friendly compute units string into ct.ComputeUnit."""
+    normalized = str(name).strip().upper()
+    mapping = {
+        "ALL": ct.ComputeUnit.ALL,
+        "CPU_ONLY": ct.ComputeUnit.CPU_ONLY,
+        "CPU_AND_GPU": ct.ComputeUnit.CPU_AND_GPU,
+        "CPU_AND_NE": ct.ComputeUnit.CPU_AND_NE,
+        "CPU_AND_NEURALENGINE": ct.ComputeUnit.CPU_AND_NE,
+    }
+    if normalized not in mapping:
+        raise typer.BadParameter(
+            f"Unknown compute units '{name}'. Choose from: " + ", ".join(mapping.keys())
+        )
+    return mapping[normalized]
+
+
+def _parse_compute_precision(name: Optional[str]) -> Optional[ct.precision]:
+    """Parse compute precision string into ct.precision or None."""
+    if name is None:
+        return None
+    normalized = str(name).strip().upper()
+    if normalized == "":
+        return None
+    mapping = {
+        "FLOAT32": ct.precision.FLOAT32,
+        "FLOAT16": ct.precision.FLOAT16,
+    }
+    if normalized not in mapping:
+        raise typer.BadParameter(
+            f"Unknown compute precision '{name}'. Choose from: "
+            + ", ".join(mapping.keys())
+        )
+    return mapping[normalized]
+
+
+app = typer.Typer(add_completion=False, pretty_exceptions_show_locals=False)
+
+
+@app.command()
+def convert(
+    nemo_path: Optional[Path] = typer.Option(
+        None,
+        "--nemo-path",
+        exists=True,
+        resolve_path=True,
+        help="Path to parakeet_realtime_eou_120m-v1.nemo checkpoint (skip to auto-download)",
+    ),
+    model_id: str = typer.Option(
+        DEFAULT_MODEL_ID,
+        "--model-id",
+        help="Model identifier to download when --nemo-path is omitted",
+    ),
+    output_dir: Path = typer.Option(
+        Path("parakeet_eou_coreml"),
+        help="Directory where mlpackages and metadata will be written",
+    ),
+    preprocessor_cu: str = typer.Option(
+        "CPU_ONLY",
+        "--preprocessor-cu",
+        help="Compute units for preprocessor (default CPU_ONLY)",
+    ),
+    mel_encoder_cu: str = typer.Option(
+        "CPU_ONLY",
+        "--mel-encoder-cu",
+        help="Compute units for fused mel+encoder (default CPU_ONLY)",
+    ),
+    compute_precision: Optional[str] = typer.Option(
+        None,
+        "--compute-precision",
+        help="Export precision: FLOAT32 (default) or FLOAT16 to shrink non-quantized weights.",
+    ),
+    max_audio_seconds: float = typer.Option(
+        15.0,
+        "--max-audio-seconds",
+        help="Maximum audio duration in seconds for the fixed window export",
+    ),
+    validation_audio: Optional[Path] = typer.Option(
+        None,
+        "--validation-audio",
+        exists=True,
+        resolve_path=True,
+        help="Path to a 16kHz WAV file for tracing (uses random if not provided)",
+    ),
+) -> None:
+    """Export all Parakeet Realtime EOU sub-modules to CoreML.
+
+    This exports the cache-aware streaming FastConformer-RNNT model for
+    low-latency speech recognition with end-of-utterance detection.
+    """
+    export_settings = ExportSettings(
+        output_dir=output_dir,
+        compute_units=ct.ComputeUnit.CPU_ONLY,
+        deployment_target=ct.target.iOS17,
+        compute_precision=_parse_compute_precision(compute_precision),
+        max_audio_seconds=max_audio_seconds,
+        max_symbol_steps=1,
+    )
+
+    typer.echo("Export configuration:")
+    typer.echo(asdict(export_settings))
+
+    output_dir.mkdir(parents=True, exist_ok=True)
+    pre_cu = _parse_compute_units(preprocessor_cu)
+    melenc_cu = _parse_compute_units(mel_encoder_cu)
+
+    if nemo_path is not None:
+        typer.echo(f"Loading NeMo model from {nemo_path}…")
+        # Try loading as generic ASRModel first, then specific class
+        try:
+            asr_model = nemo_asr.models.ASRModel.restore_from(
+                str(nemo_path), map_location="cpu"
+            )
+        except Exception:
+            # Fallback to EncDecRNNTBPEModel
+            asr_model = nemo_asr.models.EncDecRNNTBPEModel.restore_from(
+                str(nemo_path), map_location="cpu"
+            )
+        checkpoint_meta = {
+            "type": "file",
+            "path": str(nemo_path),
+        }
+    else:
+        typer.echo(f"Downloading NeMo model via {model_id}…")
+        # Use ASRModel.from_pretrained as recommended for this model
+        try:
+            asr_model = nemo_asr.models.ASRModel.from_pretrained(
+                model_id, map_location="cpu"
+            )
+        except Exception:
+            # Fallback to EncDecRNNTBPEModel
+            asr_model = nemo_asr.models.EncDecRNNTBPEModel.from_pretrained(
+                model_id, map_location="cpu"
+            )
+        checkpoint_meta = {
+            "type": "pretrained",
+            "model_id": model_id,
+        }
+    asr_model.eval()
+
+    # Print model info
+    typer.echo(f"Model class: {type(asr_model).__name__}")
+    typer.echo(f"Encoder class: {type(asr_model.encoder).__name__}")
+
+    sample_rate = int(asr_model.cfg.preprocessor.sample_rate)
+    max_samples = _compute_length(export_settings.max_audio_seconds, sample_rate)
+
+    # Prepare audio for tracing
+    if validation_audio is not None:
+        typer.echo(f"Using validation audio: {validation_audio}")
+        audio_tensor = _prepare_audio(validation_audio, sample_rate, max_samples, seed=None)
+    else:
+        typer.echo("Using random audio for tracing (seed=42)")
+        audio_tensor = _prepare_audio(None, sample_rate, max_samples, seed=42)
+
+    audio_length = torch.tensor([max_samples], dtype=torch.int32)
+
+    preprocessor = PreprocessorWrapper(asr_model.preprocessor.eval())
+    encoder = EncoderWrapper(asr_model.encoder.eval())
+    decoder = DecoderWrapper(asr_model.decoder.eval())
+    joint = JointWrapper(asr_model.joint.eval())
+
+    decoder_export_flag = getattr(asr_model.decoder, "_rnnt_export", False)
+    asr_model.decoder._rnnt_export = True
+
+    try:
+        with torch.no_grad():
+            mel_ref, mel_length_ref = preprocessor(audio_tensor, audio_length)
+            mel_length_ref = mel_length_ref.to(dtype=torch.int32)
+            encoder_ref, encoder_length_ref, frame_times_ref = encoder(
+                mel_ref, mel_length_ref
+            )
+            encoder_length_ref = encoder_length_ref.to(dtype=torch.int32)
+
+            # Clone tensors to drop inference flags
+            mel_ref = mel_ref.clone().detach()
+            mel_length_ref = mel_length_ref.clone().detach()
+            encoder_ref = encoder_ref.clone().detach()
+            encoder_length_ref = encoder_length_ref.clone().detach()
+            frame_times_ref = frame_times_ref.clone().detach()
+
+        vocab_size = int(asr_model.tokenizer.vocab_size)
+        decoder_hidden = int(asr_model.decoder.pred_hidden)
+        decoder_layers = int(asr_model.decoder.pred_rnn_layers)
+
+        # Check if model has extra outputs (TDT-style duration)
+        num_extra = getattr(asr_model.joint, "num_extra_outputs", 0)
+        typer.echo(f"Vocab size: {vocab_size}, num_extra_outputs: {num_extra}")
+
+        targets = torch.full(
+            (1, export_settings.max_symbol_steps),
+            fill_value=asr_model.decoder.blank_idx,
+            dtype=torch.int32,
+        )
+        target_lengths = torch.tensor(
+            [export_settings.max_symbol_steps], dtype=torch.int32
+        )
+        zero_state = torch.zeros(
+            decoder_layers,
+            1,
+            decoder_hidden,
+            dtype=torch.float32,
+        )
+
+        with torch.no_grad():
+            decoder_ref, h_ref, c_ref = decoder(
+                targets, target_lengths, zero_state, zero_state
+            )
+            joint_ref = joint(encoder_ref, decoder_ref)
+
+        decoder_ref = decoder_ref.clone()
+        h_ref = h_ref.clone()
+        c_ref = c_ref.clone()
+        joint_ref = joint_ref.clone()
+
+        typer.echo(f"Encoder output shape: {encoder_ref.shape}")
+        typer.echo(f"Decoder output shape: {decoder_ref.shape}")
+        typer.echo(f"Joint output shape: {joint_ref.shape}")
+
+        # === Export Preprocessor ===
+        typer.echo("Tracing and converting preprocessor…")
+        preprocessor = preprocessor.cpu()
+        audio_tensor = audio_tensor.cpu()
+        audio_length = audio_length.cpu()
+        traced_preprocessor = torch.jit.trace(
+            preprocessor, (audio_tensor, audio_length), strict=False
+        )
+        traced_preprocessor.eval()
+        preprocessor_inputs = [
+            ct.TensorType(
+                name="audio_signal",
+                shape=(1, ct.RangeDim(1, max_samples)),
+                dtype=np.float32,
+            ),
+            ct.TensorType(name="audio_length", shape=(1,), dtype=np.int32),
+        ]
+        preprocessor_outputs = [
+            ct.TensorType(name="mel", dtype=np.float32),
+            ct.TensorType(name="mel_length", dtype=np.int32),
+        ]
+        preprocessor_model = _coreml_convert(
+            traced_preprocessor,
+            preprocessor_inputs,
+            preprocessor_outputs,
+            export_settings,
+            compute_units_override=pre_cu,
+        )
+        preprocessor_path = output_dir / "parakeet_eou_preprocessor.mlpackage"
+        _save_mlpackage(
+            preprocessor_model,
+            preprocessor_path,
+            f"Parakeet EOU preprocessor ({max_audio_seconds}s window)",
+        )
+
+        # === Export Encoder ===
+        typer.echo("Tracing and converting encoder…")
+        traced_encoder = torch.jit.trace(
+            encoder, (mel_ref, mel_length_ref), strict=False
+        )
+        traced_encoder.eval()
+        encoder_inputs = [
+            ct.TensorType(
+                name="mel", shape=_tensor_shape(mel_ref), dtype=np.float32
+            ),
+            ct.TensorType(name="mel_length", shape=(1,), dtype=np.int32),
+        ]
+        encoder_outputs = [
+            ct.TensorType(name="encoder", dtype=np.float32),
+            ct.TensorType(name="encoder_length", dtype=np.int32),
+            ct.TensorType(name="frame_times", dtype=np.float32),
+        ]
+        encoder_model = _coreml_convert(
+            traced_encoder,
+            encoder_inputs,
+            encoder_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        encoder_path = output_dir / "parakeet_eou_encoder.mlpackage"
+        _save_mlpackage(
+            encoder_model,
+            encoder_path,
+            f"Parakeet EOU encoder ({max_audio_seconds}s window)",
+        )
+
+        # === Export Fused Mel+Encoder ===
+        typer.echo("Tracing and converting fused mel+encoder…")
+        mel_encoder = MelEncoderWrapper(preprocessor, encoder)
+        traced_mel_encoder = torch.jit.trace(
+            mel_encoder, (audio_tensor, audio_length), strict=False
+        )
+        traced_mel_encoder.eval()
+        mel_encoder_inputs = [
+            ct.TensorType(
+                name="audio_signal", shape=(1, max_samples), dtype=np.float32
+            ),
+            ct.TensorType(name="audio_length", shape=(1,), dtype=np.int32),
+        ]
+        mel_encoder_outputs = [
+            ct.TensorType(name="encoder", dtype=np.float32),
+            ct.TensorType(name="encoder_length", dtype=np.int32),
+            ct.TensorType(name="frame_times", dtype=np.float32),
+        ]
+        mel_encoder_model = _coreml_convert(
+            traced_mel_encoder,
+            mel_encoder_inputs,
+            mel_encoder_outputs,
+            export_settings,
+            compute_units_override=melenc_cu,
+        )
+        mel_encoder_path = output_dir / "parakeet_eou_mel_encoder.mlpackage"
+        _save_mlpackage(
+            mel_encoder_model,
+            mel_encoder_path,
+            f"Parakeet EOU fused Mel+Encoder ({max_audio_seconds}s window)",
+        )
+
+        # === Export Decoder ===
+        typer.echo("Tracing and converting decoder…")
+        traced_decoder = torch.jit.trace(
+            decoder,
+            (targets, target_lengths, zero_state, zero_state),
+            strict=False,
+        )
+        traced_decoder.eval()
+        decoder_inputs = [
+            ct.TensorType(
+                name="targets", shape=_tensor_shape(targets), dtype=np.int32
+            ),
+            ct.TensorType(name="target_length", shape=(1,), dtype=np.int32),
+            ct.TensorType(
+                name="h_in", shape=_tensor_shape(zero_state), dtype=np.float32
+            ),
+            ct.TensorType(
+                name="c_in", shape=_tensor_shape(zero_state), dtype=np.float32
+            ),
+        ]
+        decoder_outputs = [
+            ct.TensorType(name="decoder", dtype=np.float32),
+            ct.TensorType(name="h_out", dtype=np.float32),
+            ct.TensorType(name="c_out", dtype=np.float32),
+        ]
+        decoder_model = _coreml_convert(
+            traced_decoder,
+            decoder_inputs,
+            decoder_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        decoder_path = output_dir / "parakeet_eou_decoder.mlpackage"
+        _save_mlpackage(
+            decoder_model,
+            decoder_path,
+            "Parakeet EOU decoder (RNNT prediction network)",
+        )
+
+        # === Export Joint ===
+        typer.echo("Tracing and converting joint…")
+        traced_joint = torch.jit.trace(
+            joint,
+            (encoder_ref, decoder_ref),
+            strict=False,
+        )
+        traced_joint.eval()
+        joint_inputs = [
+            ct.TensorType(
+                name="encoder", shape=_tensor_shape(encoder_ref), dtype=np.float32
+            ),
+            ct.TensorType(
+                name="decoder", shape=_tensor_shape(decoder_ref), dtype=np.float32
+            ),
+        ]
+        joint_outputs = [
+            ct.TensorType(name="logits", dtype=np.float32),
+        ]
+        joint_model = _coreml_convert(
+            traced_joint,
+            joint_inputs,
+            joint_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        joint_path = output_dir / "parakeet_eou_joint.mlpackage"
+        _save_mlpackage(
+            joint_model,
+            joint_path,
+            "Parakeet EOU joint network (RNNT)",
+        )
+
+        # === Export Joint Decision Head ===
+        typer.echo("Tracing and converting joint decision head…")
+        joint_decision = JointDecisionWrapper(joint, vocab_size=vocab_size)
+        traced_joint_decision = torch.jit.trace(
+            joint_decision,
+            (encoder_ref, decoder_ref),
+            strict=False,
+        )
+        traced_joint_decision.eval()
+        joint_decision_inputs = [
+            ct.TensorType(
+                name="encoder", shape=_tensor_shape(encoder_ref), dtype=np.float32
+            ),
+            ct.TensorType(
+                name="decoder", shape=_tensor_shape(decoder_ref), dtype=np.float32
+            ),
+        ]
+        joint_decision_outputs = [
+            ct.TensorType(name="token_id", dtype=np.int32),
+            ct.TensorType(name="token_prob", dtype=np.float32),
+        ]
+        joint_decision_model = _coreml_convert(
+            traced_joint_decision,
+            joint_decision_inputs,
+            joint_decision_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        joint_decision_path = output_dir / "parakeet_eou_joint_decision.mlpackage"
+        _save_mlpackage(
+            joint_decision_model,
+            joint_decision_path,
+            "Parakeet EOU joint + decision head (softmax, argmax)",
+        )
+
+        # === Export Single-Step Joint Decision ===
+        typer.echo("Tracing and converting single-step joint decision…")
+        jd_single = JointDecisionSingleStep(joint, vocab_size=vocab_size)
+        # Create single-step slices from refs
+        enc_step = encoder_ref[:, :, :1].contiguous()
+        dec_step = decoder_ref[:, :, :1].contiguous()
+        traced_jd_single = torch.jit.trace(
+            jd_single,
+            (enc_step, dec_step),
+            strict=False,
+        )
+        traced_jd_single.eval()
+        jd_single_inputs = [
+            ct.TensorType(
+                name="encoder_step",
+                shape=(1, enc_step.shape[1], 1),
+                dtype=np.float32,
+            ),
+            ct.TensorType(
+                name="decoder_step",
+                shape=(1, dec_step.shape[1], 1),
+                dtype=np.float32,
+            ),
+        ]
+        jd_single_outputs = [
+            ct.TensorType(name="token_id", dtype=np.int32),
+            ct.TensorType(name="token_prob", dtype=np.float32),
+            ct.TensorType(name="top_k_ids", dtype=np.int32),
+            ct.TensorType(name="top_k_logits", dtype=np.float32),
+        ]
+        jd_single_model = _coreml_convert(
+            traced_jd_single,
+            jd_single_inputs,
+            jd_single_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        jd_single_path = output_dir / "parakeet_eou_joint_decision_single_step.mlpackage"
+        _save_mlpackage(
+            jd_single_model,
+            jd_single_path,
+            "Parakeet EOU single-step joint decision (current frame)",
+        )
+
+        # === Save Metadata ===
+        metadata: Dict[str, object] = {
+            "model_id": model_id,
+            "model_name": "parakeet_realtime_eou_120m-v1",
+            "model_class": type(asr_model).__name__,
+            "encoder_class": type(asr_model.encoder).__name__,
+            "sample_rate": sample_rate,
+            "max_audio_seconds": export_settings.max_audio_seconds,
+            "max_audio_samples": max_samples,
+            "max_symbol_steps": export_settings.max_symbol_steps,
+            "vocab_size": vocab_size,
+            "vocab_with_blank": vocab_size + 1,
+            "decoder_hidden": decoder_hidden,
+            "decoder_layers": decoder_layers,
+            "num_extra_outputs": num_extra,
+            "has_eou_token": True,
+            "checkpoint": checkpoint_meta,
+            "coreml": {
+                "compute_units": export_settings.compute_units.name,
+                "compute_precision": (
+                    export_settings.compute_precision.name
+                    if export_settings.compute_precision is not None
+                    else "FLOAT32"
+                ),
+            },
+            "components": {
+                "preprocessor": {
+                    "inputs": {
+                        "audio_signal": [1, max_samples],
+                        "audio_length": [1],
+                    },
+                    "outputs": {
+                        "mel": list(_tensor_shape(mel_ref)),
+                        "mel_length": [1],
+                    },
+                    "path": preprocessor_path.name,
+                },
+                "encoder": {
+                    "inputs": {
+                        "mel": list(_tensor_shape(mel_ref)),
+                        "mel_length": [1],
+                    },
+                    "outputs": {
+                        "encoder": list(_tensor_shape(encoder_ref)),
+                        "encoder_length": [1],
+                        "frame_times": [1, _tensor_shape(encoder_ref)[2]],
+                    },
+                    "path": encoder_path.name,
+                },
+                "mel_encoder": {
+                    "inputs": {
+                        "audio_signal": [1, max_samples],
+                        "audio_length": [1],
+                    },
+                    "outputs": {
+                        "encoder": list(_tensor_shape(encoder_ref)),
+                        "encoder_length": [1],
+                        "frame_times": [1, _tensor_shape(encoder_ref)[2]],
+                    },
+                    "path": mel_encoder_path.name,
+                },
+                "decoder": {
+                    "inputs": {
+                        "targets": list(_tensor_shape(targets)),
+                        "target_length": [1],
+                        "h_in": list(_tensor_shape(zero_state)),
+                        "c_in": list(_tensor_shape(zero_state)),
+                    },
+                    "outputs": {
+                        "decoder": list(_tensor_shape(decoder_ref)),
+                        "h_out": list(_tensor_shape(h_ref)),
+                        "c_out": list(_tensor_shape(c_ref)),
+                    },
+                    "path": decoder_path.name,
+                },
+                "joint": {
+                    "inputs": {
+                        "encoder": list(_tensor_shape(encoder_ref)),
+                        "decoder": list(_tensor_shape(decoder_ref)),
+                    },
+                    "outputs": {
+                        "logits": list(_tensor_shape(joint_ref)),
+                    },
+                    "path": joint_path.name,
+                },
+                "joint_decision": {
+                    "inputs": {
+                        "encoder": list(_tensor_shape(encoder_ref)),
+                        "decoder": list(_tensor_shape(decoder_ref)),
+                    },
+                    "outputs": {
+                        "token_id": [
+                            _tensor_shape(encoder_ref)[0],
+                            _tensor_shape(encoder_ref)[2],
+                            _tensor_shape(decoder_ref)[2],
+                        ],
+                        "token_prob": [
+                            _tensor_shape(encoder_ref)[0],
+                            _tensor_shape(encoder_ref)[2],
+                            _tensor_shape(decoder_ref)[2],
+                        ],
+                    },
+                    "path": joint_decision_path.name,
+                },
+                "joint_decision_single_step": {
+                    "inputs": {
+                        "encoder_step": [1, _tensor_shape(encoder_ref)[1], 1],
+                        "decoder_step": [1, _tensor_shape(decoder_ref)[1], 1],
+                    },
+                    "outputs": {
+                        "token_id": [1, 1, 1],
+                        "token_prob": [1, 1, 1],
+                        "top_k_ids": [1, 1, 1, 64],
+                        "top_k_logits": [1, 1, 1, 64],
+                    },
+                    "path": jd_single_path.name,
+                },
+            },
+        }
+
+        # Export tokenizer vocab if available
+        try:
+            tokenizer = asr_model.tokenizer
+            vocab = {
+                "blank_id": int(asr_model.decoder.blank_idx),
+                "vocab_size": vocab_size,
+            }
+            # Try to get special tokens
+            if hasattr(tokenizer, "tokenizer"):
+                inner_tokenizer = tokenizer.tokenizer
+                if hasattr(inner_tokenizer, "get_vocab"):
+                    full_vocab = inner_tokenizer.get_vocab()
+                    # Find EOU token
+                    eou_token = None
+                    for token, idx in full_vocab.items():
+                        if "<EOU>" in token.upper() or "eou" in token.lower():
+                            eou_token = {"token": token, "id": idx}
+                            break
+                    if eou_token:
+                        vocab["eou_token"] = eou_token
+            metadata["tokenizer"] = vocab
+        except Exception as e:
+            typer.echo(f"Warning: Could not export tokenizer info: {e}")
+
+        metadata_path = output_dir / "metadata.json"
+        metadata_path.write_text(json.dumps(metadata, indent=2))
+        typer.echo(f"\nExport complete. Metadata written to {metadata_path}")
+        typer.echo(f"Output directory: {output_dir}")
+
+    finally:
+        asr_model.decoder._rnnt_export = decoder_export_flag
+
+
+if __name__ == "__main__":
+    app()
diff --git a/320ms/convert_streaming_encoder.py b/320ms/convert_streaming_encoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..a1c199c8a6c540dab48b973ab4232db2901b8e72
--- /dev/null
+++ b/320ms/convert_streaming_encoder.py
@@ -0,0 +1,193 @@
+
+import torch
+import torch.nn as nn
+import coremltools as ct
+import numpy as np
+import typer
+from pathlib import Path
+from typing import Tuple, List, Optional
+import json
+import shutil
+
+# Iimport torch
+import coremltools as ct
+import numpy as np
+import argparse
+from nemo.collections.asr.models import EncDecRNNTBPEModel
+
+app = typer.Typer()
+
+class LoopbackEncoderWrapper(nn.Module):
+    """
+    Wraps the entire Parakeet Encoder (PreEncode + Conformer) for CoreML Loopback Streaming.
+    
+    Inputs:
+      - audio_signal: [B, D, T] (Mel spectrogram chunk)
+      - audio_length: [B]
+      - pre_cache: [B, D, pre_cache_size] (Previous audio context)
+      - cache_last_channel: [layers, B, cache_size, hidden]
+      - cache_last_time: [layers, B, hidden, time_cache]
+      - cache_last_channel_len: [B]
+      
+    Outputs:
+      - encoded_output: [B, D_out, T_out]
+      - encoded_length: [B]
+      - new_pre_cache: [B, D, pre_cache_size]
+      - new_cache_last_channel
+      - new_cache_last_time
+      - new_cache_last_channel_len
+    """
+    def __init__(self, encoder, pre_cache_size=16):
+        super().__init__()
+        self.encoder = encoder
+        self.pre_cache_size = pre_cache_size
+        
+    def forward(
+        self, 
+        audio_signal: torch.Tensor, 
+        audio_length: torch.Tensor,
+        pre_cache: torch.Tensor,
+        cache_last_channel: torch.Tensor,
+        cache_last_time: torch.Tensor,
+        cache_last_channel_len: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        
+        # 1. Prepend pre_cache to audio_signal
+        # audio_signal: [B, D, T]
+        # pre_cache: [B, D, T_cache]
+        full_input = torch.cat([pre_cache, audio_signal], dim=2)
+        full_length = audio_length + self.pre_cache_size
+        
+        # 2. Extract NEW pre_cache (last N frames of full_input)
+        # Note: We do this BEFORE processing because we want the raw audio context
+        new_pre_cache = full_input[:, :, -self.pre_cache_size:]
+        
+        # 3. Process with Encoder
+        # Reconstruct NeMo cache object
+        current_cache = [cache_last_channel, cache_last_time, cache_last_channel_len]
+        
+        encoded, encoded_len, new_cache_channel, new_cache_time, new_cache_len = self.encoder.cache_aware_stream_step(
+            processed_signal=full_input,
+            processed_signal_length=full_length,
+            cache_last_channel=cache_last_channel,
+            cache_last_time=cache_last_time,
+            cache_last_channel_len=cache_last_channel_len
+        )
+        
+        # 4. Drop the first few frames corresponding to pre_cache?
+        # NeMo's cache_aware_stream_step usually handles the "valid" output frames.
+        # But since we manually prepended, we might get extra output frames.
+        # However, for streaming, we usually want the model to see the context but only output the new tokens.
+        # Let's trust NeMo's streaming logic for now, or check if we need to slice.
+        # Given we are using 'cache_aware_stream_step', it expects the full context window?
+        # Actually, standard usage is: input IS the new chunk, but internal convolution looks at past.
+        # But since we are stateless, we MUST provide the past.
+        # So passing (pre_cache + chunk) is correct.
+        
+        # Cast lengths to Int32 for CoreML
+        encoded_len_32 = encoded_len.to(dtype=torch.int32)
+        new_channel_len_32 = new_cache_len.to(dtype=torch.int32)
+        
+        return encoded, encoded_len_32, new_pre_cache, new_cache_channel, new_cache_time, new_channel_len_32
+
+def _coreml_convert(
+    traced_model,
+    inputs,
+    outputs,
+    compute_units=ct.ComputeUnit.CPU_ONLY
+):
+    return ct.convert(
+        traced_model,
+        inputs=inputs,
+        outputs=outputs,
+        compute_units=compute_units,
+        minimum_deployment_target=ct.target.macOS14,
+    )
+
+def main():
+    model_id: str = "nvidia/parakeet_realtime_eou_120m-v1"
+    output_dir: str = "temp_swift_models/StreamingLoopback"
+    output_path = Path(output_dir)
+    output_path.mkdir(parents=True, exist_ok=True)
+    
+    print(f"Loading model: {model_id}...")
+    asr_model = EncDecRNNTBPEModel.from_pretrained(model_name=model_id)
+    asr_model.eval()
+    
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--chunk-frames", type=int, default=17, help="Number of frames in the input chunk (e.g. 17 for 160ms, 129 for 1.28s)")
+    args = parser.parse_args()
+
+    encoder = asr_model.encoder
+    
+    # --- Configuration ---
+    # 160ms chunk = 16 frames (but preprocessor produces 17 with padding/centering)
+    # 1.28s chunk = 128 frames (preprocessor produces 129)
+    chunk_size_in = args.chunk_frames 
+    mel_dim = 128
+    hidden_dim = encoder.d_model # 512
+    num_layers = len(encoder.layers) # 17
+    
+    # Cache sizes
+    cache_channel_size = 70
+    cache_time_size = 8
+    pre_cache_size = 16
+    
+    print(f"Config: Chunk={chunk_size_in}, Mel={mel_dim}, Hidden={hidden_dim}, Layers={num_layers}")
+    print(f"Cache: Channel={cache_channel_size}, Time={cache_time_size}, Pre={pre_cache_size}")
+
+    # --- Wrapper ---
+    wrapper = LoopbackEncoderWrapper(encoder, pre_cache_size=pre_cache_size)
+    wrapper.eval()
+    
+    # --- Test Inputs (for Tracing) ---
+    batch_size = 1
+    test_mel = torch.randn(batch_size, mel_dim, chunk_size_in)
+    test_mel_len = torch.tensor([chunk_size_in], dtype=torch.int32)
+    test_pre_cache = torch.zeros(batch_size, mel_dim, pre_cache_size)
+    
+    # Initial Cache (Zeros)
+    test_cache_channel = torch.zeros(num_layers, batch_size, cache_channel_size, hidden_dim)
+    test_cache_time = torch.zeros(num_layers, batch_size, hidden_dim, cache_time_size)
+    test_cache_len = torch.zeros(batch_size, dtype=torch.int32) 
+    
+    print("Tracing model...")
+    traced_model = torch.jit.trace(
+        wrapper,
+        (test_mel, test_mel_len, test_pre_cache, test_cache_channel, test_cache_time, test_cache_len),
+        strict=False
+    )
+    
+    # --- CoreML Conversion ---
+    print("Converting to CoreML...")
+    
+    inputs = [
+        ct.TensorType(name="audio_signal", shape=(1, 128, chunk_size_in), dtype=np.float32),
+        ct.TensorType(name="audio_length", shape=(1,), dtype=np.int32),
+        ct.TensorType(name="pre_cache", shape=(1, 128, pre_cache_size), dtype=np.float32),
+        ct.TensorType(name="cache_last_channel", shape=(num_layers, 1, cache_channel_size, hidden_dim), dtype=np.float32),
+        ct.TensorType(name="cache_last_time", shape=(num_layers, 1, hidden_dim, cache_time_size), dtype=np.float32),
+        ct.TensorType(name="cache_last_channel_len", shape=(1,), dtype=np.int32),
+    ]
+    
+    outputs = [
+        ct.TensorType(name="encoded_output", dtype=np.float32),
+        ct.TensorType(name="encoded_length", dtype=np.int32),
+        ct.TensorType(name="new_pre_cache", dtype=np.float32),
+        ct.TensorType(name="new_cache_last_channel", dtype=np.float32),
+        ct.TensorType(name="new_cache_last_time", dtype=np.float32),
+        ct.TensorType(name="new_cache_last_channel_len", dtype=np.int32),
+    ]
+    
+    mlmodel = _coreml_convert(traced_model, inputs, outputs)
+    
+    save_path = output_path / "streaming_encoder.mlpackage"
+    mlmodel.save(str(save_path))
+    print(f"Saved: {save_path}")
+    
+    # Also export Preprocessor, Decoder, Joint for completeness? 
+    # For now, let's assume we reuse the existing ones or export them separately if needed.
+    # But the user asked specifically for the Encoder loopback.
+
+if __name__ == "__main__":
+    main()
diff --git a/320ms/convert_streaming_encoder_unified.py b/320ms/convert_streaming_encoder_unified.py
new file mode 100644
index 0000000000000000000000000000000000000000..f457f84634ed52af0aeb3cb4907c121b884849d7
--- /dev/null
+++ b/320ms/convert_streaming_encoder_unified.py
@@ -0,0 +1,322 @@
+#!/usr/bin/env python3
+"""
+Unified CoreML conversion script for Parakeet EOU streaming encoder.
+
+Supports 160ms, 320ms, and 1600ms chunk sizes by properly configuring NeMo's
+streaming parameters before tracing.
+
+Usage:
+    # 160ms (default)
+    python convert_streaming_encoder_unified.py --chunk-ms 160 --output-dir Models/160ms/160ms
+
+    # 320ms
+    python convert_streaming_encoder_unified.py --chunk-ms 320 --output-dir Models/320ms
+
+    # 1600ms
+    python convert_streaming_encoder_unified.py --chunk-ms 1600 --output-dir Models/1600ms
+"""
+
+import argparse
+import json
+from pathlib import Path
+from typing import Tuple
+
+import coremltools as ct
+import numpy as np
+import torch
+import torch.nn as nn
+from nemo.collections.asr.models import EncDecRNNTBPEModel
+
+
+class LoopbackEncoderWrapper(nn.Module):
+    """
+    Wraps the Parakeet Encoder for CoreML Loopback Streaming.
+
+    This wrapper handles the pre_cache concatenation and cache management
+    that NeMo does internally in its streaming pipeline.
+
+    Inputs:
+      - audio_signal: [B, D, T] (Mel spectrogram chunk)
+      - audio_length: [B]
+      - pre_cache: [B, D, pre_cache_size] (Previous mel context)
+      - cache_last_channel: [layers, B, cache_size, hidden]
+      - cache_last_time: [layers, B, hidden, time_cache]
+      - cache_last_channel_len: [B]
+
+    Outputs:
+      - encoded_output: [B, D_out, T_out]
+      - encoded_length: [B]
+      - new_pre_cache: [B, D, pre_cache_size]
+      - new_cache_last_channel
+      - new_cache_last_time
+      - new_cache_last_channel_len
+    """
+
+    def __init__(self, encoder, pre_cache_size: int):
+        super().__init__()
+        self.encoder = encoder
+        self.pre_cache_size = pre_cache_size
+
+    def forward(
+        self,
+        audio_signal: torch.Tensor,
+        audio_length: torch.Tensor,
+        pre_cache: torch.Tensor,
+        cache_last_channel: torch.Tensor,
+        cache_last_time: torch.Tensor,
+        cache_last_channel_len: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+
+        # 1. Prepend pre_cache to audio_signal
+        full_input = torch.cat([pre_cache, audio_signal], dim=2)
+        full_length = audio_length + self.pre_cache_size
+
+        # 2. Extract NEW pre_cache (last N frames of full_input)
+        new_pre_cache = full_input[:, :, -self.pre_cache_size :]
+
+        # 3. Process with Encoder using cache_aware_stream_step
+        encoded, encoded_len, new_cache_channel, new_cache_time, new_cache_len = (
+            self.encoder.cache_aware_stream_step(
+                processed_signal=full_input,
+                processed_signal_length=full_length,
+                cache_last_channel=cache_last_channel,
+                cache_last_time=cache_last_time,
+                cache_last_channel_len=cache_last_channel_len,
+            )
+        )
+
+        # Cast lengths to Int32 for CoreML
+        encoded_len_32 = encoded_len.to(dtype=torch.int32)
+        new_channel_len_32 = new_cache_len.to(dtype=torch.int32)
+
+        return (
+            encoded,
+            encoded_len_32,
+            new_pre_cache,
+            new_cache_channel,
+            new_cache_time,
+            new_channel_len_32,
+        )
+
+
+def get_streaming_config(encoder, chunk_ms: int):
+    """
+    Get the correct streaming configuration for the given chunk size.
+
+    Returns:
+        dict with:
+        - chunk_size: encoder output steps
+        - shift_size: shift steps
+        - mel_frames: input mel frames for the chunk
+        - pre_cache_size: pre-encode cache size in mel frames
+        - valid_out_len: number of valid output frames per chunk
+    """
+    if chunk_ms == 160:
+        # Default 160ms config - no need to call setup_streaming_params
+        # Uses: chunk_size=[9, 16], valid_out_len=2
+        # With pre_cache=16, input is 17 mel frames
+        return {
+            "chunk_size": 4,  # encoder output steps
+            "shift_size": 2,
+            "mel_frames": 17,  # 16 + 1 for padding
+            "pre_cache_size": 16,
+            "valid_out_len": 2,
+            "samples": 2560,  # 160ms * 16000
+        }
+    elif chunk_ms == 320:
+        # 320ms config - need to call setup_streaming_params(chunk_size=8, shift_size=4)
+        # After setup: chunk_size=[57, 64], shift_size=[25, 32], pre_encode_cache_size=[0, 9]
+        # valid_out_len=4
+        encoder.setup_streaming_params(chunk_size=8, shift_size=4)
+        cfg = encoder.streaming_cfg
+        print(f"320ms streaming_cfg: {cfg}")
+
+        # Use NeMo's exact values from streaming_cfg:
+        # - chunk_size[1] = 64 mel frames
+        # - shift_size[1] = 32 mel frames (320ms latency)
+        # - pre_encode_cache_size[1] = 9
+        return {
+            "chunk_size": 8,  # encoder output steps
+            "shift_size": 4,
+            "mel_frames": 64,  # From NeMo cfg.chunk_size[1] = 64 mel frames
+            "pre_cache_size": 9,  # From NeMo cfg.pre_encode_cache_size[1] = 9
+            "valid_out_len": 4,
+            "samples": 10240,  # 64 mel frames * 160 hop_length = 10240 samples (~640ms audio)
+            "shift_samples": 5120,  # 32 mel frames * 160 = 5120 samples (320ms latency)
+        }
+    elif chunk_ms == 1600:
+        # 1600ms config - need to call setup_streaming_params(chunk_size=40, shift_size=20)
+        # After setup: chunk_size=[313, 320] mel frames, shift_size=[153, 160] mel frames
+        # The "1600ms" refers to the latency (shift), not the chunk size!
+        # - Chunk size: 320 mel frames (3183ms of audio)
+        # - Shift: 160 mel frames (1600ms latency)
+        # - Output: valid_out_len=20 encoder frames
+        encoder.setup_streaming_params(chunk_size=40, shift_size=20)
+        cfg = encoder.streaming_cfg
+        print(f"1600ms streaming_cfg: {cfg}")
+
+        # Audio samples for 320 mel frames: (320-1)*160 + 400 - 512 = 50928 samples (~3183ms)
+        # Audio shift for 160 mel frames: 160*160 = 25600 samples (1600ms)
+        return {
+            "chunk_size": 40,  # encoder output steps
+            "shift_size": 20,
+            "mel_frames": 320,  # From NeMo cfg.chunk_size[1] = 320 mel frames
+            "pre_cache_size": 9,  # From NeMo cfg.pre_encode_cache_size[1] = 9
+            "valid_out_len": 20,
+            "samples": 50928,  # (320-1)*160 + 400 - 512 = 50928 samples (~3183ms)
+            "shift_samples": 25600,  # 160*160 = 25600 samples (1600ms)
+        }
+    else:
+        raise ValueError(f"Unsupported chunk size: {chunk_ms}ms. Use 160, 320, or 1600.")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Convert Parakeet EOU encoder to CoreML")
+    parser.add_argument(
+        "--chunk-ms",
+        type=int,
+        default=160,
+        choices=[160, 320, 1600],
+        help="Chunk size in milliseconds (160, 320, or 1600)",
+    )
+    parser.add_argument(
+        "--output-dir",
+        type=str,
+        default=None,
+        help="Output directory for the CoreML model",
+    )
+    parser.add_argument(
+        "--model-id",
+        type=str,
+        default="nvidia/parakeet_realtime_eou_120m-v1",
+        help="HuggingFace model ID",
+    )
+    args = parser.parse_args()
+
+    # Default output directory
+    if args.output_dir is None:
+        args.output_dir = f"Models/{args.chunk_ms}ms"
+
+    output_path = Path(args.output_dir)
+    output_path.mkdir(parents=True, exist_ok=True)
+
+    print(f"Loading model: {args.model_id}...")
+    asr_model = EncDecRNNTBPEModel.from_pretrained(model_name=args.model_id)
+    asr_model.eval()
+
+    encoder = asr_model.encoder
+
+    # Get streaming configuration
+    config = get_streaming_config(encoder, args.chunk_ms)
+
+    mel_dim = 128
+    hidden_dim = encoder.d_model  # 512
+    num_layers = len(encoder.layers)  # 17
+
+    # Cache sizes
+    cache_channel_size = 70
+    cache_time_size = 8
+    pre_cache_size = config["pre_cache_size"]
+    chunk_size_in = config["mel_frames"]
+
+    print(f"\n=== Configuration for {args.chunk_ms}ms ===")
+    print(f"Mel frames: {chunk_size_in}")
+    print(f"Pre-cache: {pre_cache_size}")
+    print(f"Valid output len: {config['valid_out_len']}")
+    print(f"Hidden dim: {hidden_dim}, Layers: {num_layers}")
+    print(f"Cache: Channel={cache_channel_size}, Time={cache_time_size}")
+
+    # Create wrapper
+    wrapper = LoopbackEncoderWrapper(encoder, pre_cache_size=pre_cache_size)
+    wrapper.eval()
+
+    # Test inputs for tracing
+    batch_size = 1
+    test_mel = torch.randn(batch_size, mel_dim, chunk_size_in)
+    test_mel_len = torch.tensor([chunk_size_in], dtype=torch.int32)
+    test_pre_cache = torch.zeros(batch_size, mel_dim, pre_cache_size)
+    test_cache_channel = torch.zeros(num_layers, batch_size, cache_channel_size, hidden_dim)
+    test_cache_time = torch.zeros(num_layers, batch_size, hidden_dim, cache_time_size)
+    test_cache_len = torch.zeros(batch_size, dtype=torch.int32)
+
+    print("\nTracing model...")
+    traced_model = torch.jit.trace(
+        wrapper,
+        (test_mel, test_mel_len, test_pre_cache, test_cache_channel, test_cache_time, test_cache_len),
+        strict=False,
+    )
+
+    # Test output shape
+    with torch.no_grad():
+        out = traced_model(
+            test_mel, test_mel_len, test_pre_cache, test_cache_channel, test_cache_time, test_cache_len
+        )
+        print(f"Encoder output shape: {out[0].shape}")  # [B, D, T_out]
+
+    # CoreML conversion
+    print("\nConverting to CoreML...")
+    inputs = [
+        ct.TensorType(name="audio_signal", shape=(1, mel_dim, chunk_size_in), dtype=np.float32),
+        ct.TensorType(name="audio_length", shape=(1,), dtype=np.int32),
+        ct.TensorType(name="pre_cache", shape=(1, mel_dim, pre_cache_size), dtype=np.float32),
+        ct.TensorType(
+            name="cache_last_channel",
+            shape=(num_layers, 1, cache_channel_size, hidden_dim),
+            dtype=np.float32,
+        ),
+        ct.TensorType(
+            name="cache_last_time",
+            shape=(num_layers, 1, hidden_dim, cache_time_size),
+            dtype=np.float32,
+        ),
+        ct.TensorType(name="cache_last_channel_len", shape=(1,), dtype=np.int32),
+    ]
+
+    outputs = [
+        ct.TensorType(name="encoded_output", dtype=np.float32),
+        ct.TensorType(name="encoded_length", dtype=np.int32),
+        ct.TensorType(name="new_pre_cache", dtype=np.float32),
+        ct.TensorType(name="new_cache_last_channel", dtype=np.float32),
+        ct.TensorType(name="new_cache_last_time", dtype=np.float32),
+        ct.TensorType(name="new_cache_last_channel_len", dtype=np.int32),
+    ]
+
+    mlmodel = ct.convert(
+        traced_model,
+        inputs=inputs,
+        outputs=outputs,
+        compute_units=ct.ComputeUnit.CPU_ONLY,
+        minimum_deployment_target=ct.target.macOS14,
+    )
+
+    save_path = output_path / "streaming_encoder.mlpackage"
+    mlmodel.save(str(save_path))
+    print(f"Saved: {save_path}")
+
+    # Save metadata
+    metadata = {
+        "model_id": args.model_id,
+        "chunk_ms": args.chunk_ms,
+        "mel_frames": chunk_size_in,
+        "pre_cache_size": pre_cache_size,
+        "valid_out_len": config["valid_out_len"],
+        "samples_per_chunk": config["samples"],
+        "hidden_dim": hidden_dim,
+        "num_layers": num_layers,
+        "cache_channel_size": cache_channel_size,
+        "cache_time_size": cache_time_size,
+    }
+
+    metadata_path = output_path / "streaming_encoder_metadata.json"
+    with open(metadata_path, "w") as f:
+        json.dump(metadata, f, indent=2)
+    print(f"Saved metadata: {metadata_path}")
+
+    print(f"\n=== Export complete for {args.chunk_ms}ms ===")
+    print(f"Output: {output_path}")
+    print("\nNote: Decoder and Joint models are shared between chunk sizes.")
+    print("Copy decoder.mlmodelc, joint_decision.mlmodelc, and vocab.json from 160ms directory.")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/320ms/decoder.mlmodelc/analytics/coremldata.bin b/320ms/decoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..c6f51dedd2638184c5a2d71512339ceea086abc0
--- /dev/null
+++ b/320ms/decoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3996975a8cbc1949159c55605b3132b39b2484f51acbd55d796d93c70de02b49
+size 243
diff --git a/320ms/decoder.mlmodelc/coremldata.bin b/320ms/decoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..71936e61ffe7697e4af55bae8870d95c1fbb66c5
--- /dev/null
+++ b/320ms/decoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3ccbff963d8cf07e2be2bd56ea3384a89ea49628922c6bd95ff62e2ae57dc34
+size 497
diff --git a/320ms/decoder.mlmodelc/metadata.json b/320ms/decoder.mlmodelc/metadata.json
new file mode 100644
index 0000000000000000000000000000000000000000..3581673bcb062f7e0e1e0efa9f59edc5f218d50e
--- /dev/null
+++ b/320ms/decoder.mlmodelc/metadata.json
@@ -0,0 +1,118 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Parakeet EOU decoder (RNNT prediction network)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 640 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 640, 1]",
+        "name" : "decoder",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 640]",
+        "name" : "h_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 640]",
+        "name" : "c_out",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.squeeze" : 2,
+      "Ios17.gather" : 1,
+      "Ios17.cast" : 6,
+      "Ios17.lstm" : 1,
+      "Ios17.transpose" : 2,
+      "Identity" : 1,
+      "Ios17.expandDims" : 2
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int16, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1]",
+        "name" : "targets",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "target_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 640]",
+        "name" : "h_in",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 640]",
+        "name" : "c_in",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.4.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "parakeet_eou_decoder",
+    "method" : "predict"
+  }
+]
\ No newline at end of file
diff --git a/320ms/decoder.mlmodelc/model.mil b/320ms/decoder.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..5ead417fbdc13f8bf0c05a26039b2828503d4eca
--- /dev/null
+++ b/320ms/decoder.mlmodelc/model.mil
@@ -0,0 +1,45 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.4.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<fp32, [1, 1, 640]> c_in, tensor<fp32, [1, 1, 640]> h_in, tensor<int32, [1]> target_length, tensor<int32, [1, 1]> targets) {
+            tensor<int32, []> y_axis_0 = const()[name = tensor<string, []>("y_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> y_batch_dims_0 = const()[name = tensor<string, []>("y_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> y_validate_indices_0 = const()[name = tensor<string, []>("y_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1027, 640]> module_prediction_embed_weight_to_fp16 = const()[name = tensor<string, []>("module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [1027, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<string, []> targets_to_int16_dtype_0 = const()[name = tensor<string, []>("targets_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 1]> targets_to_int16 = cast(dtype = targets_to_int16_dtype_0, x = targets)[name = tensor<string, []>("cast_8")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = targets_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = tensor<string, []>("y_cast_fp16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [1]> input_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<string, []> h_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("h_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = tensor<string, []>("cast_7")];
+            tensor<fp16, [1, 640]> input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = h_in_to_fp16)[name = tensor<string, []>("input_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<string, []> c_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("c_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = tensor<string, []>("cast_6")];
+            tensor<fp16, [1, 640]> input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = c_in_to_fp16)[name = tensor<string, []>("input_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> input_direction_0 = const()[name = tensor<string, []>("input_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_output_sequence_0 = const()[name = tensor<string, []>("input_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_recurrent_activation_0 = const()[name = tensor<string, []>("input_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_cell_activation_0 = const()[name = tensor<string, []>("input_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_activation_0 = const()[name = tensor<string, []>("input_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = tensor<string, []>("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1314688)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = tensor<string, []>("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4591552)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = tensor<string, []>("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7868416)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_2")];
+            tensor<fp16, [1, 1, 640]> input_cast_fp16_0, tensor<fp16, [1, 640]> input_cast_fp16_1, tensor<fp16, [1, 640]> input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_0_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [1]> obj_3_axes_0 = const()[name = tensor<string, []>("obj_3_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 640]> obj_3_cast_fp16 = expand_dims(axes = obj_3_axes_0, x = input_cast_fp16_1)[name = tensor<string, []>("obj_3_cast_fp16")];
+            tensor<string, []> obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_3_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> obj_axes_0 = const()[name = tensor<string, []>("obj_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 640]> obj_cast_fp16 = expand_dims(axes = obj_axes_0, x = input_cast_fp16_2)[name = tensor<string, []>("obj_cast_fp16")];
+            tensor<string, []> obj_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = tensor<string, []>("transpose_0_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
+            tensor<string, []> transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("transpose_0_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp16, [1, 640, 1]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = tensor<string, []>("transpose_1")];
+            tensor<fp32, [1, 640, 1]> decoder = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = tensor<string, []>("cast_3")];
+            tensor<fp32, [1, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = tensor<string, []>("cast_4")];
+            tensor<fp32, [1, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = tensor<string, []>("cast_5")];
+            tensor<int32, [1]> target_length_tmp = identity(x = target_length)[name = tensor<string, []>("target_length_tmp")];
+        } -> (decoder, h_out, c_out);
+}
\ No newline at end of file
diff --git a/320ms/decoder.mlmodelc/weights/weight.bin b/320ms/decoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..cde1618cbed235b421984a302289e3bdd7e3df02
--- /dev/null
+++ b/320ms/decoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b4cacecdcd9df79ab1e56de67230baf5a8664d2afe0bb8f3408eefa972cb2f4
+size 7873600
diff --git a/320ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/320ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..0631f4ecb012ab42d13312be15475aff506055c5
--- /dev/null
+++ b/320ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:09f2dbd1f6a06faa6995f71d4b25d7c446996b6059cfac5ecc889853bdc7c6e5
+size 6728
diff --git a/320ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/320ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..cde1618cbed235b421984a302289e3bdd7e3df02
--- /dev/null
+++ b/320ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b4cacecdcd9df79ab1e56de67230baf5a8664d2afe0bb8f3408eefa972cb2f4
+size 7873600
diff --git a/320ms/decoder.mlpackage/Manifest.json b/320ms/decoder.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..4e498499a0a4026095bbc46d4e1a47ecbc39b0ca
--- /dev/null
+++ b/320ms/decoder.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "8201D73A-2B5D-488C-9C2B-7E2E75DF700D": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "F8EEBE8D-F17D-4556-B8DF-9BC11701B36D": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "F8EEBE8D-F17D-4556-B8DF-9BC11701B36D"
+}
diff --git a/320ms/individual_components.py b/320ms/individual_components.py
new file mode 100644
index 0000000000000000000000000000000000000000..47271397bc8d9d17cc0fabcf4bb63be7e7c2109c
--- /dev/null
+++ b/320ms/individual_components.py
@@ -0,0 +1,250 @@
+#!/usr/bin/env python3
+"""Export Parakeet Realtime EOU 120M RNNT components into CoreML."""
+from __future__ import annotations
+
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Optional, Tuple
+
+import coremltools as ct
+import torch
+
+
+@dataclass
+class ExportSettings:
+    output_dir: Path
+    compute_units: ct.ComputeUnit
+    deployment_target: Optional[ct.target]
+    compute_precision: Optional[ct.precision]
+    max_audio_seconds: float
+    max_symbol_steps: int
+
+
+class PreprocessorWrapper(torch.nn.Module):
+    """Wrapper for the audio preprocessor (mel spectrogram extraction)."""
+
+    def __init__(self, module: torch.nn.Module) -> None:
+        super().__init__()
+        self.module = module
+
+    def forward(
+        self, audio_signal: torch.Tensor, length: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        mel, mel_length = self.module(
+            input_signal=audio_signal, length=length.to(dtype=torch.long)
+        )
+        return mel, mel_length
+
+
+class EncoderWrapper(torch.nn.Module):
+    """Wrapper for the cache-aware FastConformer encoder.
+
+    Note: For the realtime EOU model, the encoder is cache-aware which means
+    it can operate in a streaming fashion. For CoreML export, we export
+    without cache state for simplicity (full-context mode).
+    """
+
+    def __init__(self, module: torch.nn.Module) -> None:
+        super().__init__()
+        self.module = module
+
+    def forward(
+        self, features: torch.Tensor, length: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        encoded, encoded_lengths = self.module(
+            audio_signal=features, length=length.to(dtype=torch.long)
+        )
+        # Synthesize per-frame timestamps (seconds) using the 80 ms encoder stride.
+        # Shape: [B, T_enc]
+        frame_times = (
+            torch.arange(encoded.shape[-1], device=encoded.device, dtype=torch.float32)
+            * 0.08
+        )
+        return encoded, encoded_lengths, frame_times
+
+
+class DecoderWrapper(torch.nn.Module):
+    """Wrapper for the RNNT prediction network (decoder)."""
+
+    def __init__(self, module: torch.nn.Module) -> None:
+        super().__init__()
+        self.module = module
+
+    def forward(
+        self,
+        targets: torch.Tensor,
+        target_lengths: torch.Tensor,
+        h_in: torch.Tensor,
+        c_in: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        state = [h_in, c_in]
+        decoder_output, _, new_state = self.module(
+            targets=targets.to(dtype=torch.long),
+            target_length=target_lengths.to(dtype=torch.long),
+            states=state,
+        )
+        return decoder_output, new_state[0], new_state[1]
+
+
+class JointWrapper(torch.nn.Module):
+    """Wrapper for the RNNT joint network.
+
+    Note: Unlike Parakeet TDT v3, the realtime EOU model does NOT have
+    duration outputs (num_extra_outputs). The joint network outputs only
+    token logits over the vocabulary + blank.
+    """
+
+    def __init__(self, module: torch.nn.Module) -> None:
+        super().__init__()
+        self.module = module
+
+    def forward(
+        self, encoder_outputs: torch.Tensor, decoder_outputs: torch.Tensor
+    ) -> torch.Tensor:
+        # Input: encoder_outputs [B, D, T], decoder_outputs [B, D, U]
+        # Transpose to match what projection layers expect
+        encoder_outputs = encoder_outputs.transpose(1, 2)  # [B, T, D]
+        decoder_outputs = decoder_outputs.transpose(1, 2)  # [B, U, D]
+
+        # Apply projections
+        enc_proj = self.module.enc(encoder_outputs)  # [B, T, joint_hidden]
+        dec_proj = self.module.pred(decoder_outputs)  # [B, U, joint_hidden]
+
+        # Explicit broadcasting along T and U to avoid converter ambiguity
+        x = enc_proj.unsqueeze(2) + dec_proj.unsqueeze(1)  # [B, T, U, joint_hidden]
+        x = self.module.joint_net[0](x)  # ReLU
+        x = self.module.joint_net[1](x)  # Dropout (no-op in eval)
+        out = self.module.joint_net[2](x)  # Linear -> logits [B, T, U, vocab+blank]
+        return out
+
+
+class MelEncoderWrapper(torch.nn.Module):
+    """Fused wrapper: waveform -> mel -> encoder.
+
+    Inputs:
+      - audio_signal: [B, S]
+      - audio_length: [B]
+
+    Outputs:
+      - encoder: [B, D, T_enc]
+      - encoder_length: [B]
+      - frame_times: [T_enc]
+    """
+
+    def __init__(
+        self, preprocessor: PreprocessorWrapper, encoder: EncoderWrapper
+    ) -> None:
+        super().__init__()
+        self.preprocessor = preprocessor
+        self.encoder = encoder
+
+    def forward(
+        self, audio_signal: torch.Tensor, audio_length: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        mel, mel_length = self.preprocessor(audio_signal, audio_length)
+        encoded, enc_len, frame_times = self.encoder(mel, mel_length.to(dtype=torch.int32))
+        return encoded, enc_len, frame_times
+
+
+class JointDecisionWrapper(torch.nn.Module):
+    """Joint + decision head: outputs label id and label prob.
+
+    Unlike Parakeet TDT v3, this model does NOT have duration outputs.
+
+    Inputs:
+      - encoder_outputs: [B, D, T]
+      - decoder_outputs: [B, D, U]
+
+    Returns:
+      - token_id: [B, T, U] int32
+      - token_prob: [B, T, U] float32
+    """
+
+    def __init__(self, joint: JointWrapper, vocab_size: int) -> None:
+        super().__init__()
+        self.joint = joint
+        self.vocab_with_blank = int(vocab_size) + 1
+
+    def forward(self, encoder_outputs: torch.Tensor, decoder_outputs: torch.Tensor):
+        logits = self.joint(encoder_outputs, decoder_outputs)
+
+        # Token selection
+        token_ids = torch.argmax(logits, dim=-1).to(dtype=torch.int32)
+        token_probs_all = torch.softmax(logits, dim=-1)
+        # gather expects int64 (long) indices; cast only for gather
+        token_prob = torch.gather(
+            token_probs_all, dim=-1, index=token_ids.long().unsqueeze(-1)
+        ).squeeze(-1)
+
+        return token_ids, token_prob
+
+
+class JointDecisionSingleStep(torch.nn.Module):
+    """Single-step variant for streaming: encoder_step -> token decision.
+
+    Inputs:
+      - encoder_step: [B=1, D, T=1]
+      - decoder_step: [B=1, D, U=1]
+
+    Returns:
+      - token_id: [1, 1, 1] int32
+      - token_prob: [1, 1, 1] float32
+      - top_k_ids: [1, 1, 1, K] int32
+      - top_k_logits: [1, 1, 1, K] float32
+    """
+
+    def __init__(self, joint: JointWrapper, vocab_size: int, top_k: int = 64) -> None:
+        super().__init__()
+        self.joint = joint
+        self.vocab_with_blank = int(vocab_size) + 1
+        self.top_k = int(top_k)
+
+    def forward(self, encoder_step: torch.Tensor, decoder_step: torch.Tensor):
+        # Reuse JointWrapper which expects [B, D, T] and [B, D, U]
+        logits = self.joint(encoder_step, decoder_step)  # [1, 1, 1, V+blank]
+
+        token_ids = torch.argmax(logits, dim=-1, keepdim=False).to(dtype=torch.int32)
+        token_probs_all = torch.softmax(logits, dim=-1)
+        token_prob = torch.gather(
+            token_probs_all, dim=-1, index=token_ids.long().unsqueeze(-1)
+        ).squeeze(-1)
+
+        # Also expose top-K candidates for host-side processing
+        topk_logits, topk_ids_long = torch.topk(
+            logits, k=min(self.top_k, logits.shape[-1]), dim=-1
+        )
+        topk_ids = topk_ids_long.to(dtype=torch.int32)
+        return token_ids, token_prob, topk_ids, topk_logits
+
+
+def _coreml_convert(
+    traced: torch.jit.ScriptModule,
+    inputs,
+    outputs,
+    settings: ExportSettings,
+    compute_units_override: Optional[ct.ComputeUnit] = None,
+    compute_precision: Optional[ct.precision] = None,
+) -> ct.models.MLModel:
+    cu = (
+        compute_units_override
+        if compute_units_override is not None
+        else settings.compute_units
+    )
+    kwargs = {
+        "convert_to": "mlprogram",
+        "inputs": inputs,
+        "outputs": outputs,
+        "compute_units": cu,
+    }
+    print("Converting:", traced.__class__.__name__)
+    print("Conversion kwargs:", kwargs)
+    if settings.deployment_target is not None:
+        kwargs["minimum_deployment_target"] = settings.deployment_target
+    
+    # Priority: explicit argument > settings
+    if compute_precision is not None:
+        kwargs["compute_precision"] = compute_precision
+    elif settings.compute_precision is not None:
+        kwargs["compute_precision"] = settings.compute_precision
+        
+    return ct.convert(traced, **kwargs)
diff --git a/320ms/joint_decision.mlmodelc/analytics/coremldata.bin b/320ms/joint_decision.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..ff9929e12bae25c509655bda0ebda6b9fe13fade
--- /dev/null
+++ b/320ms/joint_decision.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5bca32ad130dcad6605cc00044c752aa5b45ef57d14c17f2d1a2fa49d6cf55b5
+size 243
diff --git a/320ms/joint_decision.mlmodelc/coremldata.bin b/320ms/joint_decision.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..1aa5afe7c02e93f9baba2757cad5c21f563157bf
--- /dev/null
+++ b/320ms/joint_decision.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22d4abc4625b935ee035b5f8ce7cb28d1041b9b01c12173e287bf4b5f5d99625
+size 493
diff --git a/320ms/joint_decision.mlmodelc/metadata.json b/320ms/joint_decision.mlmodelc/metadata.json
new file mode 100644
index 0000000000000000000000000000000000000000..a0a319857f429f5ec4b106397cd52ec354ef2e96
--- /dev/null
+++ b/320ms/joint_decision.mlmodelc/metadata.json
@@ -0,0 +1,112 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Parakeet EOU single-step joint decision",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1]",
+        "name" : "token_id",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1]",
+        "name" : "token_prob",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1 × 1 × 64)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1, 64]",
+        "name" : "top_k_ids",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 1 × 64)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1, 64]",
+        "name" : "top_k_logits",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.reduceArgmax" : 1,
+      "Ios17.squeeze" : 1,
+      "Ios17.cast" : 6,
+      "Ios17.linear" : 3,
+      "Ios17.transpose" : 2,
+      "Ios17.add" : 1,
+      "Ios16.relu" : 1,
+      "Ios16.softmax" : 1,
+      "Ios17.gatherAlongAxis" : 1,
+      "Ios17.topk" : 1,
+      "Ios17.expandDims" : 3
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int16, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 1]",
+        "name" : "encoder_step",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 640 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 640, 1]",
+        "name" : "decoder_step",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.4.0"
+    },
+    "generatedClassName" : "parakeet_eou_joint_decision_single_step",
+    "method" : "predict"
+  }
+]
\ No newline at end of file
diff --git a/320ms/joint_decision.mlmodelc/model.mil b/320ms/joint_decision.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..172a7579866c84407b32b5f746b7f1ae132599d8
--- /dev/null
+++ b/320ms/joint_decision.mlmodelc/model.mil
@@ -0,0 +1,57 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.4.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<fp32, [1, 640, 1]> decoder_step, tensor<fp32, [1, 512, 1]> encoder_step) {
+            tensor<int32, [3]> input_1_perm_0 = const()[name = tensor<string, []>("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> encoder_step_to_fp16_dtype_0 = const()[name = tensor<string, []>("encoder_step_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> decoder_step_to_fp16_dtype_0 = const()[name = tensor<string, []>("decoder_step_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [640, 512]> joint_module_enc_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_enc_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [640]> joint_module_enc_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(655488)))];
+            tensor<fp16, [1, 512, 1]> encoder_step_to_fp16 = cast(dtype = encoder_step_to_fp16_dtype_0, x = encoder_step)[name = tensor<string, []>("cast_8")];
+            tensor<fp16, [1, 1, 512]> input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_step_to_fp16)[name = tensor<string, []>("transpose_1")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(656832)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1476096)))];
+            tensor<fp16, [1, 640, 1]> decoder_step_to_fp16 = cast(dtype = decoder_step_to_fp16_dtype_0, x = decoder_step)[name = tensor<string, []>("cast_7")];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_step_to_fp16)[name = tensor<string, []>("transpose_0")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_23_axes_0 = const()[name = tensor<string, []>("op_23_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_23_cast_fp16")];
+            tensor<int32, [1]> var_24_axes_0 = const()[name = tensor<string, []>("op_24_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_24_cast_fp16 = expand_dims(axes = var_24_axes_0, x = linear_1_cast_fp16)[name = tensor<string, []>("op_24_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_24_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<fp16, [1027, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [1027, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1477440)))];
+            tensor<fp16, [1027]> joint_module_joint_net_2_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [1027]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2792064)))];
+            tensor<fp16, [1, 1, 1, 1027]> linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<int32, []> var_38_axis_0 = const()[name = tensor<string, []>("op_38_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> var_38_keep_dims_0 = const()[name = tensor<string, []>("op_38_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> var_38_output_dtype_0 = const()[name = tensor<string, []>("op_38_output_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 1, 1]> token_id = reduce_argmax(axis = var_38_axis_0, keep_dims = var_38_keep_dims_0, output_dtype = var_38_output_dtype_0, x = linear_2_cast_fp16)[name = tensor<string, []>("op_38_cast_fp16")];
+            tensor<int32, []> var_44 = const()[name = tensor<string, []>("op_44"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [1, 1, 1, 1027]> token_probs_all_cast_fp16 = softmax(axis = var_44, x = linear_2_cast_fp16)[name = tensor<string, []>("token_probs_all_cast_fp16")];
+            tensor<int32, [1]> var_53_axes_0 = const()[name = tensor<string, []>("op_53_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 1, 1, 1]> var_53 = expand_dims(axes = var_53_axes_0, x = token_id)[name = tensor<string, []>("op_53")];
+            tensor<int32, []> var_54 = const()[name = tensor<string, []>("op_54"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> var_56_validate_indices_0 = const()[name = tensor<string, []>("op_56_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> var_53_to_int16_dtype_0 = const()[name = tensor<string, []>("op_53_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 1, 1, 1]> var_53_to_int16 = cast(dtype = var_53_to_int16_dtype_0, x = var_53)[name = tensor<string, []>("cast_6")];
+            tensor<fp16, [1, 1, 1, 1]> var_56_cast_fp16_cast_int16 = gather_along_axis(axis = var_54, indices = var_53_to_int16, validate_indices = var_56_validate_indices_0, x = token_probs_all_cast_fp16)[name = tensor<string, []>("op_56_cast_fp16_cast_int16")];
+            tensor<int32, [1]> var_58_axes_0 = const()[name = tensor<string, []>("op_58_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 1, 1]> var_58_cast_fp16 = squeeze(axes = var_58_axes_0, x = var_56_cast_fp16_cast_int16)[name = tensor<string, []>("op_58_cast_fp16")];
+            tensor<string, []> var_58_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_58_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, []> var_59 = const()[name = tensor<string, []>("op_59"), val = tensor<int32, []>(64)];
+            tensor<int32, []> var_63_axis_0 = const()[name = tensor<string, []>("op_63_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> var_63_ascending_0 = const()[name = tensor<string, []>("op_63_ascending_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_63_sort_0 = const()[name = tensor<string, []>("op_63_sort_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_63_return_indices_0 = const()[name = tensor<string, []>("op_63_return_indices_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> var_63_cast_fp16_cast_int16_output_indices_dtype_0 = const()[name = tensor<string, []>("op_63_cast_fp16_cast_int16_output_indices_dtype_0"), val = tensor<string, []>("uint16")];
+            tensor<fp16, [1, 1, 1, 64]> var_63_cast_fp16_cast_int16_0, tensor<uint16, [1, 1, 1, 64]> var_63_cast_fp16_cast_int16_1 = topk(ascending = var_63_ascending_0, axis = var_63_axis_0, k = var_59, output_indices_dtype = var_63_cast_fp16_cast_int16_output_indices_dtype_0, return_indices = var_63_return_indices_0, sort = var_63_sort_0, x = linear_2_cast_fp16)[name = tensor<string, []>("op_63_cast_fp16_cast_int16")];
+            tensor<string, []> var_63_cast_fp16_cast_int16_1_to_int32_dtype_0 = const()[name = tensor<string, []>("op_63_cast_fp16_cast_int16_1_to_int32_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<string, []> var_63_cast_fp16_0_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_63_cast_fp16_0_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 1, 1, 64]> top_k_logits = cast(dtype = var_63_cast_fp16_0_to_fp32_dtype_0, x = var_63_cast_fp16_cast_int16_0)[name = tensor<string, []>("cast_3")];
+            tensor<int32, [1, 1, 1, 64]> top_k_ids = cast(dtype = var_63_cast_fp16_cast_int16_1_to_int32_dtype_0, x = var_63_cast_fp16_cast_int16_1)[name = tensor<string, []>("cast_4")];
+            tensor<fp32, [1, 1, 1]> token_prob = cast(dtype = var_58_cast_fp16_to_fp32_dtype_0, x = var_58_cast_fp16)[name = tensor<string, []>("cast_5")];
+        } -> (token_id, token_prob, top_k_ids, top_k_logits);
+}
\ No newline at end of file
diff --git a/320ms/joint_decision.mlmodelc/weights/weight.bin b/320ms/joint_decision.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..ec3d2c1dea18b332fa1a3cd8d31981bae4e3f649
--- /dev/null
+++ b/320ms/joint_decision.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7039b2010a269153f5a96edf28637f921a86ef8822f248f2d6712f7a6bce84b4
+size 2794182
diff --git a/320ms/joint_decision.mlpackage/Data/com.apple.CoreML/model.mlmodel b/320ms/joint_decision.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..678be6fae0d594ff4bb923f973b0459139c8b370
--- /dev/null
+++ b/320ms/joint_decision.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25d4d7be6eeb60c7de1d3a1278a5a4700cbe34017e1a8c1cab33204ddb2e4d5e
+size 8701
diff --git a/320ms/joint_decision.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/320ms/joint_decision.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..ec3d2c1dea18b332fa1a3cd8d31981bae4e3f649
--- /dev/null
+++ b/320ms/joint_decision.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7039b2010a269153f5a96edf28637f921a86ef8822f248f2d6712f7a6bce84b4
+size 2794182
diff --git a/320ms/joint_decision.mlpackage/Manifest.json b/320ms/joint_decision.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..fffd90a77691d335c8153fd9c38d49adb725c949
--- /dev/null
+++ b/320ms/joint_decision.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "634E266B-4447-41D3-879E-F3611888F54B": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "C7F40527-180B-45CD-BC12-4F054F2E5D9A": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "C7F40527-180B-45CD-BC12-4F054F2E5D9A"
+}
diff --git a/320ms/streaming_encoder.mlmodelc/analytics/coremldata.bin b/320ms/streaming_encoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..eec5bd02b4d1075885ea77fb1d5526138526e198
--- /dev/null
+++ b/320ms/streaming_encoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae0a559f8d53c43fe13c5c142b15ca237b97812febde71a36475b7b9612baec5
+size 243
diff --git a/320ms/streaming_encoder.mlmodelc/coremldata.bin b/320ms/streaming_encoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..4bd36c23d7283099ead6c7857659a4e1a4f70dc5
--- /dev/null
+++ b/320ms/streaming_encoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b57dfe32ed2acb0081f97d7a9e410117005823e3d2d995e21feb9e60775443c3
+size 670
diff --git a/320ms/streaming_encoder.mlmodelc/metadata.json b/320ms/streaming_encoder.mlmodelc/metadata.json
new file mode 100644
index 0000000000000000000000000000000000000000..4830f432c8a7079ceced0f4dbb661a775224824d
--- /dev/null
+++ b/320ms/streaming_encoder.mlmodelc/metadata.json
@@ -0,0 +1,187 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512 × 8)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 8]",
+        "name" : "encoded_output",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "encoded_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128 × 9)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 9]",
+        "name" : "new_pre_cache",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 17 × 1 × 70 × 512)",
+        "shortDescription" : "",
+        "shape" : "[17, 1, 70, 512]",
+        "name" : "new_cache_last_channel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 17 × 1 × 512 × 8)",
+        "shortDescription" : "",
+        "shape" : "[17, 1, 512, 8]",
+        "name" : "new_cache_last_time",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "new_cache_last_channel_len",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.logicalAnd" : 3,
+      "Ios17.reshape" : 103,
+      "Ios16.softmax" : 17,
+      "Ios17.matmul" : 51,
+      "Ios17.transpose" : 157,
+      "Split" : 17,
+      "Ios17.expandDims" : 18,
+      "Select" : 51,
+      "Ios17.add" : 132,
+      "Tile" : 8,
+      "Ios17.sliceByIndex" : 140,
+      "Ios16.sigmoid" : 17,
+      "Pad" : 20,
+      "Ios17.logicalNot" : 2,
+      "Ios17.layerNorm" : 102,
+      "Ios17.less" : 5,
+      "Ios17.sub" : 4,
+      "Ios17.conv" : 56,
+      "Ios16.relu" : 3,
+      "Ios17.clip" : 2,
+      "Ios17.linear" : 137,
+      "Ios17.concat" : 52,
+      "Ios17.floorDiv" : 3,
+      "Ios17.greaterEqual" : 1,
+      "Ios17.cast" : 20,
+      "Ios16.silu" : 51,
+      "Stack" : 2,
+      "Ios17.mul" : 78
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.4.0",
+      "com.github.apple.coremltools.version" : "8.3.0"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128 × 64)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 64]",
+        "name" : "audio_signal",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "audio_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128 × 9)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 9]",
+        "name" : "pre_cache",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 17 × 1 × 70 × 512)",
+        "shortDescription" : "",
+        "shape" : "[17, 1, 70, 512]",
+        "name" : "cache_last_channel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 17 × 1 × 512 × 8)",
+        "shortDescription" : "",
+        "shape" : "[17, 1, 512, 8]",
+        "name" : "cache_last_time",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "cache_last_channel_len",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "streaming_encoder",
+    "method" : "predict"
+  }
+]
\ No newline at end of file
diff --git a/320ms/streaming_encoder.mlmodelc/model.mil b/320ms/streaming_encoder.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..7902d6696813c64c694517418f9a9bdd2ee40119
--- /dev/null
+++ b/320ms/streaming_encoder.mlmodelc/model.mil
@@ -0,0 +1,3195 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.4.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<int32, [1]> audio_length, tensor<fp32, [1, 128, 64]> audio_signal, tensor<fp32, [17, 1, 70, 512]> cache_last_channel, tensor<int32, [1]> cache_last_channel_len, tensor<fp32, [17, 1, 512, 8]> cache_last_time, tensor<fp32, [1, 128, 9]> pre_cache) {
+            tensor<int32, []> var_9 = const()[name = tensor<string, []>("op_9"), val = tensor<int32, []>(2)];
+            tensor<bool, []> full_input_interleave_0 = const()[name = tensor<string, []>("full_input_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> pre_cache_to_fp16_dtype_0 = const()[name = tensor<string, []>("pre_cache_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<string, []> audio_signal_to_fp16_dtype_0 = const()[name = tensor<string, []>("audio_signal_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 128, 64]> audio_signal_to_fp16 = cast(dtype = audio_signal_to_fp16_dtype_0, x = audio_signal)[name = tensor<string, []>("cast_199")];
+            tensor<fp16, [1, 128, 9]> pre_cache_to_fp16 = cast(dtype = pre_cache_to_fp16_dtype_0, x = pre_cache)[name = tensor<string, []>("cast_200")];
+            tensor<fp16, [1, 128, 73]> full_input_cast_fp16 = concat(axis = var_9, interleave = full_input_interleave_0, values = (pre_cache_to_fp16, audio_signal_to_fp16))[name = tensor<string, []>("full_input_cast_fp16")];
+            tensor<int32, []> var_12 = const()[name = tensor<string, []>("op_12"), val = tensor<int32, []>(9)];
+            tensor<int32, [1]> value_1 = add(x = audio_length, y = var_12)[name = tensor<string, []>("value_1")];
+            tensor<int32, [3]> var_28_begin_0 = const()[name = tensor<string, []>("op_28_begin_0"), val = tensor<int32, [3]>([0, 0, 64])];
+            tensor<int32, [3]> var_28_end_0 = const()[name = tensor<string, []>("op_28_end_0"), val = tensor<int32, [3]>([1, 128, 73])];
+            tensor<bool, [3]> var_28_end_mask_0 = const()[name = tensor<string, []>("op_28_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 128, 9]> var_28_cast_fp16 = slice_by_index(begin = var_28_begin_0, end = var_28_end_0, end_mask = var_28_end_mask_0, x = full_input_cast_fp16)[name = tensor<string, []>("op_28_cast_fp16")];
+            tensor<string, []> var_28_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_28_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, []> var_56 = const()[name = tensor<string, []>("op_56"), val = tensor<int32, []>(-1)];
+            tensor<int32, []> var_65 = const()[name = tensor<string, []>("op_65"), val = tensor<int32, []>(1)];
+            tensor<int32, [3]> x_1_perm_0 = const()[name = tensor<string, []>("x_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> tensor_1_axes_0 = const()[name = tensor<string, []>("tensor_1_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 73, 128]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = full_input_cast_fp16)[name = tensor<string, []>("transpose_241")];
+            tensor<fp16, [1, 1, 73, 128]> tensor_1_cast_fp16 = expand_dims(axes = tensor_1_axes_0, x = x_1_cast_fp16)[name = tensor<string, []>("tensor_1_cast_fp16")];
+            tensor<int32, [1, 73]> expand_dims_0 = const()[name = tensor<string, []>("expand_dims_0"), val = tensor<int32, [1, 73]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72]])];
+            tensor<int32, [1]> var_120_axes_0 = const()[name = tensor<string, []>("op_120_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_120 = expand_dims(axes = var_120_axes_0, x = value_1)[name = tensor<string, []>("op_120")];
+            tensor<bool, [1, 73]> time_mask_1 = less(x = expand_dims_0, y = var_120)[name = tensor<string, []>("time_mask_1")];
+            tensor<int32, [1]> var_122_axes_0 = const()[name = tensor<string, []>("op_122_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 73, 1]> var_122 = expand_dims(axes = var_122_axes_0, x = time_mask_1)[name = tensor<string, []>("op_122")];
+            tensor<int32, [3]> var_124_reps_0 = const()[name = tensor<string, []>("op_124_reps_0"), val = tensor<int32, [3]>([1, 1, 128])];
+            tensor<bool, [1, 73, 128]> var_124 = tile(reps = var_124_reps_0, x = var_122)[name = tensor<string, []>("op_124")];
+            tensor<int32, [1]> var_130_axes_0 = const()[name = tensor<string, []>("op_130_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<string, []> cast_2_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 73, 128]> var_124_to_fp16 = cast(dtype = cast_2_to_fp16_dtype_0, x = var_124)[name = tensor<string, []>("cast_197")];
+            tensor<fp16, [1, 1, 73, 128]> var_130_cast_fp16 = expand_dims(axes = var_130_axes_0, x = var_124_to_fp16)[name = tensor<string, []>("op_130_cast_fp16")];
+            tensor<fp16, [1, 1, 73, 128]> input_1_cast_fp16 = mul(x = tensor_1_cast_fp16, y = var_130_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
+            tensor<int32, [8]> input_3_pad_0 = const()[name = tensor<string, []>("input_3_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            tensor<string, []> input_3_mode_0 = const()[name = tensor<string, []>("input_3_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_9_to_fp16 = const()[name = tensor<string, []>("const_9_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 1, 76, 131]> input_3_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<string, []> tensor_3_pad_type_0 = const()[name = tensor<string, []>("tensor_3_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_3_strides_0 = const()[name = tensor<string, []>("tensor_3_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, [4]> tensor_3_pad_0 = const()[name = tensor<string, []>("tensor_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_3_dilations_0 = const()[name = tensor<string, []>("tensor_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> tensor_3_groups_0 = const()[name = tensor<string, []>("tensor_3_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_0_weight_to_fp16"), val = tensor<fp16, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4736)))];
+            tensor<fp16, [1, 256, 37, 65]> tensor_3_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = tensor_3_dilations_0, groups = tensor_3_groups_0, pad = tensor_3_pad_0, pad_type = tensor_3_pad_type_0, strides = tensor_3_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("tensor_3_cast_fp16")];
+            tensor<string, []> cast_0_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_0_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> var_143_promoted_to_fp16 = const()[name = tensor<string, []>("op_143_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1]> value_1_to_fp16 = cast(dtype = cast_0_to_fp16_dtype_0, x = value_1)[name = tensor<string, []>("cast_196")];
+            tensor<fp16, [1]> var_144_cast_fp16 = add(x = value_1_to_fp16, y = var_143_promoted_to_fp16)[name = tensor<string, []>("op_144_cast_fp16")];
+            tensor<fp16, []> var_145_promoted_to_fp16 = const()[name = tensor<string, []>("op_145_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> var_146_cast_fp16 = add(x = var_144_cast_fp16, y = var_145_promoted_to_fp16)[name = tensor<string, []>("op_146_cast_fp16")];
+            tensor<fp16, []> var_147_promoted_to_fp16 = const()[name = tensor<string, []>("op_147_promoted_to_fp16"), val = tensor<fp16, []>(0x1.8p+1)];
+            tensor<fp16, [1]> var_148_cast_fp16 = sub(x = var_146_cast_fp16, y = var_147_promoted_to_fp16)[name = tensor<string, []>("op_148_cast_fp16")];
+            tensor<fp16, []> var_53_promoted_to_fp16 = const()[name = tensor<string, []>("op_53_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1]> floor_div_0_cast_fp16 = floor_div(x = var_148_cast_fp16, y = var_53_promoted_to_fp16)[name = tensor<string, []>("floor_div_0_cast_fp16")];
+            tensor<fp16, []> var_150_promoted_to_fp16 = const()[name = tensor<string, []>("op_150_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_3_cast_fp16 = add(x = floor_div_0_cast_fp16, y = var_150_promoted_to_fp16)[name = tensor<string, []>("current_lengths_3_cast_fp16")];
+            tensor<string, []> cast_3_dtype_0 = const()[name = tensor<string, []>("cast_3_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 37]> expand_dims_1 = const()[name = tensor<string, []>("expand_dims_1"), val = tensor<int32, [1, 37]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36]])];
+            tensor<int32, [1]> var_159_axes_0 = const()[name = tensor<string, []>("op_159_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_3_cast_fp16_to_int32 = cast(dtype = cast_3_dtype_0, x = current_lengths_3_cast_fp16)[name = tensor<string, []>("cast_195")];
+            tensor<int32, [1, 1]> var_159 = expand_dims(axes = var_159_axes_0, x = current_lengths_3_cast_fp16_to_int32)[name = tensor<string, []>("op_159")];
+            tensor<bool, [1, 37]> time_mask_3 = less(x = expand_dims_1, y = var_159)[name = tensor<string, []>("time_mask_3")];
+            tensor<int32, [1]> var_161_axes_0 = const()[name = tensor<string, []>("op_161_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 37, 1]> var_161 = expand_dims(axes = var_161_axes_0, x = time_mask_3)[name = tensor<string, []>("op_161")];
+            tensor<int32, [3]> var_163_reps_0 = const()[name = tensor<string, []>("op_163_reps_0"), val = tensor<int32, [3]>([1, 1, 65])];
+            tensor<bool, [1, 37, 65]> var_163 = tile(reps = var_163_reps_0, x = var_161)[name = tensor<string, []>("op_163")];
+            tensor<int32, [1]> var_169_axes_0 = const()[name = tensor<string, []>("op_169_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<string, []> cast_4_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_4_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 37, 65]> var_163_to_fp16 = cast(dtype = cast_4_to_fp16_dtype_0, x = var_163)[name = tensor<string, []>("cast_194")];
+            tensor<fp16, [1, 1, 37, 65]> var_169_cast_fp16 = expand_dims(axes = var_169_axes_0, x = var_163_to_fp16)[name = tensor<string, []>("op_169_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_3_reps_0 = const()[name = tensor<string, []>("expanded_mask_3_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 37, 65]> expanded_mask_3_cast_fp16 = tile(reps = expanded_mask_3_reps_0, x = var_169_cast_fp16)[name = tensor<string, []>("expanded_mask_3_cast_fp16")];
+            tensor<fp16, [1, 256, 37, 65]> input_5_cast_fp16 = mul(x = tensor_3_cast_fp16, y = expanded_mask_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [1, 256, 37, 65]> tensor_5_cast_fp16 = relu(x = input_5_cast_fp16)[name = tensor<string, []>("tensor_5_cast_fp16")];
+            tensor<fp16, [1, 256, 37, 65]> input_7_cast_fp16 = mul(x = tensor_5_cast_fp16, y = expanded_mask_3_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<int32, [8]> input_9_pad_0 = const()[name = tensor<string, []>("input_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            tensor<string, []> input_9_mode_0 = const()[name = tensor<string, []>("input_9_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_23_to_fp16 = const()[name = tensor<string, []>("const_23_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 256, 40, 68]> input_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
+            tensor<string, []> tensor_7_pad_type_0 = const()[name = tensor<string, []>("tensor_7_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_7_strides_0 = const()[name = tensor<string, []>("tensor_7_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, []> tensor_7_groups_0 = const()[name = tensor<string, []>("tensor_7_groups_0"), val = tensor<int32, []>(256)];
+            tensor<int32, [4]> tensor_7_pad_0 = const()[name = tensor<string, []>("tensor_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_7_dilations_0 = const()[name = tensor<string, []>("tensor_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_2_weight_to_fp16"), val = tensor<fp16, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5312)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9984)))];
+            tensor<fp16, [1, 256, 19, 33]> tensor_7_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = tensor_7_dilations_0, groups = tensor_7_groups_0, pad = tensor_7_pad_0, pad_type = tensor_7_pad_type_0, strides = tensor_7_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("tensor_7_cast_fp16")];
+            tensor<fp16, []> var_191_promoted_to_fp16 = const()[name = tensor<string, []>("op_191_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1]> var_192_cast_fp16 = add(x = current_lengths_3_cast_fp16, y = var_191_promoted_to_fp16)[name = tensor<string, []>("op_192_cast_fp16")];
+            tensor<fp16, []> var_193_promoted_to_fp16 = const()[name = tensor<string, []>("op_193_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> var_194_cast_fp16 = add(x = var_192_cast_fp16, y = var_193_promoted_to_fp16)[name = tensor<string, []>("op_194_cast_fp16")];
+            tensor<fp16, []> var_195_promoted_to_fp16 = const()[name = tensor<string, []>("op_195_promoted_to_fp16"), val = tensor<fp16, []>(0x1.8p+1)];
+            tensor<fp16, [1]> var_196_cast_fp16 = sub(x = var_194_cast_fp16, y = var_195_promoted_to_fp16)[name = tensor<string, []>("op_196_cast_fp16")];
+            tensor<fp16, []> var_53_promoted_1_to_fp16 = const()[name = tensor<string, []>("op_53_promoted_1_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1]> floor_div_1_cast_fp16 = floor_div(x = var_196_cast_fp16, y = var_53_promoted_1_to_fp16)[name = tensor<string, []>("floor_div_1_cast_fp16")];
+            tensor<fp16, []> var_198_promoted_to_fp16 = const()[name = tensor<string, []>("op_198_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_5_cast_fp16 = add(x = floor_div_1_cast_fp16, y = var_198_promoted_to_fp16)[name = tensor<string, []>("current_lengths_5_cast_fp16")];
+            tensor<string, []> cast_5_dtype_0 = const()[name = tensor<string, []>("cast_5_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 19]> expand_dims_2 = const()[name = tensor<string, []>("expand_dims_2"), val = tensor<int32, [1, 19]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18]])];
+            tensor<int32, [1]> var_207_axes_0 = const()[name = tensor<string, []>("op_207_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_5_cast_fp16_to_int32 = cast(dtype = cast_5_dtype_0, x = current_lengths_5_cast_fp16)[name = tensor<string, []>("cast_193")];
+            tensor<int32, [1, 1]> var_207 = expand_dims(axes = var_207_axes_0, x = current_lengths_5_cast_fp16_to_int32)[name = tensor<string, []>("op_207")];
+            tensor<bool, [1, 19]> time_mask_5 = less(x = expand_dims_2, y = var_207)[name = tensor<string, []>("time_mask_5")];
+            tensor<int32, [1]> var_209_axes_0 = const()[name = tensor<string, []>("op_209_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 19, 1]> var_209 = expand_dims(axes = var_209_axes_0, x = time_mask_5)[name = tensor<string, []>("op_209")];
+            tensor<int32, [3]> var_211_reps_0 = const()[name = tensor<string, []>("op_211_reps_0"), val = tensor<int32, [3]>([1, 1, 33])];
+            tensor<bool, [1, 19, 33]> var_211 = tile(reps = var_211_reps_0, x = var_209)[name = tensor<string, []>("op_211")];
+            tensor<int32, [1]> var_217_axes_0 = const()[name = tensor<string, []>("op_217_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<string, []> cast_6_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_6_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 19, 33]> var_211_to_fp16 = cast(dtype = cast_6_to_fp16_dtype_0, x = var_211)[name = tensor<string, []>("cast_192")];
+            tensor<fp16, [1, 1, 19, 33]> var_217_cast_fp16 = expand_dims(axes = var_217_axes_0, x = var_211_to_fp16)[name = tensor<string, []>("op_217_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_7_reps_0 = const()[name = tensor<string, []>("expanded_mask_7_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 19, 33]> expanded_mask_7_cast_fp16 = tile(reps = expanded_mask_7_reps_0, x = var_217_cast_fp16)[name = tensor<string, []>("expanded_mask_7_cast_fp16")];
+            tensor<fp16, [1, 256, 19, 33]> input_11_cast_fp16 = mul(x = tensor_7_cast_fp16, y = expanded_mask_7_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<string, []> tensor_9_pad_type_0 = const()[name = tensor<string, []>("tensor_9_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_9_strides_0 = const()[name = tensor<string, []>("tensor_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> tensor_9_pad_0 = const()[name = tensor<string, []>("tensor_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_9_dilations_0 = const()[name = tensor<string, []>("tensor_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> tensor_9_groups_0 = const()[name = tensor<string, []>("tensor_9_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_3_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_3_weight_to_fp16"), val = tensor<fp16, [256, 256, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10560)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141696)))];
+            tensor<fp16, [1, 256, 19, 33]> tensor_9_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = tensor_9_dilations_0, groups = tensor_9_groups_0, pad = tensor_9_pad_0, pad_type = tensor_9_pad_type_0, strides = tensor_9_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("tensor_9_cast_fp16")];
+            tensor<fp16, [1, 256, 19, 33]> input_13_cast_fp16 = mul(x = tensor_9_cast_fp16, y = expanded_mask_7_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
+            tensor<fp16, [1, 256, 19, 33]> tensor_11_cast_fp16 = relu(x = input_13_cast_fp16)[name = tensor<string, []>("tensor_11_cast_fp16")];
+            tensor<fp16, [1, 256, 19, 33]> input_15_cast_fp16 = mul(x = tensor_11_cast_fp16, y = expanded_mask_7_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
+            tensor<int32, [8]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            tensor<string, []> input_17_mode_0 = const()[name = tensor<string, []>("input_17_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_41_to_fp16 = const()[name = tensor<string, []>("const_41_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 256, 22, 36]> input_17_cast_fp16 = pad(constant_val = const_41_to_fp16, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
+            tensor<string, []> tensor_13_pad_type_0 = const()[name = tensor<string, []>("tensor_13_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_13_strides_0 = const()[name = tensor<string, []>("tensor_13_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, []> tensor_13_groups_0 = const()[name = tensor<string, []>("tensor_13_groups_0"), val = tensor<int32, []>(256)];
+            tensor<int32, [4]> tensor_13_pad_0 = const()[name = tensor<string, []>("tensor_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_13_dilations_0 = const()[name = tensor<string, []>("tensor_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_5_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_5_weight_to_fp16"), val = tensor<fp16, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142272)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146944)))];
+            tensor<fp16, [1, 256, 10, 17]> tensor_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = tensor_13_dilations_0, groups = tensor_13_groups_0, pad = tensor_13_pad_0, pad_type = tensor_13_pad_type_0, strides = tensor_13_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("tensor_13_cast_fp16")];
+            tensor<fp16, []> var_254_promoted_to_fp16 = const()[name = tensor<string, []>("op_254_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1]> var_255_cast_fp16 = add(x = current_lengths_5_cast_fp16, y = var_254_promoted_to_fp16)[name = tensor<string, []>("op_255_cast_fp16")];
+            tensor<fp16, []> var_256_promoted_to_fp16 = const()[name = tensor<string, []>("op_256_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> var_257_cast_fp16 = add(x = var_255_cast_fp16, y = var_256_promoted_to_fp16)[name = tensor<string, []>("op_257_cast_fp16")];
+            tensor<fp16, []> var_258_promoted_to_fp16 = const()[name = tensor<string, []>("op_258_promoted_to_fp16"), val = tensor<fp16, []>(0x1.8p+1)];
+            tensor<fp16, [1]> var_259_cast_fp16 = sub(x = var_257_cast_fp16, y = var_258_promoted_to_fp16)[name = tensor<string, []>("op_259_cast_fp16")];
+            tensor<fp16, []> var_53_promoted_2_to_fp16 = const()[name = tensor<string, []>("op_53_promoted_2_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1]> floor_div_2_cast_fp16 = floor_div(x = var_259_cast_fp16, y = var_53_promoted_2_to_fp16)[name = tensor<string, []>("floor_div_2_cast_fp16")];
+            tensor<fp16, []> var_261_promoted_to_fp16 = const()[name = tensor<string, []>("op_261_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_cast_fp16 = add(x = floor_div_2_cast_fp16, y = var_261_promoted_to_fp16)[name = tensor<string, []>("current_lengths_cast_fp16")];
+            tensor<string, []> cast_7_dtype_0 = const()[name = tensor<string, []>("cast_7_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 10]> expand_dims_3 = const()[name = tensor<string, []>("expand_dims_3"), val = tensor<int32, [1, 10]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]])];
+            tensor<int32, [1]> var_270_axes_0 = const()[name = tensor<string, []>("op_270_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_cast_fp16_to_int32 = cast(dtype = cast_7_dtype_0, x = current_lengths_cast_fp16)[name = tensor<string, []>("cast_191")];
+            tensor<int32, [1, 1]> var_270 = expand_dims(axes = var_270_axes_0, x = current_lengths_cast_fp16_to_int32)[name = tensor<string, []>("op_270")];
+            tensor<bool, [1, 10]> time_mask = less(x = expand_dims_3, y = var_270)[name = tensor<string, []>("time_mask")];
+            tensor<int32, [1]> var_272_axes_0 = const()[name = tensor<string, []>("op_272_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 10, 1]> var_272 = expand_dims(axes = var_272_axes_0, x = time_mask)[name = tensor<string, []>("op_272")];
+            tensor<int32, [3]> var_274_reps_0 = const()[name = tensor<string, []>("op_274_reps_0"), val = tensor<int32, [3]>([1, 1, 17])];
+            tensor<bool, [1, 10, 17]> var_274 = tile(reps = var_274_reps_0, x = var_272)[name = tensor<string, []>("op_274")];
+            tensor<int32, [1]> var_280_axes_0 = const()[name = tensor<string, []>("op_280_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<string, []> cast_8_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_8_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 10, 17]> var_274_to_fp16 = cast(dtype = cast_8_to_fp16_dtype_0, x = var_274)[name = tensor<string, []>("cast_190")];
+            tensor<fp16, [1, 1, 10, 17]> var_280_cast_fp16 = expand_dims(axes = var_280_axes_0, x = var_274_to_fp16)[name = tensor<string, []>("op_280_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_13_reps_0 = const()[name = tensor<string, []>("expanded_mask_13_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 10, 17]> expanded_mask_13_cast_fp16 = tile(reps = expanded_mask_13_reps_0, x = var_280_cast_fp16)[name = tensor<string, []>("expanded_mask_13_cast_fp16")];
+            tensor<fp16, [1, 256, 10, 17]> input_19_cast_fp16 = mul(x = tensor_13_cast_fp16, y = expanded_mask_13_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
+            tensor<string, []> tensor_15_pad_type_0 = const()[name = tensor<string, []>("tensor_15_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_15_strides_0 = const()[name = tensor<string, []>("tensor_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> tensor_15_pad_0 = const()[name = tensor<string, []>("tensor_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_15_dilations_0 = const()[name = tensor<string, []>("tensor_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> tensor_15_groups_0 = const()[name = tensor<string, []>("tensor_15_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_6_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_6_weight_to_fp16"), val = tensor<fp16, [256, 256, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147520)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278656)))];
+            tensor<fp16, [1, 256, 10, 17]> tensor_15_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = tensor_15_dilations_0, groups = tensor_15_groups_0, pad = tensor_15_pad_0, pad_type = tensor_15_pad_type_0, strides = tensor_15_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("tensor_15_cast_fp16")];
+            tensor<fp16, [1, 256, 10, 17]> input_21_cast_fp16 = mul(x = tensor_15_cast_fp16, y = expanded_mask_13_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
+            tensor<fp16, [1, 256, 10, 17]> tensor_cast_fp16 = relu(x = input_21_cast_fp16)[name = tensor<string, []>("tensor_cast_fp16")];
+            tensor<fp16, [1, 256, 10, 17]> x_3_cast_fp16 = mul(x = tensor_cast_fp16, y = expanded_mask_13_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
+            tensor<int32, [4]> var_314_perm_0 = const()[name = tensor<string, []>("op_314_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_315 = const()[name = tensor<string, []>("op_315"), val = tensor<int32, [3]>([1, 10, -1])];
+            tensor<fp16, [1, 10, 256, 17]> var_314_cast_fp16 = transpose(perm = var_314_perm_0, x = x_3_cast_fp16)[name = tensor<string, []>("transpose_240")];
+            tensor<fp16, [1, 10, 4352]> input_23_cast_fp16 = reshape(shape = var_315, x = var_314_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
+            tensor<fp16, [512, 4352]> encoder_pre_encode_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_out_weight_to_fp16"), val = tensor<fp16, [512, 4352]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279232)))];
+            tensor<fp16, [512]> encoder_pre_encode_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4735744)))];
+            tensor<fp16, [1, 10, 512]> linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<int32, [3]> var_325_begin_0 = const()[name = tensor<string, []>("op_325_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_325_end_0 = const()[name = tensor<string, []>("op_325_end_0"), val = tensor<int32, [3]>([1, 10, 512])];
+            tensor<bool, [3]> var_325_end_mask_0 = const()[name = tensor<string, []>("op_325_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 8, 512]> var_325_cast_fp16 = slice_by_index(begin = var_325_begin_0, end = var_325_end_0, end_mask = var_325_end_mask_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_325_cast_fp16")];
+            tensor<int32, []> var_327 = const()[name = tensor<string, []>("op_327"), val = tensor<int32, []>(2)];
+            tensor<int32, [1]> var_328 = sub(x = current_lengths_cast_fp16_to_int32, y = var_327)[name = tensor<string, []>("op_328")];
+            tensor<string, []> var_328_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_328_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> var_59_promoted_to_fp16 = const()[name = tensor<string, []>("op_59_promoted_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, []> const_61_to_fp16 = const()[name = tensor<string, []>("const_61_to_fp16"), val = tensor<fp16, []>(inf)];
+            tensor<fp16, [1]> var_328_to_fp16 = cast(dtype = var_328_promoted_to_fp16_dtype_0, x = var_328)[name = tensor<string, []>("cast_189")];
+            tensor<fp16, [1]> clip_0_cast_fp16 = clip(alpha = var_59_promoted_to_fp16, beta = const_61_to_fp16, x = var_328_to_fp16)[name = tensor<string, []>("clip_0_cast_fp16")];
+            tensor<int32, [1]> cache_keep_size = const()[name = tensor<string, []>("cache_keep_size"), val = tensor<int32, [1]>([4])];
+            tensor<fp16, []> var_344_promoted_to_fp16 = const()[name = tensor<string, []>("op_344_promoted_to_fp16"), val = tensor<fp16, []>(0x1.18p+6)];
+            tensor<fp16, [1]> padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_344_promoted_to_fp16)[name = tensor<string, []>("padding_length_cast_fp16")];
+            tensor<int32, []> const_63 = const()[name = tensor<string, []>("const_63"), val = tensor<int32, []>(-1)];
+            tensor<int32, [1]> var_346 = mul(x = cache_last_channel_len, y = const_63)[name = tensor<string, []>("op_346")];
+            tensor<int32, []> var_347 = const()[name = tensor<string, []>("op_347"), val = tensor<int32, []>(70)];
+            tensor<int32, [1]> offset = add(x = var_346, y = var_347)[name = tensor<string, []>("offset")];
+            tensor<int32, [1]> var_387_axes_0 = const()[name = tensor<string, []>("op_387_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 1]> var_387_cast_fp16 = expand_dims(axes = var_387_axes_0, x = padding_length_cast_fp16)[name = tensor<string, []>("op_387_cast_fp16")];
+            tensor<fp16, [1, 78]> var_386_promoted_to_fp16 = const()[name = tensor<string, []>("op_386_promoted_to_fp16"), val = tensor<fp16, [1, 78]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4736832)))];
+            tensor<bool, [1, 78]> pad_mask_1_cast_fp16 = less(x = var_386_promoted_to_fp16, y = var_387_cast_fp16)[name = tensor<string, []>("pad_mask_1_cast_fp16")];
+            tensor<int32, [1, 78]> expand_dims_5 = const()[name = tensor<string, []>("expand_dims_5"), val = tensor<int32, [1, 78]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77]])];
+            tensor<int32, [1]> var_393_axes_0 = const()[name = tensor<string, []>("op_393_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 1]> var_393 = expand_dims(axes = var_393_axes_0, x = offset)[name = tensor<string, []>("op_393")];
+            tensor<bool, [1, 78]> pad_mask_off = greater_equal(x = expand_dims_5, y = var_393)[name = tensor<string, []>("pad_mask_off")];
+            tensor<bool, [1, 78]> pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = tensor<string, []>("pad_mask_3")];
+            tensor<int32, [1]> var_396_axes_0 = const()[name = tensor<string, []>("op_396_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 78]> var_396 = expand_dims(axes = var_396_axes_0, x = pad_mask_3)[name = tensor<string, []>("op_396")];
+            tensor<int32, [3]> var_397 = const()[name = tensor<string, []>("op_397"), val = tensor<int32, [3]>([1, 78, 1])];
+            tensor<bool, [1, 78, 78]> pad_mask_for_att_mask_1 = tile(reps = var_397, x = var_396)[name = tensor<string, []>("pad_mask_for_att_mask_1")];
+            tensor<int32, [3]> var_399_perm_0 = const()[name = tensor<string, []>("op_399_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<bool, [1, 78, 78]> var_399 = transpose(perm = var_399_perm_0, x = pad_mask_for_att_mask_1)[name = tensor<string, []>("transpose_239")];
+            tensor<bool, [1, 78, 78]> pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_399)[name = tensor<string, []>("pad_mask_for_att_mask")];
+            tensor<bool, [1, 78, 78]> const_71 = const()[name = tensor<string, []>("const_71"), val = tensor<bool, [1, 78, 78]>([[[true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false], [false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false], [false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false], [false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false], [false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false], [false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])];
+            tensor<bool, [1, 78, 78]> att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_71)[name = tensor<string, []>("att_mask_9")];
+            tensor<bool, [1, 78, 78]> att_mask = logical_not(x = att_mask_9)[name = tensor<string, []>("att_mask")];
+            tensor<bool, [1, 78]> pad_mask_5 = logical_not(x = pad_mask_3)[name = tensor<string, []>("pad_mask_5")];
+            tensor<int32, [2]> pad_mask_begin_0 = const()[name = tensor<string, []>("pad_mask_begin_0"), val = tensor<int32, [2]>([0, 70])];
+            tensor<int32, [2]> pad_mask_end_0 = const()[name = tensor<string, []>("pad_mask_end_0"), val = tensor<int32, [2]>([1, 78])];
+            tensor<bool, [2]> pad_mask_end_mask_0 = const()[name = tensor<string, []>("pad_mask_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<bool, [1, 8]> pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = tensor<string, []>("pad_mask")];
+            tensor<int32, [3]> mask_9_begin_0 = const()[name = tensor<string, []>("mask_9_begin_0"), val = tensor<int32, [3]>([0, 70, 0])];
+            tensor<int32, [3]> mask_9_end_0 = const()[name = tensor<string, []>("mask_9_end_0"), val = tensor<int32, [3]>([1, 78, 78])];
+            tensor<bool, [3]> mask_9_end_mask_0 = const()[name = tensor<string, []>("mask_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<bool, [1, 8, 78]> mask_9 = slice_by_index(begin = mask_9_begin_0, end = mask_9_end_0, end_mask = mask_9_end_mask_0, x = att_mask)[name = tensor<string, []>("mask_9")];
+            tensor<int32, [4]> cache_1_begin_0 = const()[name = tensor<string, []>("cache_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_1_end_0 = const()[name = tensor<string, []>("cache_1_end_0"), val = tensor<int32, [4]>([1, 1, 70, 512])];
+            tensor<bool, [4]> cache_1_end_mask_0 = const()[name = tensor<string, []>("cache_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_1_squeeze_mask_0 = const()[name = tensor<string, []>("cache_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<string, []> cache_last_channel_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_last_channel_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [17, 1, 70, 512]> cache_last_channel_to_fp16 = cast(dtype = cache_last_channel_to_fp16_dtype_0, x = cache_last_channel)[name = tensor<string, []>("cast_188")];
+            tensor<fp16, [1, 70, 512]> cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_1_cast_fp16")];
+            tensor<int32, [4]> cache_3_begin_0 = const()[name = tensor<string, []>("cache_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_3_end_0 = const()[name = tensor<string, []>("cache_3_end_0"), val = tensor<int32, [4]>([1, 1, 512, 8])];
+            tensor<bool, [4]> cache_3_end_mask_0 = const()[name = tensor<string, []>("cache_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_3_squeeze_mask_0 = const()[name = tensor<string, []>("cache_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<string, []> cache_last_time_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_last_time_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [17, 1, 512, 8]> cache_last_time_to_fp16 = cast(dtype = cache_last_time_to_fp16_dtype_0, x = cache_last_time)[name = tensor<string, []>("cast_187")];
+            tensor<fp16, [1, 512, 8]> cache_3_cast_fp16 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_3_cast_fp16")];
+            tensor<int32, [1]> input_27_axes_0 = const()[name = tensor<string, []>("input_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4737088)))];
+            tensor<fp16, [512]> encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4738176)))];
+            tensor<fp16, []> var_38_to_fp16 = const()[name = tensor<string, []>("op_38_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 8, 512]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_325_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_0_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4739264)))];
+            tensor<fp16, [2048]> linear_1_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_1_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6836480)))];
+            tensor<fp16, [1, 8, 2048]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_0_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6840640)))];
+            tensor<fp16, [512]> linear_2_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_2_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8937856)))];
+            tensor<fp16, [1, 8, 512]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<fp16, []> var_436_to_fp16 = const()[name = tensor<string, []>("op_436_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_437_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_436_to_fp16)[name = tensor<string, []>("op_437_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_37_cast_fp16 = add(x = var_325_cast_fp16, y = var_437_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
+            tensor<int32, [1]> key_1_axes_0 = const()[name = tensor<string, []>("key_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8938944)))];
+            tensor<fp16, [512]> encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8940032)))];
+            tensor<fp16, [1, 8, 512]> key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")];
+            tensor<bool, []> input_39_interleave_0 = const()[name = tensor<string, []>("input_39_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 78, 512]> input_39_cast_fp16 = concat(axis = var_65, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = tensor<string, []>("input_39_cast_fp16")];
+            tensor<int32, [3]> var_459_begin_0 = const()[name = tensor<string, []>("op_459_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
+            tensor<int32, [3]> var_459_end_0 = const()[name = tensor<string, []>("op_459_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_459_end_mask_0 = const()[name = tensor<string, []>("op_459_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 66, 512]> var_459_cast_fp16 = slice_by_index(begin = var_459_begin_0, end = var_459_end_0, end_mask = var_459_end_mask_0, x = cache_1_cast_fp16)[name = tensor<string, []>("op_459_cast_fp16")];
+            tensor<int32, [3]> var_462_begin_0 = const()[name = tensor<string, []>("op_462_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_462_end_0 = const()[name = tensor<string, []>("op_462_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
+            tensor<bool, [3]> var_462_end_mask_0 = const()[name = tensor<string, []>("op_462_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 4, 512]> var_462_cast_fp16 = slice_by_index(begin = var_462_begin_0, end = var_462_end_0, end_mask = var_462_end_mask_0, x = key_1_cast_fp16)[name = tensor<string, []>("op_462_cast_fp16")];
+            tensor<bool, []> var_465_interleave_0 = const()[name = tensor<string, []>("op_465_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_465_cast_fp16 = concat(axis = var_65, interleave = var_465_interleave_0, values = (var_459_cast_fp16, var_462_cast_fp16))[name = tensor<string, []>("op_465_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_0_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8941120)))];
+            tensor<fp16, [1, 8, 512]> linear_3_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16, x = key_1_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
+            tensor<int32, [4]> var_469 = const()[name = tensor<string, []>("op_469"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 8, 8, 64]> q_1_cast_fp16 = reshape(shape = var_469, x = linear_3_cast_fp16)[name = tensor<string, []>("q_1_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_0_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9465472)))];
+            tensor<fp16, [1, 78, 512]> linear_4_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
+            tensor<int32, [4]> var_473 = const()[name = tensor<string, []>("op_473"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> k_1_cast_fp16 = reshape(shape = var_473, x = linear_4_cast_fp16)[name = tensor<string, []>("k_1_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_0_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9989824)))];
+            tensor<fp16, [1, 78, 512]> linear_5_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
+            tensor<int32, [4]> var_477 = const()[name = tensor<string, []>("op_477"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> v_1_cast_fp16 = reshape(shape = var_477, x = linear_5_cast_fp16)[name = tensor<string, []>("v_1_cast_fp16")];
+            tensor<int32, [4]> value_3_perm_0 = const()[name = tensor<string, []>("value_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10514176)))];
+            tensor<fp16, [1, 8, 8, 64]> var_489_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_489_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10515264)))];
+            tensor<fp16, [1, 8, 8, 64]> var_491_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_491_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_1_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_7_transpose_x_0 = const()[name = tensor<string, []>("x_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_7_transpose_y_0 = const()[name = tensor<string, []>("x_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 155]> var_493_to_fp16 = const()[name = tensor<string, []>("op_493_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10516352)))];
+            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_491_cast_fp16)[name = tensor<string, []>("transpose_237")];
+            tensor<fp16, [1, 8, 8, 155]> x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = var_493_to_fp16)[name = tensor<string, []>("x_7_cast_fp16")];
+            tensor<int32, [8]> x_9_pad_0 = const()[name = tensor<string, []>("x_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_9_mode_0 = const()[name = tensor<string, []>("x_9_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_79_to_fp16 = const()[name = tensor<string, []>("const_79_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 8, 156]> x_9_cast_fp16 = pad(constant_val = const_79_to_fp16, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7_cast_fp16)[name = tensor<string, []>("x_9_cast_fp16")];
+            tensor<int32, [4]> var_501 = const()[name = tensor<string, []>("op_501"), val = tensor<int32, [4]>([1, 8, -1, 8])];
+            tensor<fp16, [1, 8, 156, 8]> x_11_cast_fp16 = reshape(shape = var_501, x = x_9_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
+            tensor<int32, [4]> var_505_begin_0 = const()[name = tensor<string, []>("op_505_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_505_end_0 = const()[name = tensor<string, []>("op_505_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
+            tensor<bool, [4]> var_505_end_mask_0 = const()[name = tensor<string, []>("op_505_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 155, 8]> var_505_cast_fp16 = slice_by_index(begin = var_505_begin_0, end = var_505_end_0, end_mask = var_505_end_mask_0, x = x_11_cast_fp16)[name = tensor<string, []>("op_505_cast_fp16")];
+            tensor<int32, [4]> var_506 = const()[name = tensor<string, []>("op_506"), val = tensor<int32, [4]>([1, 8, 8, 155])];
+            tensor<fp16, [1, 8, 8, 155]> matrix_bd_1_cast_fp16 = reshape(shape = var_506, x = var_505_cast_fp16)[name = tensor<string, []>("matrix_bd_1_cast_fp16")];
+            tensor<bool, []> matrix_ac_1_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_1_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_51_perm_0 = const()[name = tensor<string, []>("transpose_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_52_perm_0 = const()[name = tensor<string, []>("transpose_52_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 78]> transpose_52 = transpose(perm = transpose_52_perm_0, x = k_1_cast_fp16)[name = tensor<string, []>("transpose_235")];
+            tensor<fp16, [1, 8, 8, 64]> transpose_51 = transpose(perm = transpose_51_perm_0, x = var_489_cast_fp16)[name = tensor<string, []>("transpose_236")];
+            tensor<fp16, [1, 8, 8, 78]> matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_51, y = transpose_52)[name = tensor<string, []>("matrix_ac_1_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_3_begin_0 = const()[name = tensor<string, []>("matrix_bd_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_3_end_0 = const()[name = tensor<string, []>("matrix_bd_3_end_0"), val = tensor<int32, [4]>([1, 8, 8, 78])];
+            tensor<bool, [4]> matrix_bd_3_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 8, 78]> matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1_cast_fp16)[name = tensor<string, []>("matrix_bd_3_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_515_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = tensor<string, []>("op_515_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_1_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_1_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 8, 78]> _inversed_scores_1_cast_fp16 = mul(x = var_515_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_1_cast_fp16")];
+            tensor<int32, [1]> mask_11_axes_0 = const()[name = tensor<string, []>("mask_11_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 8, 78]> mask_11 = expand_dims(axes = mask_11_axes_0, x = mask_9)[name = tensor<string, []>("mask_11")];
+            tensor<fp16, []> var_41_to_fp16 = const()[name = tensor<string, []>("op_41_to_fp16"), val = tensor<fp16, []>(-0x1.388p+13)];
+            tensor<fp16, [1, 8, 8, 78]> scores_3_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_3_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_521_cast_fp16 = softmax(axis = var_56, x = scores_3_cast_fp16)[name = tensor<string, []>("op_521_cast_fp16")];
+            tensor<fp16, []> var_40_to_fp16 = const()[name = tensor<string, []>("op_40_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 8, 78]> input_41_cast_fp16 = select(a = var_40_to_fp16, b = var_521_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_41_cast_fp16")];
+            tensor<bool, []> x_13_transpose_x_0 = const()[name = tensor<string, []>("x_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_13_transpose_y_0 = const()[name = tensor<string, []>("x_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 78, 64]> value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = v_1_cast_fp16)[name = tensor<string, []>("transpose_238")];
+            tensor<fp16, [1, 8, 8, 64]> x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41_cast_fp16, y = value_3_cast_fp16)[name = tensor<string, []>("x_13_cast_fp16")];
+            tensor<int32, [4]> var_525_perm_0 = const()[name = tensor<string, []>("op_525_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_526 = const()[name = tensor<string, []>("op_526"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 8, 8, 64]> var_525_cast_fp16 = transpose(perm = var_525_perm_0, x = x_13_cast_fp16)[name = tensor<string, []>("transpose_234")];
+            tensor<fp16, [1, 8, 512]> input_43_cast_fp16 = reshape(shape = var_526, x = var_525_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_0_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10675136)))];
+            tensor<fp16, [1, 8, 512]> linear_7_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
+            tensor<int32, [1]> x_17_axes_0 = const()[name = tensor<string, []>("x_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11199488)))];
+            tensor<fp16, [512]> encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11200576)))];
+            tensor<fp16, [1, 8, 512]> x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("x_17_cast_fp16")];
+            tensor<int32, [3]> input_49_perm_0 = const()[name = tensor<string, []>("input_49_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_51_pad_type_0 = const()[name = tensor<string, []>("input_51_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_51_strides_0 = const()[name = tensor<string, []>("input_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_51_pad_0 = const()[name = tensor<string, []>("input_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_51_dilations_0 = const()[name = tensor<string, []>("input_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_51_groups_0 = const()[name = tensor<string, []>("input_51_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_0_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11201664)))];
+            tensor<fp16, [1, 512, 8]> input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = tensor<string, []>("transpose_233")];
+            tensor<fp16, [1, 1024, 8]> input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
+            tensor<int32, []> x_19_split_num_splits_0 = const()[name = tensor<string, []>("x_19_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_19_split_axis_0 = const()[name = tensor<string, []>("x_19_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 8]> x_19_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51_cast_fp16)[name = tensor<string, []>("x_19_split_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = tensor<string, []>("x_19_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")];
+            tensor<int32, [1]> var_551_axes_0 = const()[name = tensor<string, []>("op_551_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 8]> var_551 = expand_dims(axes = var_551_axes_0, x = pad_mask)[name = tensor<string, []>("op_551")];
+            tensor<fp16, [1, 512, 8]> input_53_cast_fp16 = select(a = var_40_to_fp16, b = x_19_cast_fp16, cond = var_551)[name = tensor<string, []>("input_53_cast_fp16")];
+            tensor<bool, []> new_x_3_interleave_0 = const()[name = tensor<string, []>("new_x_3_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 16]> new_x_3_cast_fp16 = concat(axis = var_56, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = tensor<string, []>("new_x_3_cast_fp16")];
+            tensor<int32, [3]> next_cache_1_begin_0 = const()[name = tensor<string, []>("next_cache_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_1_end_0 = const()[name = tensor<string, []>("next_cache_1_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> next_cache_1_end_mask_0 = const()[name = tensor<string, []>("next_cache_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 12]> next_cache_1_cast_fp16 = slice_by_index(begin = next_cache_1_begin_0, end = next_cache_1_end_0, end_mask = next_cache_1_end_mask_0, x = new_x_3_cast_fp16)[name = tensor<string, []>("next_cache_1_cast_fp16")];
+            tensor<int32, [3]> var_567_begin_0 = const()[name = tensor<string, []>("op_567_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
+            tensor<int32, [3]> var_567_end_0 = const()[name = tensor<string, []>("op_567_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> var_567_end_mask_0 = const()[name = tensor<string, []>("op_567_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_567_cast_fp16 = slice_by_index(begin = var_567_begin_0, end = var_567_end_0, end_mask = var_567_end_mask_0, x = next_cache_1_cast_fp16)[name = tensor<string, []>("op_567_cast_fp16")];
+            tensor<string, []> x_21_pad_type_0 = const()[name = tensor<string, []>("x_21_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_21_groups_0 = const()[name = tensor<string, []>("x_21_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_21_strides_0 = const()[name = tensor<string, []>("x_21_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_21_pad_0 = const()[name = tensor<string, []>("x_21_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_21_dilations_0 = const()[name = tensor<string, []>("x_21_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_0_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12250304)))];
+            tensor<fp16, [1, 512, 8]> x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16, x = new_x_3_cast_fp16)[name = tensor<string, []>("x_21_cast_fp16")];
+            tensor<int32, [3]> input_55_perm_0 = const()[name = tensor<string, []>("input_55_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_23_axes_0 = const()[name = tensor<string, []>("x_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12259584)))];
+            tensor<fp16, [512]> encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12260672)))];
+            tensor<fp16, [1, 8, 512]> input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = tensor<string, []>("transpose_232")];
+            tensor<fp16, [1, 8, 512]> x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")];
+            tensor<int32, [3]> input_57_perm_0 = const()[name = tensor<string, []>("input_57_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 8]> input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = tensor<string, []>("transpose_231")];
+            tensor<fp16, [1, 512, 8]> input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
+            tensor<string, []> x_25_pad_type_0 = const()[name = tensor<string, []>("x_25_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_25_strides_0 = const()[name = tensor<string, []>("x_25_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_25_pad_0 = const()[name = tensor<string, []>("x_25_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_25_dilations_0 = const()[name = tensor<string, []>("x_25_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_25_groups_0 = const()[name = tensor<string, []>("x_25_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_0_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12261760)))];
+            tensor<fp16, [1, 512, 8]> x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("x_25_cast_fp16")];
+            tensor<int32, [3]> input_61_perm_0 = const()[name = tensor<string, []>("input_61_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 8, 512]> input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = tensor<string, []>("transpose_230")];
+            tensor<fp16, [1, 8, 512]> input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
+            tensor<int32, [1]> input_65_axes_0 = const()[name = tensor<string, []>("input_65_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12786112)))];
+            tensor<fp16, [512]> encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12787200)))];
+            tensor<fp16, [1, 8, 512]> input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_0_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12788288)))];
+            tensor<fp16, [1, 8, 2048]> linear_8_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_0_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14885504)))];
+            tensor<fp16, [1, 8, 512]> linear_9_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
+            tensor<fp16, []> var_608_to_fp16 = const()[name = tensor<string, []>("op_608_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_609_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_608_to_fp16)[name = tensor<string, []>("op_609_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_609_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
+            tensor<int32, [1]> input_77_axes_0 = const()[name = tensor<string, []>("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_0_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16982720)))];
+            tensor<fp16, [512]> encoder_layers_0_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16983808)))];
+            tensor<fp16, [1, 8, 512]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
+            tensor<int32, [4]> cache_5_begin_0 = const()[name = tensor<string, []>("cache_5_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_5_end_0 = const()[name = tensor<string, []>("cache_5_end_0"), val = tensor<int32, [4]>([2, 1, 70, 512])];
+            tensor<bool, [4]> cache_5_end_mask_0 = const()[name = tensor<string, []>("cache_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_5_squeeze_mask_0 = const()[name = tensor<string, []>("cache_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_5_cast_fp16 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_5_cast_fp16")];
+            tensor<int32, [4]> cache_7_begin_0 = const()[name = tensor<string, []>("cache_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_7_end_0 = const()[name = tensor<string, []>("cache_7_end_0"), val = tensor<int32, [4]>([2, 1, 512, 8])];
+            tensor<bool, [4]> cache_7_end_mask_0 = const()[name = tensor<string, []>("cache_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_7_squeeze_mask_0 = const()[name = tensor<string, []>("cache_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_7_cast_fp16 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_7_cast_fp16")];
+            tensor<int32, [1]> input_79_axes_0 = const()[name = tensor<string, []>("input_79_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16984896)))];
+            tensor<fp16, [512]> encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16985984)))];
+            tensor<fp16, [1, 8, 512]> input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_1_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16987072)))];
+            tensor<fp16, [1, 8, 2048]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_1_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19084288)))];
+            tensor<fp16, [1, 8, 512]> linear_11_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
+            tensor<fp16, []> var_643_to_fp16 = const()[name = tensor<string, []>("op_643_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_644_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_643_to_fp16)[name = tensor<string, []>("op_644_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_644_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
+            tensor<int32, [1]> key_3_axes_0 = const()[name = tensor<string, []>("key_3_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21181504)))];
+            tensor<fp16, [512]> encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21182592)))];
+            tensor<fp16, [1, 8, 512]> key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("key_3_cast_fp16")];
+            tensor<bool, []> input_91_interleave_0 = const()[name = tensor<string, []>("input_91_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 78, 512]> input_91_cast_fp16 = concat(axis = var_65, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = tensor<string, []>("input_91_cast_fp16")];
+            tensor<int32, [3]> var_666_begin_0 = const()[name = tensor<string, []>("op_666_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
+            tensor<int32, [3]> var_666_end_0 = const()[name = tensor<string, []>("op_666_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_666_end_mask_0 = const()[name = tensor<string, []>("op_666_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 66, 512]> var_666_cast_fp16 = slice_by_index(begin = var_666_begin_0, end = var_666_end_0, end_mask = var_666_end_mask_0, x = cache_5_cast_fp16)[name = tensor<string, []>("op_666_cast_fp16")];
+            tensor<int32, [3]> var_669_begin_0 = const()[name = tensor<string, []>("op_669_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_669_end_0 = const()[name = tensor<string, []>("op_669_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
+            tensor<bool, [3]> var_669_end_mask_0 = const()[name = tensor<string, []>("op_669_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 4, 512]> var_669_cast_fp16 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = key_3_cast_fp16)[name = tensor<string, []>("op_669_cast_fp16")];
+            tensor<bool, []> var_672_interleave_0 = const()[name = tensor<string, []>("op_672_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_672_cast_fp16 = concat(axis = var_65, interleave = var_672_interleave_0, values = (var_666_cast_fp16, var_669_cast_fp16))[name = tensor<string, []>("op_672_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_1_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21183680)))];
+            tensor<fp16, [1, 8, 512]> linear_12_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16, x = key_3_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
+            tensor<int32, [4]> var_676 = const()[name = tensor<string, []>("op_676"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 8, 8, 64]> q_7_cast_fp16 = reshape(shape = var_676, x = linear_12_cast_fp16)[name = tensor<string, []>("q_7_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_1_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21708032)))];
+            tensor<fp16, [1, 78, 512]> linear_13_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
+            tensor<int32, [4]> var_680 = const()[name = tensor<string, []>("op_680"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> k_5_cast_fp16 = reshape(shape = var_680, x = linear_13_cast_fp16)[name = tensor<string, []>("k_5_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_1_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22232384)))];
+            tensor<fp16, [1, 78, 512]> linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
+            tensor<int32, [4]> var_684 = const()[name = tensor<string, []>("op_684"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> v_3_cast_fp16 = reshape(shape = var_684, x = linear_14_cast_fp16)[name = tensor<string, []>("v_3_cast_fp16")];
+            tensor<int32, [4]> value_5_perm_0 = const()[name = tensor<string, []>("value_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22756736)))];
+            tensor<fp16, [1, 8, 8, 64]> var_696_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_696_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22757824)))];
+            tensor<fp16, [1, 8, 8, 64]> var_698_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_698_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_3_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_33_transpose_x_0 = const()[name = tensor<string, []>("x_33_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_33_transpose_y_0 = const()[name = tensor<string, []>("x_33_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 155]> var_700_to_fp16 = const()[name = tensor<string, []>("op_700_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22758912)))];
+            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_698_cast_fp16)[name = tensor<string, []>("transpose_228")];
+            tensor<fp16, [1, 8, 8, 155]> x_33_cast_fp16 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = var_700_to_fp16)[name = tensor<string, []>("x_33_cast_fp16")];
+            tensor<int32, [8]> x_35_pad_0 = const()[name = tensor<string, []>("x_35_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_35_mode_0 = const()[name = tensor<string, []>("x_35_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_92_to_fp16 = const()[name = tensor<string, []>("const_92_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 8, 156]> x_35_cast_fp16 = pad(constant_val = const_92_to_fp16, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")];
+            tensor<int32, [4]> var_708 = const()[name = tensor<string, []>("op_708"), val = tensor<int32, [4]>([1, 8, -1, 8])];
+            tensor<fp16, [1, 8, 156, 8]> x_37_cast_fp16 = reshape(shape = var_708, x = x_35_cast_fp16)[name = tensor<string, []>("x_37_cast_fp16")];
+            tensor<int32, [4]> var_712_begin_0 = const()[name = tensor<string, []>("op_712_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_712_end_0 = const()[name = tensor<string, []>("op_712_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
+            tensor<bool, [4]> var_712_end_mask_0 = const()[name = tensor<string, []>("op_712_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 155, 8]> var_712_cast_fp16 = slice_by_index(begin = var_712_begin_0, end = var_712_end_0, end_mask = var_712_end_mask_0, x = x_37_cast_fp16)[name = tensor<string, []>("op_712_cast_fp16")];
+            tensor<int32, [4]> var_713 = const()[name = tensor<string, []>("op_713"), val = tensor<int32, [4]>([1, 8, 8, 155])];
+            tensor<fp16, [1, 8, 8, 155]> matrix_bd_5_cast_fp16 = reshape(shape = var_713, x = var_712_cast_fp16)[name = tensor<string, []>("matrix_bd_5_cast_fp16")];
+            tensor<bool, []> matrix_ac_3_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_3_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_3_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_3_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_53_perm_0 = const()[name = tensor<string, []>("transpose_53_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_54_perm_0 = const()[name = tensor<string, []>("transpose_54_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 78]> transpose_54 = transpose(perm = transpose_54_perm_0, x = k_5_cast_fp16)[name = tensor<string, []>("transpose_226")];
+            tensor<fp16, [1, 8, 8, 64]> transpose_53 = transpose(perm = transpose_53_perm_0, x = var_696_cast_fp16)[name = tensor<string, []>("transpose_227")];
+            tensor<fp16, [1, 8, 8, 78]> matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_53, y = transpose_54)[name = tensor<string, []>("matrix_ac_3_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_7_begin_0 = const()[name = tensor<string, []>("matrix_bd_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_7_end_0 = const()[name = tensor<string, []>("matrix_bd_7_end_0"), val = tensor<int32, [4]>([1, 8, 8, 78])];
+            tensor<bool, [4]> matrix_bd_7_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 8, 78]> matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5_cast_fp16)[name = tensor<string, []>("matrix_bd_7_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_722_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = tensor<string, []>("op_722_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_5_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_5_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 8, 78]> _inversed_scores_5_cast_fp16 = mul(x = var_722_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_5_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> scores_7_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_7_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_728_cast_fp16 = softmax(axis = var_56, x = scores_7_cast_fp16)[name = tensor<string, []>("op_728_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> input_93_cast_fp16 = select(a = var_40_to_fp16, b = var_728_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_93_cast_fp16")];
+            tensor<bool, []> x_39_transpose_x_0 = const()[name = tensor<string, []>("x_39_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_39_transpose_y_0 = const()[name = tensor<string, []>("x_39_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 78, 64]> value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = v_3_cast_fp16)[name = tensor<string, []>("transpose_229")];
+            tensor<fp16, [1, 8, 8, 64]> x_39_cast_fp16 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93_cast_fp16, y = value_5_cast_fp16)[name = tensor<string, []>("x_39_cast_fp16")];
+            tensor<int32, [4]> var_732_perm_0 = const()[name = tensor<string, []>("op_732_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_733 = const()[name = tensor<string, []>("op_733"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 8, 8, 64]> var_732_cast_fp16 = transpose(perm = var_732_perm_0, x = x_39_cast_fp16)[name = tensor<string, []>("transpose_225")];
+            tensor<fp16, [1, 8, 512]> input_95_cast_fp16 = reshape(shape = var_733, x = var_732_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_1_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22917696)))];
+            tensor<fp16, [1, 8, 512]> linear_16_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
+            tensor<int32, [1]> x_43_axes_0 = const()[name = tensor<string, []>("x_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23442048)))];
+            tensor<fp16, [512]> encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23443136)))];
+            tensor<fp16, [1, 8, 512]> x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")];
+            tensor<int32, [3]> input_101_perm_0 = const()[name = tensor<string, []>("input_101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_103_pad_type_0 = const()[name = tensor<string, []>("input_103_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_103_strides_0 = const()[name = tensor<string, []>("input_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_103_pad_0 = const()[name = tensor<string, []>("input_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_103_dilations_0 = const()[name = tensor<string, []>("input_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_103_groups_0 = const()[name = tensor<string, []>("input_103_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_1_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23444224)))];
+            tensor<fp16, [1, 512, 8]> input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = tensor<string, []>("transpose_224")];
+            tensor<fp16, [1, 1024, 8]> input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
+            tensor<int32, []> x_45_split_num_splits_0 = const()[name = tensor<string, []>("x_45_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_45_split_axis_0 = const()[name = tensor<string, []>("x_45_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 8]> x_45_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> x_45_split_cast_fp16_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103_cast_fp16)[name = tensor<string, []>("x_45_split_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = tensor<string, []>("x_45_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_45_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> input_105_cast_fp16 = select(a = var_40_to_fp16, b = x_45_cast_fp16, cond = var_551)[name = tensor<string, []>("input_105_cast_fp16")];
+            tensor<bool, []> new_x_7_interleave_0 = const()[name = tensor<string, []>("new_x_7_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 16]> new_x_7_cast_fp16 = concat(axis = var_56, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = tensor<string, []>("new_x_7_cast_fp16")];
+            tensor<int32, [3]> next_cache_3_begin_0 = const()[name = tensor<string, []>("next_cache_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_3_end_0 = const()[name = tensor<string, []>("next_cache_3_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> next_cache_3_end_mask_0 = const()[name = tensor<string, []>("next_cache_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 12]> next_cache_3_cast_fp16 = slice_by_index(begin = next_cache_3_begin_0, end = next_cache_3_end_0, end_mask = next_cache_3_end_mask_0, x = new_x_7_cast_fp16)[name = tensor<string, []>("next_cache_3_cast_fp16")];
+            tensor<int32, [3]> var_774_begin_0 = const()[name = tensor<string, []>("op_774_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
+            tensor<int32, [3]> var_774_end_0 = const()[name = tensor<string, []>("op_774_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> var_774_end_mask_0 = const()[name = tensor<string, []>("op_774_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_774_cast_fp16 = slice_by_index(begin = var_774_begin_0, end = var_774_end_0, end_mask = var_774_end_mask_0, x = next_cache_3_cast_fp16)[name = tensor<string, []>("op_774_cast_fp16")];
+            tensor<string, []> x_47_pad_type_0 = const()[name = tensor<string, []>("x_47_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_47_groups_0 = const()[name = tensor<string, []>("x_47_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_47_strides_0 = const()[name = tensor<string, []>("x_47_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_47_pad_0 = const()[name = tensor<string, []>("x_47_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_47_dilations_0 = const()[name = tensor<string, []>("x_47_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_1_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24492864)))];
+            tensor<fp16, [1, 512, 8]> x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16, x = new_x_7_cast_fp16)[name = tensor<string, []>("x_47_cast_fp16")];
+            tensor<int32, [3]> input_107_perm_0 = const()[name = tensor<string, []>("input_107_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_49_axes_0 = const()[name = tensor<string, []>("x_49_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24502144)))];
+            tensor<fp16, [512]> encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24503232)))];
+            tensor<fp16, [1, 8, 512]> input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = tensor<string, []>("transpose_223")];
+            tensor<fp16, [1, 8, 512]> x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = tensor<string, []>("x_49_cast_fp16")];
+            tensor<int32, [3]> input_109_perm_0 = const()[name = tensor<string, []>("input_109_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 8]> input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = tensor<string, []>("transpose_222")];
+            tensor<fp16, [1, 512, 8]> input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
+            tensor<string, []> x_51_pad_type_0 = const()[name = tensor<string, []>("x_51_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_51_strides_0 = const()[name = tensor<string, []>("x_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_51_pad_0 = const()[name = tensor<string, []>("x_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_51_dilations_0 = const()[name = tensor<string, []>("x_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_51_groups_0 = const()[name = tensor<string, []>("x_51_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_1_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24504320)))];
+            tensor<fp16, [1, 512, 8]> x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("x_51_cast_fp16")];
+            tensor<int32, [3]> input_113_perm_0 = const()[name = tensor<string, []>("input_113_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 8, 512]> input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = tensor<string, []>("transpose_221")];
+            tensor<fp16, [1, 8, 512]> input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
+            tensor<int32, [1]> input_117_axes_0 = const()[name = tensor<string, []>("input_117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25028672)))];
+            tensor<fp16, [512]> encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25029760)))];
+            tensor<fp16, [1, 8, 512]> input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_1_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25030848)))];
+            tensor<fp16, [1, 8, 2048]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16, x = input_117_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_1_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27128064)))];
+            tensor<fp16, [1, 8, 512]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
+            tensor<fp16, []> var_815_to_fp16 = const()[name = tensor<string, []>("op_815_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_816_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_815_to_fp16)[name = tensor<string, []>("op_816_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_816_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
+            tensor<int32, [1]> input_129_axes_0 = const()[name = tensor<string, []>("input_129_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_1_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29225280)))];
+            tensor<fp16, [512]> encoder_layers_1_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29226368)))];
+            tensor<fp16, [1, 8, 512]> input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
+            tensor<int32, [4]> cache_9_begin_0 = const()[name = tensor<string, []>("cache_9_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_9_end_0 = const()[name = tensor<string, []>("cache_9_end_0"), val = tensor<int32, [4]>([3, 1, 70, 512])];
+            tensor<bool, [4]> cache_9_end_mask_0 = const()[name = tensor<string, []>("cache_9_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_9_squeeze_mask_0 = const()[name = tensor<string, []>("cache_9_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_9_cast_fp16 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_9_cast_fp16")];
+            tensor<int32, [4]> cache_11_begin_0 = const()[name = tensor<string, []>("cache_11_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_11_end_0 = const()[name = tensor<string, []>("cache_11_end_0"), val = tensor<int32, [4]>([3, 1, 512, 8])];
+            tensor<bool, [4]> cache_11_end_mask_0 = const()[name = tensor<string, []>("cache_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_11_squeeze_mask_0 = const()[name = tensor<string, []>("cache_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_11_cast_fp16 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_11_cast_fp16")];
+            tensor<int32, [1]> input_131_axes_0 = const()[name = tensor<string, []>("input_131_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29227456)))];
+            tensor<fp16, [512]> encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29228544)))];
+            tensor<fp16, [1, 8, 512]> input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("input_131_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_2_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29229632)))];
+            tensor<fp16, [1, 8, 2048]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16, x = input_131_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_2_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31326848)))];
+            tensor<fp16, [1, 8, 512]> linear_20_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16, x = input_135_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
+            tensor<fp16, []> var_850_to_fp16 = const()[name = tensor<string, []>("op_850_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_851_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_850_to_fp16)[name = tensor<string, []>("op_851_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_851_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
+            tensor<int32, [1]> key_5_axes_0 = const()[name = tensor<string, []>("key_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33424064)))];
+            tensor<fp16, [512]> encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33425152)))];
+            tensor<fp16, [1, 8, 512]> key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")];
+            tensor<bool, []> input_143_interleave_0 = const()[name = tensor<string, []>("input_143_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 78, 512]> input_143_cast_fp16 = concat(axis = var_65, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = tensor<string, []>("input_143_cast_fp16")];
+            tensor<int32, [3]> var_873_begin_0 = const()[name = tensor<string, []>("op_873_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
+            tensor<int32, [3]> var_873_end_0 = const()[name = tensor<string, []>("op_873_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_873_end_mask_0 = const()[name = tensor<string, []>("op_873_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 66, 512]> var_873_cast_fp16 = slice_by_index(begin = var_873_begin_0, end = var_873_end_0, end_mask = var_873_end_mask_0, x = cache_9_cast_fp16)[name = tensor<string, []>("op_873_cast_fp16")];
+            tensor<int32, [3]> var_876_begin_0 = const()[name = tensor<string, []>("op_876_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_876_end_0 = const()[name = tensor<string, []>("op_876_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
+            tensor<bool, [3]> var_876_end_mask_0 = const()[name = tensor<string, []>("op_876_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 4, 512]> var_876_cast_fp16 = slice_by_index(begin = var_876_begin_0, end = var_876_end_0, end_mask = var_876_end_mask_0, x = key_5_cast_fp16)[name = tensor<string, []>("op_876_cast_fp16")];
+            tensor<bool, []> var_879_interleave_0 = const()[name = tensor<string, []>("op_879_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_879_cast_fp16 = concat(axis = var_65, interleave = var_879_interleave_0, values = (var_873_cast_fp16, var_876_cast_fp16))[name = tensor<string, []>("op_879_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_2_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33426240)))];
+            tensor<fp16, [1, 8, 512]> linear_21_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16, x = key_5_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
+            tensor<int32, [4]> var_883 = const()[name = tensor<string, []>("op_883"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 8, 8, 64]> q_13_cast_fp16 = reshape(shape = var_883, x = linear_21_cast_fp16)[name = tensor<string, []>("q_13_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_2_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33950592)))];
+            tensor<fp16, [1, 78, 512]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
+            tensor<int32, [4]> var_887 = const()[name = tensor<string, []>("op_887"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> k_9_cast_fp16 = reshape(shape = var_887, x = linear_22_cast_fp16)[name = tensor<string, []>("k_9_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_2_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34474944)))];
+            tensor<fp16, [1, 78, 512]> linear_23_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
+            tensor<int32, [4]> var_891 = const()[name = tensor<string, []>("op_891"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> v_5_cast_fp16 = reshape(shape = var_891, x = linear_23_cast_fp16)[name = tensor<string, []>("v_5_cast_fp16")];
+            tensor<int32, [4]> value_7_perm_0 = const()[name = tensor<string, []>("value_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34999296)))];
+            tensor<fp16, [1, 8, 8, 64]> var_903_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_903_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35000384)))];
+            tensor<fp16, [1, 8, 8, 64]> var_905_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_905_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_5_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_59_transpose_x_0 = const()[name = tensor<string, []>("x_59_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_59_transpose_y_0 = const()[name = tensor<string, []>("x_59_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 155]> var_907_to_fp16 = const()[name = tensor<string, []>("op_907_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35001472)))];
+            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_905_cast_fp16)[name = tensor<string, []>("transpose_219")];
+            tensor<fp16, [1, 8, 8, 155]> x_59_cast_fp16 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = var_907_to_fp16)[name = tensor<string, []>("x_59_cast_fp16")];
+            tensor<int32, [8]> x_61_pad_0 = const()[name = tensor<string, []>("x_61_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_61_mode_0 = const()[name = tensor<string, []>("x_61_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_105_to_fp16 = const()[name = tensor<string, []>("const_105_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 8, 156]> x_61_cast_fp16 = pad(constant_val = const_105_to_fp16, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59_cast_fp16)[name = tensor<string, []>("x_61_cast_fp16")];
+            tensor<int32, [4]> var_915 = const()[name = tensor<string, []>("op_915"), val = tensor<int32, [4]>([1, 8, -1, 8])];
+            tensor<fp16, [1, 8, 156, 8]> x_63_cast_fp16 = reshape(shape = var_915, x = x_61_cast_fp16)[name = tensor<string, []>("x_63_cast_fp16")];
+            tensor<int32, [4]> var_919_begin_0 = const()[name = tensor<string, []>("op_919_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_919_end_0 = const()[name = tensor<string, []>("op_919_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
+            tensor<bool, [4]> var_919_end_mask_0 = const()[name = tensor<string, []>("op_919_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 155, 8]> var_919_cast_fp16 = slice_by_index(begin = var_919_begin_0, end = var_919_end_0, end_mask = var_919_end_mask_0, x = x_63_cast_fp16)[name = tensor<string, []>("op_919_cast_fp16")];
+            tensor<int32, [4]> var_920 = const()[name = tensor<string, []>("op_920"), val = tensor<int32, [4]>([1, 8, 8, 155])];
+            tensor<fp16, [1, 8, 8, 155]> matrix_bd_9_cast_fp16 = reshape(shape = var_920, x = var_919_cast_fp16)[name = tensor<string, []>("matrix_bd_9_cast_fp16")];
+            tensor<bool, []> matrix_ac_5_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_5_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_5_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_5_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_55_perm_0 = const()[name = tensor<string, []>("transpose_55_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_56_perm_0 = const()[name = tensor<string, []>("transpose_56_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 78]> transpose_56 = transpose(perm = transpose_56_perm_0, x = k_9_cast_fp16)[name = tensor<string, []>("transpose_217")];
+            tensor<fp16, [1, 8, 8, 64]> transpose_55 = transpose(perm = transpose_55_perm_0, x = var_903_cast_fp16)[name = tensor<string, []>("transpose_218")];
+            tensor<fp16, [1, 8, 8, 78]> matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_55, y = transpose_56)[name = tensor<string, []>("matrix_ac_5_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_11_begin_0 = const()[name = tensor<string, []>("matrix_bd_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_11_end_0 = const()[name = tensor<string, []>("matrix_bd_11_end_0"), val = tensor<int32, [4]>([1, 8, 8, 78])];
+            tensor<bool, [4]> matrix_bd_11_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 8, 78]> matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9_cast_fp16)[name = tensor<string, []>("matrix_bd_11_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_929_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = tensor<string, []>("op_929_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_9_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_9_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 8, 78]> _inversed_scores_9_cast_fp16 = mul(x = var_929_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_9_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> scores_11_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_11_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_935_cast_fp16 = softmax(axis = var_56, x = scores_11_cast_fp16)[name = tensor<string, []>("op_935_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> input_145_cast_fp16 = select(a = var_40_to_fp16, b = var_935_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_145_cast_fp16")];
+            tensor<bool, []> x_65_transpose_x_0 = const()[name = tensor<string, []>("x_65_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_65_transpose_y_0 = const()[name = tensor<string, []>("x_65_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 78, 64]> value_7_cast_fp16 = transpose(perm = value_7_perm_0, x = v_5_cast_fp16)[name = tensor<string, []>("transpose_220")];
+            tensor<fp16, [1, 8, 8, 64]> x_65_cast_fp16 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145_cast_fp16, y = value_7_cast_fp16)[name = tensor<string, []>("x_65_cast_fp16")];
+            tensor<int32, [4]> var_939_perm_0 = const()[name = tensor<string, []>("op_939_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_940 = const()[name = tensor<string, []>("op_940"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 8, 8, 64]> var_939_cast_fp16 = transpose(perm = var_939_perm_0, x = x_65_cast_fp16)[name = tensor<string, []>("transpose_216")];
+            tensor<fp16, [1, 8, 512]> input_147_cast_fp16 = reshape(shape = var_940, x = var_939_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_2_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35160256)))];
+            tensor<fp16, [1, 8, 512]> linear_25_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")];
+            tensor<int32, [1]> x_69_axes_0 = const()[name = tensor<string, []>("x_69_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35684608)))];
+            tensor<fp16, [512]> encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35685696)))];
+            tensor<fp16, [1, 8, 512]> x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = tensor<string, []>("x_69_cast_fp16")];
+            tensor<int32, [3]> input_153_perm_0 = const()[name = tensor<string, []>("input_153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_155_pad_type_0 = const()[name = tensor<string, []>("input_155_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_155_strides_0 = const()[name = tensor<string, []>("input_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_155_pad_0 = const()[name = tensor<string, []>("input_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_155_dilations_0 = const()[name = tensor<string, []>("input_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_155_groups_0 = const()[name = tensor<string, []>("input_155_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_2_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35686784)))];
+            tensor<fp16, [1, 512, 8]> input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = tensor<string, []>("transpose_215")];
+            tensor<fp16, [1, 1024, 8]> input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor<string, []>("input_155_cast_fp16")];
+            tensor<int32, []> x_71_split_num_splits_0 = const()[name = tensor<string, []>("x_71_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_71_split_axis_0 = const()[name = tensor<string, []>("x_71_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 8]> x_71_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> x_71_split_cast_fp16_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155_cast_fp16)[name = tensor<string, []>("x_71_split_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = tensor<string, []>("x_71_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_71_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> input_157_cast_fp16 = select(a = var_40_to_fp16, b = x_71_cast_fp16, cond = var_551)[name = tensor<string, []>("input_157_cast_fp16")];
+            tensor<bool, []> new_x_11_interleave_0 = const()[name = tensor<string, []>("new_x_11_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 16]> new_x_11_cast_fp16 = concat(axis = var_56, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = tensor<string, []>("new_x_11_cast_fp16")];
+            tensor<int32, [3]> next_cache_5_begin_0 = const()[name = tensor<string, []>("next_cache_5_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_5_end_0 = const()[name = tensor<string, []>("next_cache_5_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> next_cache_5_end_mask_0 = const()[name = tensor<string, []>("next_cache_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 12]> next_cache_5_cast_fp16 = slice_by_index(begin = next_cache_5_begin_0, end = next_cache_5_end_0, end_mask = next_cache_5_end_mask_0, x = new_x_11_cast_fp16)[name = tensor<string, []>("next_cache_5_cast_fp16")];
+            tensor<int32, [3]> var_981_begin_0 = const()[name = tensor<string, []>("op_981_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
+            tensor<int32, [3]> var_981_end_0 = const()[name = tensor<string, []>("op_981_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> var_981_end_mask_0 = const()[name = tensor<string, []>("op_981_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_981_cast_fp16 = slice_by_index(begin = var_981_begin_0, end = var_981_end_0, end_mask = var_981_end_mask_0, x = next_cache_5_cast_fp16)[name = tensor<string, []>("op_981_cast_fp16")];
+            tensor<string, []> x_73_pad_type_0 = const()[name = tensor<string, []>("x_73_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_73_groups_0 = const()[name = tensor<string, []>("x_73_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_73_strides_0 = const()[name = tensor<string, []>("x_73_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_73_pad_0 = const()[name = tensor<string, []>("x_73_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_73_dilations_0 = const()[name = tensor<string, []>("x_73_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_2_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36735424)))];
+            tensor<fp16, [1, 512, 8]> x_73_cast_fp16 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16, x = new_x_11_cast_fp16)[name = tensor<string, []>("x_73_cast_fp16")];
+            tensor<int32, [3]> input_159_perm_0 = const()[name = tensor<string, []>("input_159_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_75_axes_0 = const()[name = tensor<string, []>("x_75_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36744704)))];
+            tensor<fp16, [512]> encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36745792)))];
+            tensor<fp16, [1, 8, 512]> input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = tensor<string, []>("transpose_214")];
+            tensor<fp16, [1, 8, 512]> x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = tensor<string, []>("x_75_cast_fp16")];
+            tensor<int32, [3]> input_161_perm_0 = const()[name = tensor<string, []>("input_161_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 8]> input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = tensor<string, []>("transpose_213")];
+            tensor<fp16, [1, 512, 8]> input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
+            tensor<string, []> x_77_pad_type_0 = const()[name = tensor<string, []>("x_77_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_77_strides_0 = const()[name = tensor<string, []>("x_77_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_77_pad_0 = const()[name = tensor<string, []>("x_77_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_77_dilations_0 = const()[name = tensor<string, []>("x_77_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_77_groups_0 = const()[name = tensor<string, []>("x_77_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_2_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36746880)))];
+            tensor<fp16, [1, 512, 8]> x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16, x = input_163_cast_fp16)[name = tensor<string, []>("x_77_cast_fp16")];
+            tensor<int32, [3]> input_165_perm_0 = const()[name = tensor<string, []>("input_165_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 8, 512]> input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = tensor<string, []>("transpose_212")];
+            tensor<fp16, [1, 8, 512]> input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")];
+            tensor<int32, [1]> input_169_axes_0 = const()[name = tensor<string, []>("input_169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37271232)))];
+            tensor<fp16, [512]> encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37272320)))];
+            tensor<fp16, [1, 8, 512]> input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor<string, []>("input_169_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_2_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37273408)))];
+            tensor<fp16, [1, 8, 2048]> linear_26_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16, x = input_169_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_2_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39370624)))];
+            tensor<fp16, [1, 8, 512]> linear_27_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")];
+            tensor<fp16, []> var_1022_to_fp16 = const()[name = tensor<string, []>("op_1022_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_1023_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1022_to_fp16)[name = tensor<string, []>("op_1023_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_1023_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
+            tensor<int32, [1]> input_181_axes_0 = const()[name = tensor<string, []>("input_181_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_2_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41467840)))];
+            tensor<fp16, [512]> encoder_layers_2_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41468928)))];
+            tensor<fp16, [1, 8, 512]> input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")];
+            tensor<int32, [4]> cache_13_begin_0 = const()[name = tensor<string, []>("cache_13_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_13_end_0 = const()[name = tensor<string, []>("cache_13_end_0"), val = tensor<int32, [4]>([4, 1, 70, 512])];
+            tensor<bool, [4]> cache_13_end_mask_0 = const()[name = tensor<string, []>("cache_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_13_squeeze_mask_0 = const()[name = tensor<string, []>("cache_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_13_cast_fp16 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_13_cast_fp16")];
+            tensor<int32, [4]> cache_15_begin_0 = const()[name = tensor<string, []>("cache_15_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_15_end_0 = const()[name = tensor<string, []>("cache_15_end_0"), val = tensor<int32, [4]>([4, 1, 512, 8])];
+            tensor<bool, [4]> cache_15_end_mask_0 = const()[name = tensor<string, []>("cache_15_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_15_squeeze_mask_0 = const()[name = tensor<string, []>("cache_15_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_15_cast_fp16 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_15_cast_fp16")];
+            tensor<int32, [1]> input_183_axes_0 = const()[name = tensor<string, []>("input_183_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41470016)))];
+            tensor<fp16, [512]> encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41471104)))];
+            tensor<fp16, [1, 8, 512]> input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor<string, []>("input_183_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_3_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41472192)))];
+            tensor<fp16, [1, 8, 2048]> linear_28_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = tensor<string, []>("input_187_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_3_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43569408)))];
+            tensor<fp16, [1, 8, 512]> linear_29_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")];
+            tensor<fp16, []> var_1057_to_fp16 = const()[name = tensor<string, []>("op_1057_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_1058_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_1057_to_fp16)[name = tensor<string, []>("op_1058_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_1058_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
+            tensor<int32, [1]> key_7_axes_0 = const()[name = tensor<string, []>("key_7_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45666624)))];
+            tensor<fp16, [512]> encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45667712)))];
+            tensor<fp16, [1, 8, 512]> key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = tensor<string, []>("key_7_cast_fp16")];
+            tensor<bool, []> input_195_interleave_0 = const()[name = tensor<string, []>("input_195_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 78, 512]> input_195_cast_fp16 = concat(axis = var_65, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = tensor<string, []>("input_195_cast_fp16")];
+            tensor<int32, [3]> var_1080_begin_0 = const()[name = tensor<string, []>("op_1080_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
+            tensor<int32, [3]> var_1080_end_0 = const()[name = tensor<string, []>("op_1080_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_1080_end_mask_0 = const()[name = tensor<string, []>("op_1080_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 66, 512]> var_1080_cast_fp16 = slice_by_index(begin = var_1080_begin_0, end = var_1080_end_0, end_mask = var_1080_end_mask_0, x = cache_13_cast_fp16)[name = tensor<string, []>("op_1080_cast_fp16")];
+            tensor<int32, [3]> var_1083_begin_0 = const()[name = tensor<string, []>("op_1083_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_1083_end_0 = const()[name = tensor<string, []>("op_1083_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
+            tensor<bool, [3]> var_1083_end_mask_0 = const()[name = tensor<string, []>("op_1083_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 4, 512]> var_1083_cast_fp16 = slice_by_index(begin = var_1083_begin_0, end = var_1083_end_0, end_mask = var_1083_end_mask_0, x = key_7_cast_fp16)[name = tensor<string, []>("op_1083_cast_fp16")];
+            tensor<bool, []> var_1086_interleave_0 = const()[name = tensor<string, []>("op_1086_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_1086_cast_fp16 = concat(axis = var_65, interleave = var_1086_interleave_0, values = (var_1080_cast_fp16, var_1083_cast_fp16))[name = tensor<string, []>("op_1086_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_3_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45668800)))];
+            tensor<fp16, [1, 8, 512]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16, x = key_7_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")];
+            tensor<int32, [4]> var_1090 = const()[name = tensor<string, []>("op_1090"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 8, 8, 64]> q_19_cast_fp16 = reshape(shape = var_1090, x = linear_30_cast_fp16)[name = tensor<string, []>("q_19_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_3_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46193152)))];
+            tensor<fp16, [1, 78, 512]> linear_31_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")];
+            tensor<int32, [4]> var_1094 = const()[name = tensor<string, []>("op_1094"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> k_13_cast_fp16 = reshape(shape = var_1094, x = linear_31_cast_fp16)[name = tensor<string, []>("k_13_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_3_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46717504)))];
+            tensor<fp16, [1, 78, 512]> linear_32_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")];
+            tensor<int32, [4]> var_1098 = const()[name = tensor<string, []>("op_1098"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> v_7_cast_fp16 = reshape(shape = var_1098, x = linear_32_cast_fp16)[name = tensor<string, []>("v_7_cast_fp16")];
+            tensor<int32, [4]> value_9_perm_0 = const()[name = tensor<string, []>("value_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47241856)))];
+            tensor<fp16, [1, 8, 8, 64]> var_1110_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1110_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47242944)))];
+            tensor<fp16, [1, 8, 8, 64]> var_1112_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1112_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_7_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_85_transpose_x_0 = const()[name = tensor<string, []>("x_85_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_85_transpose_y_0 = const()[name = tensor<string, []>("x_85_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 155]> var_1114_to_fp16 = const()[name = tensor<string, []>("op_1114_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47244032)))];
+            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_1112_cast_fp16)[name = tensor<string, []>("transpose_210")];
+            tensor<fp16, [1, 8, 8, 155]> x_85_cast_fp16 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = var_1114_to_fp16)[name = tensor<string, []>("x_85_cast_fp16")];
+            tensor<int32, [8]> x_87_pad_0 = const()[name = tensor<string, []>("x_87_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_87_mode_0 = const()[name = tensor<string, []>("x_87_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_118_to_fp16 = const()[name = tensor<string, []>("const_118_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 8, 156]> x_87_cast_fp16 = pad(constant_val = const_118_to_fp16, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85_cast_fp16)[name = tensor<string, []>("x_87_cast_fp16")];
+            tensor<int32, [4]> var_1122 = const()[name = tensor<string, []>("op_1122"), val = tensor<int32, [4]>([1, 8, -1, 8])];
+            tensor<fp16, [1, 8, 156, 8]> x_89_cast_fp16 = reshape(shape = var_1122, x = x_87_cast_fp16)[name = tensor<string, []>("x_89_cast_fp16")];
+            tensor<int32, [4]> var_1126_begin_0 = const()[name = tensor<string, []>("op_1126_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1126_end_0 = const()[name = tensor<string, []>("op_1126_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
+            tensor<bool, [4]> var_1126_end_mask_0 = const()[name = tensor<string, []>("op_1126_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 155, 8]> var_1126_cast_fp16 = slice_by_index(begin = var_1126_begin_0, end = var_1126_end_0, end_mask = var_1126_end_mask_0, x = x_89_cast_fp16)[name = tensor<string, []>("op_1126_cast_fp16")];
+            tensor<int32, [4]> var_1127 = const()[name = tensor<string, []>("op_1127"), val = tensor<int32, [4]>([1, 8, 8, 155])];
+            tensor<fp16, [1, 8, 8, 155]> matrix_bd_13_cast_fp16 = reshape(shape = var_1127, x = var_1126_cast_fp16)[name = tensor<string, []>("matrix_bd_13_cast_fp16")];
+            tensor<bool, []> matrix_ac_7_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_7_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_57_perm_0 = const()[name = tensor<string, []>("transpose_57_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_58_perm_0 = const()[name = tensor<string, []>("transpose_58_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 78]> transpose_58 = transpose(perm = transpose_58_perm_0, x = k_13_cast_fp16)[name = tensor<string, []>("transpose_208")];
+            tensor<fp16, [1, 8, 8, 64]> transpose_57 = transpose(perm = transpose_57_perm_0, x = var_1110_cast_fp16)[name = tensor<string, []>("transpose_209")];
+            tensor<fp16, [1, 8, 8, 78]> matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_57, y = transpose_58)[name = tensor<string, []>("matrix_ac_7_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_15_begin_0 = const()[name = tensor<string, []>("matrix_bd_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_15_end_0 = const()[name = tensor<string, []>("matrix_bd_15_end_0"), val = tensor<int32, [4]>([1, 8, 8, 78])];
+            tensor<bool, [4]> matrix_bd_15_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 8, 78]> matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13_cast_fp16)[name = tensor<string, []>("matrix_bd_15_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_1136_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = tensor<string, []>("op_1136_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_13_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_13_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 8, 78]> _inversed_scores_13_cast_fp16 = mul(x = var_1136_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_13_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> scores_15_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_15_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_1142_cast_fp16 = softmax(axis = var_56, x = scores_15_cast_fp16)[name = tensor<string, []>("op_1142_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> input_197_cast_fp16 = select(a = var_40_to_fp16, b = var_1142_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_197_cast_fp16")];
+            tensor<bool, []> x_91_transpose_x_0 = const()[name = tensor<string, []>("x_91_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_91_transpose_y_0 = const()[name = tensor<string, []>("x_91_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 78, 64]> value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_7_cast_fp16)[name = tensor<string, []>("transpose_211")];
+            tensor<fp16, [1, 8, 8, 64]> x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197_cast_fp16, y = value_9_cast_fp16)[name = tensor<string, []>("x_91_cast_fp16")];
+            tensor<int32, [4]> var_1146_perm_0 = const()[name = tensor<string, []>("op_1146_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1147 = const()[name = tensor<string, []>("op_1147"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 8, 8, 64]> var_1146_cast_fp16 = transpose(perm = var_1146_perm_0, x = x_91_cast_fp16)[name = tensor<string, []>("transpose_207")];
+            tensor<fp16, [1, 8, 512]> input_199_cast_fp16 = reshape(shape = var_1147, x = var_1146_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_3_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47402816)))];
+            tensor<fp16, [1, 8, 512]> linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16, x = input_199_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")];
+            tensor<int32, [1]> x_95_axes_0 = const()[name = tensor<string, []>("x_95_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47927168)))];
+            tensor<fp16, [512]> encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47928256)))];
+            tensor<fp16, [1, 8, 512]> x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = tensor<string, []>("x_95_cast_fp16")];
+            tensor<int32, [3]> input_205_perm_0 = const()[name = tensor<string, []>("input_205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_207_pad_type_0 = const()[name = tensor<string, []>("input_207_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_207_strides_0 = const()[name = tensor<string, []>("input_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_207_pad_0 = const()[name = tensor<string, []>("input_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_207_dilations_0 = const()[name = tensor<string, []>("input_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_207_groups_0 = const()[name = tensor<string, []>("input_207_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_3_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47929344)))];
+            tensor<fp16, [1, 512, 8]> input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = tensor<string, []>("transpose_206")];
+            tensor<fp16, [1, 1024, 8]> input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")];
+            tensor<int32, []> x_97_split_num_splits_0 = const()[name = tensor<string, []>("x_97_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_97_split_axis_0 = const()[name = tensor<string, []>("x_97_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 8]> x_97_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207_cast_fp16)[name = tensor<string, []>("x_97_split_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = tensor<string, []>("x_97_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_97_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> input_209_cast_fp16 = select(a = var_40_to_fp16, b = x_97_cast_fp16, cond = var_551)[name = tensor<string, []>("input_209_cast_fp16")];
+            tensor<bool, []> new_x_15_interleave_0 = const()[name = tensor<string, []>("new_x_15_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 16]> new_x_15_cast_fp16 = concat(axis = var_56, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = tensor<string, []>("new_x_15_cast_fp16")];
+            tensor<int32, [3]> next_cache_7_begin_0 = const()[name = tensor<string, []>("next_cache_7_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_7_end_0 = const()[name = tensor<string, []>("next_cache_7_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> next_cache_7_end_mask_0 = const()[name = tensor<string, []>("next_cache_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 12]> next_cache_7_cast_fp16 = slice_by_index(begin = next_cache_7_begin_0, end = next_cache_7_end_0, end_mask = next_cache_7_end_mask_0, x = new_x_15_cast_fp16)[name = tensor<string, []>("next_cache_7_cast_fp16")];
+            tensor<int32, [3]> var_1188_begin_0 = const()[name = tensor<string, []>("op_1188_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
+            tensor<int32, [3]> var_1188_end_0 = const()[name = tensor<string, []>("op_1188_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> var_1188_end_mask_0 = const()[name = tensor<string, []>("op_1188_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_1188_cast_fp16 = slice_by_index(begin = var_1188_begin_0, end = var_1188_end_0, end_mask = var_1188_end_mask_0, x = next_cache_7_cast_fp16)[name = tensor<string, []>("op_1188_cast_fp16")];
+            tensor<string, []> x_99_pad_type_0 = const()[name = tensor<string, []>("x_99_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_99_groups_0 = const()[name = tensor<string, []>("x_99_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_99_strides_0 = const()[name = tensor<string, []>("x_99_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_99_pad_0 = const()[name = tensor<string, []>("x_99_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_99_dilations_0 = const()[name = tensor<string, []>("x_99_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_3_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48977984)))];
+            tensor<fp16, [1, 512, 8]> x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16, x = new_x_15_cast_fp16)[name = tensor<string, []>("x_99_cast_fp16")];
+            tensor<int32, [3]> input_211_perm_0 = const()[name = tensor<string, []>("input_211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_101_axes_0 = const()[name = tensor<string, []>("x_101_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48987264)))];
+            tensor<fp16, [512]> encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48988352)))];
+            tensor<fp16, [1, 8, 512]> input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = tensor<string, []>("transpose_205")];
+            tensor<fp16, [1, 8, 512]> x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = tensor<string, []>("x_101_cast_fp16")];
+            tensor<int32, [3]> input_213_perm_0 = const()[name = tensor<string, []>("input_213_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 8]> input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = tensor<string, []>("transpose_204")];
+            tensor<fp16, [1, 512, 8]> input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = tensor<string, []>("input_215_cast_fp16")];
+            tensor<string, []> x_103_pad_type_0 = const()[name = tensor<string, []>("x_103_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_103_strides_0 = const()[name = tensor<string, []>("x_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_103_pad_0 = const()[name = tensor<string, []>("x_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_103_dilations_0 = const()[name = tensor<string, []>("x_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_103_groups_0 = const()[name = tensor<string, []>("x_103_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_3_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48989440)))];
+            tensor<fp16, [1, 512, 8]> x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor<string, []>("x_103_cast_fp16")];
+            tensor<int32, [3]> input_217_perm_0 = const()[name = tensor<string, []>("input_217_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 8, 512]> input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = tensor<string, []>("transpose_203")];
+            tensor<fp16, [1, 8, 512]> input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
+            tensor<int32, [1]> input_221_axes_0 = const()[name = tensor<string, []>("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49513792)))];
+            tensor<fp16, [512]> encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49514880)))];
+            tensor<fp16, [1, 8, 512]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = tensor<string, []>("input_221_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_3_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49515968)))];
+            tensor<fp16, [1, 8, 2048]> linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16, x = input_221_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = tensor<string, []>("input_225_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_3_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51613184)))];
+            tensor<fp16, [1, 8, 512]> linear_36_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor<string, []>("linear_36_cast_fp16")];
+            tensor<fp16, []> var_1229_to_fp16 = const()[name = tensor<string, []>("op_1229_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_1230_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1229_to_fp16)[name = tensor<string, []>("op_1230_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1230_cast_fp16)[name = tensor<string, []>("input_231_cast_fp16")];
+            tensor<int32, [1]> input_233_axes_0 = const()[name = tensor<string, []>("input_233_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_3_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53710400)))];
+            tensor<fp16, [512]> encoder_layers_3_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53711488)))];
+            tensor<fp16, [1, 8, 512]> input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = tensor<string, []>("input_233_cast_fp16")];
+            tensor<int32, [4]> cache_17_begin_0 = const()[name = tensor<string, []>("cache_17_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_17_end_0 = const()[name = tensor<string, []>("cache_17_end_0"), val = tensor<int32, [4]>([5, 1, 70, 512])];
+            tensor<bool, [4]> cache_17_end_mask_0 = const()[name = tensor<string, []>("cache_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_17_squeeze_mask_0 = const()[name = tensor<string, []>("cache_17_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_17_cast_fp16 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_17_cast_fp16")];
+            tensor<int32, [4]> cache_19_begin_0 = const()[name = tensor<string, []>("cache_19_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_19_end_0 = const()[name = tensor<string, []>("cache_19_end_0"), val = tensor<int32, [4]>([5, 1, 512, 8])];
+            tensor<bool, [4]> cache_19_end_mask_0 = const()[name = tensor<string, []>("cache_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_19_squeeze_mask_0 = const()[name = tensor<string, []>("cache_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_19_cast_fp16 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_19_cast_fp16")];
+            tensor<int32, [1]> input_235_axes_0 = const()[name = tensor<string, []>("input_235_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53712576)))];
+            tensor<fp16, [512]> encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53713664)))];
+            tensor<fp16, [1, 8, 512]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = tensor<string, []>("input_235_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_4_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53714752)))];
+            tensor<fp16, [1, 8, 2048]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16, x = input_235_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = tensor<string, []>("input_239_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_4_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55811968)))];
+            tensor<fp16, [1, 8, 512]> linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")];
+            tensor<fp16, []> var_1264_to_fp16 = const()[name = tensor<string, []>("op_1264_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_1265_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1264_to_fp16)[name = tensor<string, []>("op_1265_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1265_cast_fp16)[name = tensor<string, []>("input_245_cast_fp16")];
+            tensor<int32, [1]> key_9_axes_0 = const()[name = tensor<string, []>("key_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57909184)))];
+            tensor<fp16, [512]> encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57910272)))];
+            tensor<fp16, [1, 8, 512]> key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")];
+            tensor<bool, []> input_247_interleave_0 = const()[name = tensor<string, []>("input_247_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 78, 512]> input_247_cast_fp16 = concat(axis = var_65, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = tensor<string, []>("input_247_cast_fp16")];
+            tensor<int32, [3]> var_1287_begin_0 = const()[name = tensor<string, []>("op_1287_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
+            tensor<int32, [3]> var_1287_end_0 = const()[name = tensor<string, []>("op_1287_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_1287_end_mask_0 = const()[name = tensor<string, []>("op_1287_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 66, 512]> var_1287_cast_fp16 = slice_by_index(begin = var_1287_begin_0, end = var_1287_end_0, end_mask = var_1287_end_mask_0, x = cache_17_cast_fp16)[name = tensor<string, []>("op_1287_cast_fp16")];
+            tensor<int32, [3]> var_1290_begin_0 = const()[name = tensor<string, []>("op_1290_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_1290_end_0 = const()[name = tensor<string, []>("op_1290_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
+            tensor<bool, [3]> var_1290_end_mask_0 = const()[name = tensor<string, []>("op_1290_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 4, 512]> var_1290_cast_fp16 = slice_by_index(begin = var_1290_begin_0, end = var_1290_end_0, end_mask = var_1290_end_mask_0, x = key_9_cast_fp16)[name = tensor<string, []>("op_1290_cast_fp16")];
+            tensor<bool, []> var_1293_interleave_0 = const()[name = tensor<string, []>("op_1293_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_1293_cast_fp16 = concat(axis = var_65, interleave = var_1293_interleave_0, values = (var_1287_cast_fp16, var_1290_cast_fp16))[name = tensor<string, []>("op_1293_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_4_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57911360)))];
+            tensor<fp16, [1, 8, 512]> linear_39_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16, x = key_9_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")];
+            tensor<int32, [4]> var_1297 = const()[name = tensor<string, []>("op_1297"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 8, 8, 64]> q_25_cast_fp16 = reshape(shape = var_1297, x = linear_39_cast_fp16)[name = tensor<string, []>("q_25_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_4_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58435712)))];
+            tensor<fp16, [1, 78, 512]> linear_40_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16, x = input_247_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")];
+            tensor<int32, [4]> var_1301 = const()[name = tensor<string, []>("op_1301"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> k_17_cast_fp16 = reshape(shape = var_1301, x = linear_40_cast_fp16)[name = tensor<string, []>("k_17_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_4_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58960064)))];
+            tensor<fp16, [1, 78, 512]> linear_41_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16, x = input_247_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")];
+            tensor<int32, [4]> var_1305 = const()[name = tensor<string, []>("op_1305"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> v_9_cast_fp16 = reshape(shape = var_1305, x = linear_41_cast_fp16)[name = tensor<string, []>("v_9_cast_fp16")];
+            tensor<int32, [4]> value_11_perm_0 = const()[name = tensor<string, []>("value_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59484416)))];
+            tensor<fp16, [1, 8, 8, 64]> var_1317_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1317_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59485504)))];
+            tensor<fp16, [1, 8, 8, 64]> var_1319_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1319_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_9_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_111_transpose_x_0 = const()[name = tensor<string, []>("x_111_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_111_transpose_y_0 = const()[name = tensor<string, []>("x_111_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 155]> var_1321_to_fp16 = const()[name = tensor<string, []>("op_1321_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59486592)))];
+            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1319_cast_fp16)[name = tensor<string, []>("transpose_201")];
+            tensor<fp16, [1, 8, 8, 155]> x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = var_1321_to_fp16)[name = tensor<string, []>("x_111_cast_fp16")];
+            tensor<int32, [8]> x_113_pad_0 = const()[name = tensor<string, []>("x_113_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_113_mode_0 = const()[name = tensor<string, []>("x_113_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_131_to_fp16 = const()[name = tensor<string, []>("const_131_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 8, 156]> x_113_cast_fp16 = pad(constant_val = const_131_to_fp16, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111_cast_fp16)[name = tensor<string, []>("x_113_cast_fp16")];
+            tensor<int32, [4]> var_1329 = const()[name = tensor<string, []>("op_1329"), val = tensor<int32, [4]>([1, 8, -1, 8])];
+            tensor<fp16, [1, 8, 156, 8]> x_115_cast_fp16 = reshape(shape = var_1329, x = x_113_cast_fp16)[name = tensor<string, []>("x_115_cast_fp16")];
+            tensor<int32, [4]> var_1333_begin_0 = const()[name = tensor<string, []>("op_1333_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1333_end_0 = const()[name = tensor<string, []>("op_1333_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
+            tensor<bool, [4]> var_1333_end_mask_0 = const()[name = tensor<string, []>("op_1333_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 155, 8]> var_1333_cast_fp16 = slice_by_index(begin = var_1333_begin_0, end = var_1333_end_0, end_mask = var_1333_end_mask_0, x = x_115_cast_fp16)[name = tensor<string, []>("op_1333_cast_fp16")];
+            tensor<int32, [4]> var_1334 = const()[name = tensor<string, []>("op_1334"), val = tensor<int32, [4]>([1, 8, 8, 155])];
+            tensor<fp16, [1, 8, 8, 155]> matrix_bd_17_cast_fp16 = reshape(shape = var_1334, x = var_1333_cast_fp16)[name = tensor<string, []>("matrix_bd_17_cast_fp16")];
+            tensor<bool, []> matrix_ac_9_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_9_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_9_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_9_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_59_perm_0 = const()[name = tensor<string, []>("transpose_59_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_60_perm_0 = const()[name = tensor<string, []>("transpose_60_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 78]> transpose_60 = transpose(perm = transpose_60_perm_0, x = k_17_cast_fp16)[name = tensor<string, []>("transpose_199")];
+            tensor<fp16, [1, 8, 8, 64]> transpose_59 = transpose(perm = transpose_59_perm_0, x = var_1317_cast_fp16)[name = tensor<string, []>("transpose_200")];
+            tensor<fp16, [1, 8, 8, 78]> matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_59, y = transpose_60)[name = tensor<string, []>("matrix_ac_9_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_19_begin_0 = const()[name = tensor<string, []>("matrix_bd_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_19_end_0 = const()[name = tensor<string, []>("matrix_bd_19_end_0"), val = tensor<int32, [4]>([1, 8, 8, 78])];
+            tensor<bool, [4]> matrix_bd_19_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 8, 78]> matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17_cast_fp16)[name = tensor<string, []>("matrix_bd_19_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_1343_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = tensor<string, []>("op_1343_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_17_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_17_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 8, 78]> _inversed_scores_17_cast_fp16 = mul(x = var_1343_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_17_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> scores_19_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_19_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_1349_cast_fp16 = softmax(axis = var_56, x = scores_19_cast_fp16)[name = tensor<string, []>("op_1349_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> input_249_cast_fp16 = select(a = var_40_to_fp16, b = var_1349_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_249_cast_fp16")];
+            tensor<bool, []> x_117_transpose_x_0 = const()[name = tensor<string, []>("x_117_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_117_transpose_y_0 = const()[name = tensor<string, []>("x_117_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 78, 64]> value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_9_cast_fp16)[name = tensor<string, []>("transpose_202")];
+            tensor<fp16, [1, 8, 8, 64]> x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249_cast_fp16, y = value_11_cast_fp16)[name = tensor<string, []>("x_117_cast_fp16")];
+            tensor<int32, [4]> var_1353_perm_0 = const()[name = tensor<string, []>("op_1353_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1354 = const()[name = tensor<string, []>("op_1354"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 8, 8, 64]> var_1353_cast_fp16 = transpose(perm = var_1353_perm_0, x = x_117_cast_fp16)[name = tensor<string, []>("transpose_198")];
+            tensor<fp16, [1, 8, 512]> input_251_cast_fp16 = reshape(shape = var_1354, x = var_1353_cast_fp16)[name = tensor<string, []>("input_251_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_4_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59645376)))];
+            tensor<fp16, [1, 8, 512]> linear_43_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16, x = input_251_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = tensor<string, []>("input_255_cast_fp16")];
+            tensor<int32, [1]> x_121_axes_0 = const()[name = tensor<string, []>("x_121_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60169728)))];
+            tensor<fp16, [512]> encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60170816)))];
+            tensor<fp16, [1, 8, 512]> x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = tensor<string, []>("x_121_cast_fp16")];
+            tensor<int32, [3]> input_257_perm_0 = const()[name = tensor<string, []>("input_257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_259_pad_type_0 = const()[name = tensor<string, []>("input_259_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_259_strides_0 = const()[name = tensor<string, []>("input_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_259_pad_0 = const()[name = tensor<string, []>("input_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_259_dilations_0 = const()[name = tensor<string, []>("input_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_259_groups_0 = const()[name = tensor<string, []>("input_259_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_4_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(60171904)))];
+            tensor<fp16, [1, 512, 8]> input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = tensor<string, []>("transpose_197")];
+            tensor<fp16, [1, 1024, 8]> input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16, x = input_257_cast_fp16)[name = tensor<string, []>("input_259_cast_fp16")];
+            tensor<int32, []> x_123_split_num_splits_0 = const()[name = tensor<string, []>("x_123_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_123_split_axis_0 = const()[name = tensor<string, []>("x_123_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 8]> x_123_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> x_123_split_cast_fp16_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259_cast_fp16)[name = tensor<string, []>("x_123_split_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = tensor<string, []>("x_123_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_123_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> input_261_cast_fp16 = select(a = var_40_to_fp16, b = x_123_cast_fp16, cond = var_551)[name = tensor<string, []>("input_261_cast_fp16")];
+            tensor<bool, []> new_x_19_interleave_0 = const()[name = tensor<string, []>("new_x_19_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 16]> new_x_19_cast_fp16 = concat(axis = var_56, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = tensor<string, []>("new_x_19_cast_fp16")];
+            tensor<int32, [3]> next_cache_9_begin_0 = const()[name = tensor<string, []>("next_cache_9_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_9_end_0 = const()[name = tensor<string, []>("next_cache_9_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> next_cache_9_end_mask_0 = const()[name = tensor<string, []>("next_cache_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 12]> next_cache_9_cast_fp16 = slice_by_index(begin = next_cache_9_begin_0, end = next_cache_9_end_0, end_mask = next_cache_9_end_mask_0, x = new_x_19_cast_fp16)[name = tensor<string, []>("next_cache_9_cast_fp16")];
+            tensor<int32, [3]> var_1395_begin_0 = const()[name = tensor<string, []>("op_1395_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
+            tensor<int32, [3]> var_1395_end_0 = const()[name = tensor<string, []>("op_1395_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> var_1395_end_mask_0 = const()[name = tensor<string, []>("op_1395_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_1395_cast_fp16 = slice_by_index(begin = var_1395_begin_0, end = var_1395_end_0, end_mask = var_1395_end_mask_0, x = next_cache_9_cast_fp16)[name = tensor<string, []>("op_1395_cast_fp16")];
+            tensor<string, []> x_125_pad_type_0 = const()[name = tensor<string, []>("x_125_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_125_groups_0 = const()[name = tensor<string, []>("x_125_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_125_strides_0 = const()[name = tensor<string, []>("x_125_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_125_pad_0 = const()[name = tensor<string, []>("x_125_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_125_dilations_0 = const()[name = tensor<string, []>("x_125_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_4_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61220544)))];
+            tensor<fp16, [1, 512, 8]> x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16, x = new_x_19_cast_fp16)[name = tensor<string, []>("x_125_cast_fp16")];
+            tensor<int32, [3]> input_263_perm_0 = const()[name = tensor<string, []>("input_263_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_127_axes_0 = const()[name = tensor<string, []>("x_127_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61229824)))];
+            tensor<fp16, [512]> encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61230912)))];
+            tensor<fp16, [1, 8, 512]> input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = tensor<string, []>("transpose_196")];
+            tensor<fp16, [1, 8, 512]> x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = tensor<string, []>("x_127_cast_fp16")];
+            tensor<int32, [3]> input_265_perm_0 = const()[name = tensor<string, []>("input_265_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 8]> input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = tensor<string, []>("transpose_195")];
+            tensor<fp16, [1, 512, 8]> input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = tensor<string, []>("input_267_cast_fp16")];
+            tensor<string, []> x_129_pad_type_0 = const()[name = tensor<string, []>("x_129_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_129_strides_0 = const()[name = tensor<string, []>("x_129_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_129_pad_0 = const()[name = tensor<string, []>("x_129_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_129_dilations_0 = const()[name = tensor<string, []>("x_129_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_129_groups_0 = const()[name = tensor<string, []>("x_129_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_4_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61232000)))];
+            tensor<fp16, [1, 512, 8]> x_129_cast_fp16 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16, x = input_267_cast_fp16)[name = tensor<string, []>("x_129_cast_fp16")];
+            tensor<int32, [3]> input_269_perm_0 = const()[name = tensor<string, []>("input_269_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 8, 512]> input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = tensor<string, []>("transpose_194")];
+            tensor<fp16, [1, 8, 512]> input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = tensor<string, []>("input_271_cast_fp16")];
+            tensor<int32, [1]> input_273_axes_0 = const()[name = tensor<string, []>("input_273_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61756352)))];
+            tensor<fp16, [512]> encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61757440)))];
+            tensor<fp16, [1, 8, 512]> input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor<string, []>("input_273_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_4_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61758528)))];
+            tensor<fp16, [1, 8, 2048]> linear_44_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16, x = input_273_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = tensor<string, []>("input_277_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_4_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63855744)))];
+            tensor<fp16, [1, 8, 512]> linear_45_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor<string, []>("linear_45_cast_fp16")];
+            tensor<fp16, []> var_1436_to_fp16 = const()[name = tensor<string, []>("op_1436_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_1437_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1436_to_fp16)[name = tensor<string, []>("op_1437_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1437_cast_fp16)[name = tensor<string, []>("input_283_cast_fp16")];
+            tensor<int32, [1]> input_285_axes_0 = const()[name = tensor<string, []>("input_285_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_4_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65952960)))];
+            tensor<fp16, [512]> encoder_layers_4_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65954048)))];
+            tensor<fp16, [1, 8, 512]> input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = tensor<string, []>("input_285_cast_fp16")];
+            tensor<int32, [4]> cache_21_begin_0 = const()[name = tensor<string, []>("cache_21_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_21_end_0 = const()[name = tensor<string, []>("cache_21_end_0"), val = tensor<int32, [4]>([6, 1, 70, 512])];
+            tensor<bool, [4]> cache_21_end_mask_0 = const()[name = tensor<string, []>("cache_21_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_21_squeeze_mask_0 = const()[name = tensor<string, []>("cache_21_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_21_cast_fp16 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_21_cast_fp16")];
+            tensor<int32, [4]> cache_23_begin_0 = const()[name = tensor<string, []>("cache_23_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_23_end_0 = const()[name = tensor<string, []>("cache_23_end_0"), val = tensor<int32, [4]>([6, 1, 512, 8])];
+            tensor<bool, [4]> cache_23_end_mask_0 = const()[name = tensor<string, []>("cache_23_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_23_squeeze_mask_0 = const()[name = tensor<string, []>("cache_23_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_23_cast_fp16 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_23_cast_fp16")];
+            tensor<int32, [1]> input_287_axes_0 = const()[name = tensor<string, []>("input_287_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65955136)))];
+            tensor<fp16, [512]> encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65956224)))];
+            tensor<fp16, [1, 8, 512]> input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = tensor<string, []>("input_287_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_5_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65957312)))];
+            tensor<fp16, [1, 8, 2048]> linear_46_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor<string, []>("linear_46_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = tensor<string, []>("input_291_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_5_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68054528)))];
+            tensor<fp16, [1, 8, 512]> linear_47_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16, x = input_291_cast_fp16)[name = tensor<string, []>("linear_47_cast_fp16")];
+            tensor<fp16, []> var_1471_to_fp16 = const()[name = tensor<string, []>("op_1471_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_1472_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1471_to_fp16)[name = tensor<string, []>("op_1472_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1472_cast_fp16)[name = tensor<string, []>("input_297_cast_fp16")];
+            tensor<int32, [1]> key_11_axes_0 = const()[name = tensor<string, []>("key_11_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70151744)))];
+            tensor<fp16, [512]> encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70152832)))];
+            tensor<fp16, [1, 8, 512]> key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = tensor<string, []>("key_11_cast_fp16")];
+            tensor<bool, []> input_299_interleave_0 = const()[name = tensor<string, []>("input_299_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 78, 512]> input_299_cast_fp16 = concat(axis = var_65, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = tensor<string, []>("input_299_cast_fp16")];
+            tensor<int32, [3]> var_1494_begin_0 = const()[name = tensor<string, []>("op_1494_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
+            tensor<int32, [3]> var_1494_end_0 = const()[name = tensor<string, []>("op_1494_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_1494_end_mask_0 = const()[name = tensor<string, []>("op_1494_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 66, 512]> var_1494_cast_fp16 = slice_by_index(begin = var_1494_begin_0, end = var_1494_end_0, end_mask = var_1494_end_mask_0, x = cache_21_cast_fp16)[name = tensor<string, []>("op_1494_cast_fp16")];
+            tensor<int32, [3]> var_1497_begin_0 = const()[name = tensor<string, []>("op_1497_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_1497_end_0 = const()[name = tensor<string, []>("op_1497_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
+            tensor<bool, [3]> var_1497_end_mask_0 = const()[name = tensor<string, []>("op_1497_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 4, 512]> var_1497_cast_fp16 = slice_by_index(begin = var_1497_begin_0, end = var_1497_end_0, end_mask = var_1497_end_mask_0, x = key_11_cast_fp16)[name = tensor<string, []>("op_1497_cast_fp16")];
+            tensor<bool, []> var_1500_interleave_0 = const()[name = tensor<string, []>("op_1500_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_1500_cast_fp16 = concat(axis = var_65, interleave = var_1500_interleave_0, values = (var_1494_cast_fp16, var_1497_cast_fp16))[name = tensor<string, []>("op_1500_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_5_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70153920)))];
+            tensor<fp16, [1, 8, 512]> linear_48_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16, x = key_11_cast_fp16)[name = tensor<string, []>("linear_48_cast_fp16")];
+            tensor<int32, [4]> var_1504 = const()[name = tensor<string, []>("op_1504"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 8, 8, 64]> q_31_cast_fp16 = reshape(shape = var_1504, x = linear_48_cast_fp16)[name = tensor<string, []>("q_31_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_5_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70678272)))];
+            tensor<fp16, [1, 78, 512]> linear_49_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16, x = input_299_cast_fp16)[name = tensor<string, []>("linear_49_cast_fp16")];
+            tensor<int32, [4]> var_1508 = const()[name = tensor<string, []>("op_1508"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> k_21_cast_fp16 = reshape(shape = var_1508, x = linear_49_cast_fp16)[name = tensor<string, []>("k_21_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_5_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71202624)))];
+            tensor<fp16, [1, 78, 512]> linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16, x = input_299_cast_fp16)[name = tensor<string, []>("linear_50_cast_fp16")];
+            tensor<int32, [4]> var_1512 = const()[name = tensor<string, []>("op_1512"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> v_11_cast_fp16 = reshape(shape = var_1512, x = linear_50_cast_fp16)[name = tensor<string, []>("v_11_cast_fp16")];
+            tensor<int32, [4]> value_13_perm_0 = const()[name = tensor<string, []>("value_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71726976)))];
+            tensor<fp16, [1, 8, 8, 64]> var_1524_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1524_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71728064)))];
+            tensor<fp16, [1, 8, 8, 64]> var_1526_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1526_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_11_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_137_transpose_x_0 = const()[name = tensor<string, []>("x_137_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_137_transpose_y_0 = const()[name = tensor<string, []>("x_137_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 155]> var_1528_to_fp16 = const()[name = tensor<string, []>("op_1528_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71729152)))];
+            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1526_cast_fp16)[name = tensor<string, []>("transpose_192")];
+            tensor<fp16, [1, 8, 8, 155]> x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = var_1528_to_fp16)[name = tensor<string, []>("x_137_cast_fp16")];
+            tensor<int32, [8]> x_139_pad_0 = const()[name = tensor<string, []>("x_139_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_139_mode_0 = const()[name = tensor<string, []>("x_139_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_144_to_fp16 = const()[name = tensor<string, []>("const_144_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 8, 156]> x_139_cast_fp16 = pad(constant_val = const_144_to_fp16, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137_cast_fp16)[name = tensor<string, []>("x_139_cast_fp16")];
+            tensor<int32, [4]> var_1536 = const()[name = tensor<string, []>("op_1536"), val = tensor<int32, [4]>([1, 8, -1, 8])];
+            tensor<fp16, [1, 8, 156, 8]> x_141_cast_fp16 = reshape(shape = var_1536, x = x_139_cast_fp16)[name = tensor<string, []>("x_141_cast_fp16")];
+            tensor<int32, [4]> var_1540_begin_0 = const()[name = tensor<string, []>("op_1540_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1540_end_0 = const()[name = tensor<string, []>("op_1540_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
+            tensor<bool, [4]> var_1540_end_mask_0 = const()[name = tensor<string, []>("op_1540_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 155, 8]> var_1540_cast_fp16 = slice_by_index(begin = var_1540_begin_0, end = var_1540_end_0, end_mask = var_1540_end_mask_0, x = x_141_cast_fp16)[name = tensor<string, []>("op_1540_cast_fp16")];
+            tensor<int32, [4]> var_1541 = const()[name = tensor<string, []>("op_1541"), val = tensor<int32, [4]>([1, 8, 8, 155])];
+            tensor<fp16, [1, 8, 8, 155]> matrix_bd_21_cast_fp16 = reshape(shape = var_1541, x = var_1540_cast_fp16)[name = tensor<string, []>("matrix_bd_21_cast_fp16")];
+            tensor<bool, []> matrix_ac_11_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_11_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_11_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_11_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_61_perm_0 = const()[name = tensor<string, []>("transpose_61_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_62_perm_0 = const()[name = tensor<string, []>("transpose_62_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 78]> transpose_62 = transpose(perm = transpose_62_perm_0, x = k_21_cast_fp16)[name = tensor<string, []>("transpose_190")];
+            tensor<fp16, [1, 8, 8, 64]> transpose_61 = transpose(perm = transpose_61_perm_0, x = var_1524_cast_fp16)[name = tensor<string, []>("transpose_191")];
+            tensor<fp16, [1, 8, 8, 78]> matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_61, y = transpose_62)[name = tensor<string, []>("matrix_ac_11_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_23_begin_0 = const()[name = tensor<string, []>("matrix_bd_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_23_end_0 = const()[name = tensor<string, []>("matrix_bd_23_end_0"), val = tensor<int32, [4]>([1, 8, 8, 78])];
+            tensor<bool, [4]> matrix_bd_23_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 8, 78]> matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21_cast_fp16)[name = tensor<string, []>("matrix_bd_23_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_1550_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = tensor<string, []>("op_1550_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_21_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_21_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 8, 78]> _inversed_scores_21_cast_fp16 = mul(x = var_1550_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_21_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> scores_23_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_23_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_1556_cast_fp16 = softmax(axis = var_56, x = scores_23_cast_fp16)[name = tensor<string, []>("op_1556_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> input_301_cast_fp16 = select(a = var_40_to_fp16, b = var_1556_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_301_cast_fp16")];
+            tensor<bool, []> x_143_transpose_x_0 = const()[name = tensor<string, []>("x_143_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_143_transpose_y_0 = const()[name = tensor<string, []>("x_143_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 78, 64]> value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_11_cast_fp16)[name = tensor<string, []>("transpose_193")];
+            tensor<fp16, [1, 8, 8, 64]> x_143_cast_fp16 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301_cast_fp16, y = value_13_cast_fp16)[name = tensor<string, []>("x_143_cast_fp16")];
+            tensor<int32, [4]> var_1560_perm_0 = const()[name = tensor<string, []>("op_1560_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1561 = const()[name = tensor<string, []>("op_1561"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 8, 8, 64]> var_1560_cast_fp16 = transpose(perm = var_1560_perm_0, x = x_143_cast_fp16)[name = tensor<string, []>("transpose_189")];
+            tensor<fp16, [1, 8, 512]> input_303_cast_fp16 = reshape(shape = var_1561, x = var_1560_cast_fp16)[name = tensor<string, []>("input_303_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_5_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71887936)))];
+            tensor<fp16, [1, 8, 512]> linear_52_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16, x = input_303_cast_fp16)[name = tensor<string, []>("linear_52_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = tensor<string, []>("input_307_cast_fp16")];
+            tensor<int32, [1]> x_147_axes_0 = const()[name = tensor<string, []>("x_147_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72412288)))];
+            tensor<fp16, [512]> encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72413376)))];
+            tensor<fp16, [1, 8, 512]> x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = tensor<string, []>("x_147_cast_fp16")];
+            tensor<int32, [3]> input_309_perm_0 = const()[name = tensor<string, []>("input_309_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_311_pad_type_0 = const()[name = tensor<string, []>("input_311_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_311_strides_0 = const()[name = tensor<string, []>("input_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_311_pad_0 = const()[name = tensor<string, []>("input_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_311_dilations_0 = const()[name = tensor<string, []>("input_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_311_groups_0 = const()[name = tensor<string, []>("input_311_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_5_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72414464)))];
+            tensor<fp16, [1, 512, 8]> input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = tensor<string, []>("transpose_188")];
+            tensor<fp16, [1, 1024, 8]> input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16, x = input_309_cast_fp16)[name = tensor<string, []>("input_311_cast_fp16")];
+            tensor<int32, []> x_149_split_num_splits_0 = const()[name = tensor<string, []>("x_149_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_149_split_axis_0 = const()[name = tensor<string, []>("x_149_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 8]> x_149_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> x_149_split_cast_fp16_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311_cast_fp16)[name = tensor<string, []>("x_149_split_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = tensor<string, []>("x_149_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_149_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> input_313_cast_fp16 = select(a = var_40_to_fp16, b = x_149_cast_fp16, cond = var_551)[name = tensor<string, []>("input_313_cast_fp16")];
+            tensor<bool, []> new_x_23_interleave_0 = const()[name = tensor<string, []>("new_x_23_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 16]> new_x_23_cast_fp16 = concat(axis = var_56, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = tensor<string, []>("new_x_23_cast_fp16")];
+            tensor<int32, [3]> next_cache_11_begin_0 = const()[name = tensor<string, []>("next_cache_11_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_11_end_0 = const()[name = tensor<string, []>("next_cache_11_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> next_cache_11_end_mask_0 = const()[name = tensor<string, []>("next_cache_11_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 12]> next_cache_11_cast_fp16 = slice_by_index(begin = next_cache_11_begin_0, end = next_cache_11_end_0, end_mask = next_cache_11_end_mask_0, x = new_x_23_cast_fp16)[name = tensor<string, []>("next_cache_11_cast_fp16")];
+            tensor<int32, [3]> var_1602_begin_0 = const()[name = tensor<string, []>("op_1602_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
+            tensor<int32, [3]> var_1602_end_0 = const()[name = tensor<string, []>("op_1602_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> var_1602_end_mask_0 = const()[name = tensor<string, []>("op_1602_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_1602_cast_fp16 = slice_by_index(begin = var_1602_begin_0, end = var_1602_end_0, end_mask = var_1602_end_mask_0, x = next_cache_11_cast_fp16)[name = tensor<string, []>("op_1602_cast_fp16")];
+            tensor<string, []> x_151_pad_type_0 = const()[name = tensor<string, []>("x_151_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_151_groups_0 = const()[name = tensor<string, []>("x_151_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_151_strides_0 = const()[name = tensor<string, []>("x_151_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_151_pad_0 = const()[name = tensor<string, []>("x_151_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_151_dilations_0 = const()[name = tensor<string, []>("x_151_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_5_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73463104)))];
+            tensor<fp16, [1, 512, 8]> x_151_cast_fp16 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16, x = new_x_23_cast_fp16)[name = tensor<string, []>("x_151_cast_fp16")];
+            tensor<int32, [3]> input_315_perm_0 = const()[name = tensor<string, []>("input_315_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_153_axes_0 = const()[name = tensor<string, []>("x_153_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73472384)))];
+            tensor<fp16, [512]> encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73473472)))];
+            tensor<fp16, [1, 8, 512]> input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = tensor<string, []>("transpose_187")];
+            tensor<fp16, [1, 8, 512]> x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = tensor<string, []>("x_153_cast_fp16")];
+            tensor<int32, [3]> input_317_perm_0 = const()[name = tensor<string, []>("input_317_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 8]> input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = tensor<string, []>("transpose_186")];
+            tensor<fp16, [1, 512, 8]> input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = tensor<string, []>("input_319_cast_fp16")];
+            tensor<string, []> x_155_pad_type_0 = const()[name = tensor<string, []>("x_155_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_155_strides_0 = const()[name = tensor<string, []>("x_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_155_pad_0 = const()[name = tensor<string, []>("x_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_155_dilations_0 = const()[name = tensor<string, []>("x_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_155_groups_0 = const()[name = tensor<string, []>("x_155_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_5_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73474560)))];
+            tensor<fp16, [1, 512, 8]> x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16, x = input_319_cast_fp16)[name = tensor<string, []>("x_155_cast_fp16")];
+            tensor<int32, [3]> input_321_perm_0 = const()[name = tensor<string, []>("input_321_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 8, 512]> input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = tensor<string, []>("transpose_185")];
+            tensor<fp16, [1, 8, 512]> input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = tensor<string, []>("input_323_cast_fp16")];
+            tensor<int32, [1]> input_325_axes_0 = const()[name = tensor<string, []>("input_325_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73998912)))];
+            tensor<fp16, [512]> encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74000000)))];
+            tensor<fp16, [1, 8, 512]> input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = tensor<string, []>("input_325_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_5_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74001088)))];
+            tensor<fp16, [1, 8, 2048]> linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16, x = input_325_cast_fp16)[name = tensor<string, []>("linear_53_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = tensor<string, []>("input_329_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_5_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76098304)))];
+            tensor<fp16, [1, 8, 512]> linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16, x = input_329_cast_fp16)[name = tensor<string, []>("linear_54_cast_fp16")];
+            tensor<fp16, []> var_1643_to_fp16 = const()[name = tensor<string, []>("op_1643_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_1644_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1643_to_fp16)[name = tensor<string, []>("op_1644_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1644_cast_fp16)[name = tensor<string, []>("input_335_cast_fp16")];
+            tensor<int32, [1]> input_337_axes_0 = const()[name = tensor<string, []>("input_337_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_5_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78195520)))];
+            tensor<fp16, [512]> encoder_layers_5_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78196608)))];
+            tensor<fp16, [1, 8, 512]> input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = tensor<string, []>("input_337_cast_fp16")];
+            tensor<int32, [4]> cache_25_begin_0 = const()[name = tensor<string, []>("cache_25_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_25_end_0 = const()[name = tensor<string, []>("cache_25_end_0"), val = tensor<int32, [4]>([7, 1, 70, 512])];
+            tensor<bool, [4]> cache_25_end_mask_0 = const()[name = tensor<string, []>("cache_25_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_25_squeeze_mask_0 = const()[name = tensor<string, []>("cache_25_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_25_cast_fp16 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_25_cast_fp16")];
+            tensor<int32, [4]> cache_27_begin_0 = const()[name = tensor<string, []>("cache_27_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_27_end_0 = const()[name = tensor<string, []>("cache_27_end_0"), val = tensor<int32, [4]>([7, 1, 512, 8])];
+            tensor<bool, [4]> cache_27_end_mask_0 = const()[name = tensor<string, []>("cache_27_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_27_squeeze_mask_0 = const()[name = tensor<string, []>("cache_27_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_27_cast_fp16 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_27_cast_fp16")];
+            tensor<int32, [1]> input_339_axes_0 = const()[name = tensor<string, []>("input_339_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78197696)))];
+            tensor<fp16, [512]> encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78198784)))];
+            tensor<fp16, [1, 8, 512]> input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = tensor<string, []>("input_339_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_6_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78199872)))];
+            tensor<fp16, [1, 8, 2048]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16, x = input_339_cast_fp16)[name = tensor<string, []>("linear_55_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = tensor<string, []>("input_343_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_6_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80297088)))];
+            tensor<fp16, [1, 8, 512]> linear_56_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16, x = input_343_cast_fp16)[name = tensor<string, []>("linear_56_cast_fp16")];
+            tensor<fp16, []> var_1678_to_fp16 = const()[name = tensor<string, []>("op_1678_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_1679_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1678_to_fp16)[name = tensor<string, []>("op_1679_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1679_cast_fp16)[name = tensor<string, []>("input_349_cast_fp16")];
+            tensor<int32, [1]> key_13_axes_0 = const()[name = tensor<string, []>("key_13_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82394304)))];
+            tensor<fp16, [512]> encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82395392)))];
+            tensor<fp16, [1, 8, 512]> key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")];
+            tensor<bool, []> input_351_interleave_0 = const()[name = tensor<string, []>("input_351_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 78, 512]> input_351_cast_fp16 = concat(axis = var_65, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = tensor<string, []>("input_351_cast_fp16")];
+            tensor<int32, [3]> var_1701_begin_0 = const()[name = tensor<string, []>("op_1701_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
+            tensor<int32, [3]> var_1701_end_0 = const()[name = tensor<string, []>("op_1701_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_1701_end_mask_0 = const()[name = tensor<string, []>("op_1701_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 66, 512]> var_1701_cast_fp16 = slice_by_index(begin = var_1701_begin_0, end = var_1701_end_0, end_mask = var_1701_end_mask_0, x = cache_25_cast_fp16)[name = tensor<string, []>("op_1701_cast_fp16")];
+            tensor<int32, [3]> var_1704_begin_0 = const()[name = tensor<string, []>("op_1704_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_1704_end_0 = const()[name = tensor<string, []>("op_1704_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
+            tensor<bool, [3]> var_1704_end_mask_0 = const()[name = tensor<string, []>("op_1704_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 4, 512]> var_1704_cast_fp16 = slice_by_index(begin = var_1704_begin_0, end = var_1704_end_0, end_mask = var_1704_end_mask_0, x = key_13_cast_fp16)[name = tensor<string, []>("op_1704_cast_fp16")];
+            tensor<bool, []> var_1707_interleave_0 = const()[name = tensor<string, []>("op_1707_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_1707_cast_fp16 = concat(axis = var_65, interleave = var_1707_interleave_0, values = (var_1701_cast_fp16, var_1704_cast_fp16))[name = tensor<string, []>("op_1707_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_6_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82396480)))];
+            tensor<fp16, [1, 8, 512]> linear_57_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16, x = key_13_cast_fp16)[name = tensor<string, []>("linear_57_cast_fp16")];
+            tensor<int32, [4]> var_1711 = const()[name = tensor<string, []>("op_1711"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 8, 8, 64]> q_37_cast_fp16 = reshape(shape = var_1711, x = linear_57_cast_fp16)[name = tensor<string, []>("q_37_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_6_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82920832)))];
+            tensor<fp16, [1, 78, 512]> linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16, x = input_351_cast_fp16)[name = tensor<string, []>("linear_58_cast_fp16")];
+            tensor<int32, [4]> var_1715 = const()[name = tensor<string, []>("op_1715"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> k_25_cast_fp16 = reshape(shape = var_1715, x = linear_58_cast_fp16)[name = tensor<string, []>("k_25_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_6_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83445184)))];
+            tensor<fp16, [1, 78, 512]> linear_59_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16, x = input_351_cast_fp16)[name = tensor<string, []>("linear_59_cast_fp16")];
+            tensor<int32, [4]> var_1719 = const()[name = tensor<string, []>("op_1719"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> v_13_cast_fp16 = reshape(shape = var_1719, x = linear_59_cast_fp16)[name = tensor<string, []>("v_13_cast_fp16")];
+            tensor<int32, [4]> value_15_perm_0 = const()[name = tensor<string, []>("value_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83969536)))];
+            tensor<fp16, [1, 8, 8, 64]> var_1731_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1731_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83970624)))];
+            tensor<fp16, [1, 8, 8, 64]> var_1733_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1733_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_13_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_163_transpose_x_0 = const()[name = tensor<string, []>("x_163_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_163_transpose_y_0 = const()[name = tensor<string, []>("x_163_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 155]> var_1735_to_fp16 = const()[name = tensor<string, []>("op_1735_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83971712)))];
+            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1733_cast_fp16)[name = tensor<string, []>("transpose_183")];
+            tensor<fp16, [1, 8, 8, 155]> x_163_cast_fp16 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = var_1735_to_fp16)[name = tensor<string, []>("x_163_cast_fp16")];
+            tensor<int32, [8]> x_165_pad_0 = const()[name = tensor<string, []>("x_165_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_165_mode_0 = const()[name = tensor<string, []>("x_165_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_157_to_fp16 = const()[name = tensor<string, []>("const_157_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 8, 156]> x_165_cast_fp16 = pad(constant_val = const_157_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = tensor<string, []>("x_165_cast_fp16")];
+            tensor<int32, [4]> var_1743 = const()[name = tensor<string, []>("op_1743"), val = tensor<int32, [4]>([1, 8, -1, 8])];
+            tensor<fp16, [1, 8, 156, 8]> x_167_cast_fp16 = reshape(shape = var_1743, x = x_165_cast_fp16)[name = tensor<string, []>("x_167_cast_fp16")];
+            tensor<int32, [4]> var_1747_begin_0 = const()[name = tensor<string, []>("op_1747_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1747_end_0 = const()[name = tensor<string, []>("op_1747_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
+            tensor<bool, [4]> var_1747_end_mask_0 = const()[name = tensor<string, []>("op_1747_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 155, 8]> var_1747_cast_fp16 = slice_by_index(begin = var_1747_begin_0, end = var_1747_end_0, end_mask = var_1747_end_mask_0, x = x_167_cast_fp16)[name = tensor<string, []>("op_1747_cast_fp16")];
+            tensor<int32, [4]> var_1748 = const()[name = tensor<string, []>("op_1748"), val = tensor<int32, [4]>([1, 8, 8, 155])];
+            tensor<fp16, [1, 8, 8, 155]> matrix_bd_25_cast_fp16 = reshape(shape = var_1748, x = var_1747_cast_fp16)[name = tensor<string, []>("matrix_bd_25_cast_fp16")];
+            tensor<bool, []> matrix_ac_13_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_13_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_63_perm_0 = const()[name = tensor<string, []>("transpose_63_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_64_perm_0 = const()[name = tensor<string, []>("transpose_64_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 78]> transpose_64 = transpose(perm = transpose_64_perm_0, x = k_25_cast_fp16)[name = tensor<string, []>("transpose_181")];
+            tensor<fp16, [1, 8, 8, 64]> transpose_63 = transpose(perm = transpose_63_perm_0, x = var_1731_cast_fp16)[name = tensor<string, []>("transpose_182")];
+            tensor<fp16, [1, 8, 8, 78]> matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_63, y = transpose_64)[name = tensor<string, []>("matrix_ac_13_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_27_begin_0 = const()[name = tensor<string, []>("matrix_bd_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_27_end_0 = const()[name = tensor<string, []>("matrix_bd_27_end_0"), val = tensor<int32, [4]>([1, 8, 8, 78])];
+            tensor<bool, [4]> matrix_bd_27_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 8, 78]> matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25_cast_fp16)[name = tensor<string, []>("matrix_bd_27_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_1757_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = tensor<string, []>("op_1757_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_25_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_25_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 8, 78]> _inversed_scores_25_cast_fp16 = mul(x = var_1757_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_25_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> scores_27_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_27_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_1763_cast_fp16 = softmax(axis = var_56, x = scores_27_cast_fp16)[name = tensor<string, []>("op_1763_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> input_353_cast_fp16 = select(a = var_40_to_fp16, b = var_1763_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_353_cast_fp16")];
+            tensor<bool, []> x_169_transpose_x_0 = const()[name = tensor<string, []>("x_169_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_169_transpose_y_0 = const()[name = tensor<string, []>("x_169_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 78, 64]> value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_13_cast_fp16)[name = tensor<string, []>("transpose_184")];
+            tensor<fp16, [1, 8, 8, 64]> x_169_cast_fp16 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353_cast_fp16, y = value_15_cast_fp16)[name = tensor<string, []>("x_169_cast_fp16")];
+            tensor<int32, [4]> var_1767_perm_0 = const()[name = tensor<string, []>("op_1767_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1768 = const()[name = tensor<string, []>("op_1768"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 8, 8, 64]> var_1767_cast_fp16 = transpose(perm = var_1767_perm_0, x = x_169_cast_fp16)[name = tensor<string, []>("transpose_180")];
+            tensor<fp16, [1, 8, 512]> input_355_cast_fp16 = reshape(shape = var_1768, x = var_1767_cast_fp16)[name = tensor<string, []>("input_355_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_6_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84130496)))];
+            tensor<fp16, [1, 8, 512]> linear_61_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16, x = input_355_cast_fp16)[name = tensor<string, []>("linear_61_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = tensor<string, []>("input_359_cast_fp16")];
+            tensor<int32, [1]> x_173_axes_0 = const()[name = tensor<string, []>("x_173_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84654848)))];
+            tensor<fp16, [512]> encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84655936)))];
+            tensor<fp16, [1, 8, 512]> x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = tensor<string, []>("x_173_cast_fp16")];
+            tensor<int32, [3]> input_361_perm_0 = const()[name = tensor<string, []>("input_361_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_363_pad_type_0 = const()[name = tensor<string, []>("input_363_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_363_strides_0 = const()[name = tensor<string, []>("input_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_363_pad_0 = const()[name = tensor<string, []>("input_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_363_dilations_0 = const()[name = tensor<string, []>("input_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_363_groups_0 = const()[name = tensor<string, []>("input_363_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_6_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84657024)))];
+            tensor<fp16, [1, 512, 8]> input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = tensor<string, []>("transpose_179")];
+            tensor<fp16, [1, 1024, 8]> input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16, x = input_361_cast_fp16)[name = tensor<string, []>("input_363_cast_fp16")];
+            tensor<int32, []> x_175_split_num_splits_0 = const()[name = tensor<string, []>("x_175_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_175_split_axis_0 = const()[name = tensor<string, []>("x_175_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 8]> x_175_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> x_175_split_cast_fp16_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363_cast_fp16)[name = tensor<string, []>("x_175_split_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = tensor<string, []>("x_175_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_175_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> input_365_cast_fp16 = select(a = var_40_to_fp16, b = x_175_cast_fp16, cond = var_551)[name = tensor<string, []>("input_365_cast_fp16")];
+            tensor<bool, []> new_x_27_interleave_0 = const()[name = tensor<string, []>("new_x_27_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 16]> new_x_27_cast_fp16 = concat(axis = var_56, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = tensor<string, []>("new_x_27_cast_fp16")];
+            tensor<int32, [3]> next_cache_13_begin_0 = const()[name = tensor<string, []>("next_cache_13_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_13_end_0 = const()[name = tensor<string, []>("next_cache_13_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> next_cache_13_end_mask_0 = const()[name = tensor<string, []>("next_cache_13_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 12]> next_cache_13_cast_fp16 = slice_by_index(begin = next_cache_13_begin_0, end = next_cache_13_end_0, end_mask = next_cache_13_end_mask_0, x = new_x_27_cast_fp16)[name = tensor<string, []>("next_cache_13_cast_fp16")];
+            tensor<int32, [3]> var_1809_begin_0 = const()[name = tensor<string, []>("op_1809_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
+            tensor<int32, [3]> var_1809_end_0 = const()[name = tensor<string, []>("op_1809_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> var_1809_end_mask_0 = const()[name = tensor<string, []>("op_1809_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_1809_cast_fp16 = slice_by_index(begin = var_1809_begin_0, end = var_1809_end_0, end_mask = var_1809_end_mask_0, x = next_cache_13_cast_fp16)[name = tensor<string, []>("op_1809_cast_fp16")];
+            tensor<string, []> x_177_pad_type_0 = const()[name = tensor<string, []>("x_177_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_177_groups_0 = const()[name = tensor<string, []>("x_177_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_177_strides_0 = const()[name = tensor<string, []>("x_177_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_177_pad_0 = const()[name = tensor<string, []>("x_177_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_177_dilations_0 = const()[name = tensor<string, []>("x_177_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_6_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85705664)))];
+            tensor<fp16, [1, 512, 8]> x_177_cast_fp16 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16, x = new_x_27_cast_fp16)[name = tensor<string, []>("x_177_cast_fp16")];
+            tensor<int32, [3]> input_367_perm_0 = const()[name = tensor<string, []>("input_367_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_179_axes_0 = const()[name = tensor<string, []>("x_179_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85714944)))];
+            tensor<fp16, [512]> encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85716032)))];
+            tensor<fp16, [1, 8, 512]> input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = tensor<string, []>("transpose_178")];
+            tensor<fp16, [1, 8, 512]> x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = tensor<string, []>("x_179_cast_fp16")];
+            tensor<int32, [3]> input_369_perm_0 = const()[name = tensor<string, []>("input_369_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 8]> input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = tensor<string, []>("transpose_177")];
+            tensor<fp16, [1, 512, 8]> input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = tensor<string, []>("input_371_cast_fp16")];
+            tensor<string, []> x_181_pad_type_0 = const()[name = tensor<string, []>("x_181_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_181_strides_0 = const()[name = tensor<string, []>("x_181_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_181_pad_0 = const()[name = tensor<string, []>("x_181_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_181_dilations_0 = const()[name = tensor<string, []>("x_181_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_181_groups_0 = const()[name = tensor<string, []>("x_181_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_6_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85717120)))];
+            tensor<fp16, [1, 512, 8]> x_181_cast_fp16 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16, x = input_371_cast_fp16)[name = tensor<string, []>("x_181_cast_fp16")];
+            tensor<int32, [3]> input_373_perm_0 = const()[name = tensor<string, []>("input_373_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 8, 512]> input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = tensor<string, []>("transpose_176")];
+            tensor<fp16, [1, 8, 512]> input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = tensor<string, []>("input_375_cast_fp16")];
+            tensor<int32, [1]> input_377_axes_0 = const()[name = tensor<string, []>("input_377_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86241472)))];
+            tensor<fp16, [512]> encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86242560)))];
+            tensor<fp16, [1, 8, 512]> input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = tensor<string, []>("input_377_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_6_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86243648)))];
+            tensor<fp16, [1, 8, 2048]> linear_62_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16, x = input_377_cast_fp16)[name = tensor<string, []>("linear_62_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = tensor<string, []>("input_381_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_6_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88340864)))];
+            tensor<fp16, [1, 8, 512]> linear_63_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16, x = input_381_cast_fp16)[name = tensor<string, []>("linear_63_cast_fp16")];
+            tensor<fp16, []> var_1850_to_fp16 = const()[name = tensor<string, []>("op_1850_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_1851_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1850_to_fp16)[name = tensor<string, []>("op_1851_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1851_cast_fp16)[name = tensor<string, []>("input_387_cast_fp16")];
+            tensor<int32, [1]> input_389_axes_0 = const()[name = tensor<string, []>("input_389_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_6_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90438080)))];
+            tensor<fp16, [512]> encoder_layers_6_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90439168)))];
+            tensor<fp16, [1, 8, 512]> input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = tensor<string, []>("input_389_cast_fp16")];
+            tensor<int32, [4]> cache_29_begin_0 = const()[name = tensor<string, []>("cache_29_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_29_end_0 = const()[name = tensor<string, []>("cache_29_end_0"), val = tensor<int32, [4]>([8, 1, 70, 512])];
+            tensor<bool, [4]> cache_29_end_mask_0 = const()[name = tensor<string, []>("cache_29_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_29_squeeze_mask_0 = const()[name = tensor<string, []>("cache_29_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_29_cast_fp16 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_29_cast_fp16")];
+            tensor<int32, [4]> cache_31_begin_0 = const()[name = tensor<string, []>("cache_31_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_31_end_0 = const()[name = tensor<string, []>("cache_31_end_0"), val = tensor<int32, [4]>([8, 1, 512, 8])];
+            tensor<bool, [4]> cache_31_end_mask_0 = const()[name = tensor<string, []>("cache_31_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_31_squeeze_mask_0 = const()[name = tensor<string, []>("cache_31_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_31_cast_fp16 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_31_cast_fp16")];
+            tensor<int32, [1]> input_391_axes_0 = const()[name = tensor<string, []>("input_391_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90440256)))];
+            tensor<fp16, [512]> encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90441344)))];
+            tensor<fp16, [1, 8, 512]> input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = tensor<string, []>("input_391_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_7_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90442432)))];
+            tensor<fp16, [1, 8, 2048]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16, x = input_391_cast_fp16)[name = tensor<string, []>("linear_64_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = tensor<string, []>("input_395_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_7_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92539648)))];
+            tensor<fp16, [1, 8, 512]> linear_65_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16, x = input_395_cast_fp16)[name = tensor<string, []>("linear_65_cast_fp16")];
+            tensor<fp16, []> var_1885_to_fp16 = const()[name = tensor<string, []>("op_1885_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_1886_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1885_to_fp16)[name = tensor<string, []>("op_1886_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1886_cast_fp16)[name = tensor<string, []>("input_401_cast_fp16")];
+            tensor<int32, [1]> key_15_axes_0 = const()[name = tensor<string, []>("key_15_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94636864)))];
+            tensor<fp16, [512]> encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94637952)))];
+            tensor<fp16, [1, 8, 512]> key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = tensor<string, []>("key_15_cast_fp16")];
+            tensor<bool, []> input_403_interleave_0 = const()[name = tensor<string, []>("input_403_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 78, 512]> input_403_cast_fp16 = concat(axis = var_65, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = tensor<string, []>("input_403_cast_fp16")];
+            tensor<int32, [3]> var_1908_begin_0 = const()[name = tensor<string, []>("op_1908_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
+            tensor<int32, [3]> var_1908_end_0 = const()[name = tensor<string, []>("op_1908_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_1908_end_mask_0 = const()[name = tensor<string, []>("op_1908_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 66, 512]> var_1908_cast_fp16 = slice_by_index(begin = var_1908_begin_0, end = var_1908_end_0, end_mask = var_1908_end_mask_0, x = cache_29_cast_fp16)[name = tensor<string, []>("op_1908_cast_fp16")];
+            tensor<int32, [3]> var_1911_begin_0 = const()[name = tensor<string, []>("op_1911_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_1911_end_0 = const()[name = tensor<string, []>("op_1911_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
+            tensor<bool, [3]> var_1911_end_mask_0 = const()[name = tensor<string, []>("op_1911_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 4, 512]> var_1911_cast_fp16 = slice_by_index(begin = var_1911_begin_0, end = var_1911_end_0, end_mask = var_1911_end_mask_0, x = key_15_cast_fp16)[name = tensor<string, []>("op_1911_cast_fp16")];
+            tensor<bool, []> var_1914_interleave_0 = const()[name = tensor<string, []>("op_1914_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_1914_cast_fp16 = concat(axis = var_65, interleave = var_1914_interleave_0, values = (var_1908_cast_fp16, var_1911_cast_fp16))[name = tensor<string, []>("op_1914_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_7_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94639040)))];
+            tensor<fp16, [1, 8, 512]> linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16, x = key_15_cast_fp16)[name = tensor<string, []>("linear_66_cast_fp16")];
+            tensor<int32, [4]> var_1918 = const()[name = tensor<string, []>("op_1918"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 8, 8, 64]> q_43_cast_fp16 = reshape(shape = var_1918, x = linear_66_cast_fp16)[name = tensor<string, []>("q_43_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_7_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(95163392)))];
+            tensor<fp16, [1, 78, 512]> linear_67_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16, x = input_403_cast_fp16)[name = tensor<string, []>("linear_67_cast_fp16")];
+            tensor<int32, [4]> var_1922 = const()[name = tensor<string, []>("op_1922"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> k_29_cast_fp16 = reshape(shape = var_1922, x = linear_67_cast_fp16)[name = tensor<string, []>("k_29_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_7_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(95687744)))];
+            tensor<fp16, [1, 78, 512]> linear_68_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16, x = input_403_cast_fp16)[name = tensor<string, []>("linear_68_cast_fp16")];
+            tensor<int32, [4]> var_1926 = const()[name = tensor<string, []>("op_1926"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> v_15_cast_fp16 = reshape(shape = var_1926, x = linear_68_cast_fp16)[name = tensor<string, []>("v_15_cast_fp16")];
+            tensor<int32, [4]> value_17_perm_0 = const()[name = tensor<string, []>("value_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96212096)))];
+            tensor<fp16, [1, 8, 8, 64]> var_1938_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1938_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96213184)))];
+            tensor<fp16, [1, 8, 8, 64]> var_1940_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1940_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_15_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_189_transpose_x_0 = const()[name = tensor<string, []>("x_189_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_189_transpose_y_0 = const()[name = tensor<string, []>("x_189_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 155]> var_1942_to_fp16 = const()[name = tensor<string, []>("op_1942_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96214272)))];
+            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_1940_cast_fp16)[name = tensor<string, []>("transpose_174")];
+            tensor<fp16, [1, 8, 8, 155]> x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = var_1942_to_fp16)[name = tensor<string, []>("x_189_cast_fp16")];
+            tensor<int32, [8]> x_191_pad_0 = const()[name = tensor<string, []>("x_191_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_191_mode_0 = const()[name = tensor<string, []>("x_191_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_170_to_fp16 = const()[name = tensor<string, []>("const_170_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 8, 156]> x_191_cast_fp16 = pad(constant_val = const_170_to_fp16, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189_cast_fp16)[name = tensor<string, []>("x_191_cast_fp16")];
+            tensor<int32, [4]> var_1950 = const()[name = tensor<string, []>("op_1950"), val = tensor<int32, [4]>([1, 8, -1, 8])];
+            tensor<fp16, [1, 8, 156, 8]> x_193_cast_fp16 = reshape(shape = var_1950, x = x_191_cast_fp16)[name = tensor<string, []>("x_193_cast_fp16")];
+            tensor<int32, [4]> var_1954_begin_0 = const()[name = tensor<string, []>("op_1954_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1954_end_0 = const()[name = tensor<string, []>("op_1954_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
+            tensor<bool, [4]> var_1954_end_mask_0 = const()[name = tensor<string, []>("op_1954_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 155, 8]> var_1954_cast_fp16 = slice_by_index(begin = var_1954_begin_0, end = var_1954_end_0, end_mask = var_1954_end_mask_0, x = x_193_cast_fp16)[name = tensor<string, []>("op_1954_cast_fp16")];
+            tensor<int32, [4]> var_1955 = const()[name = tensor<string, []>("op_1955"), val = tensor<int32, [4]>([1, 8, 8, 155])];
+            tensor<fp16, [1, 8, 8, 155]> matrix_bd_29_cast_fp16 = reshape(shape = var_1955, x = var_1954_cast_fp16)[name = tensor<string, []>("matrix_bd_29_cast_fp16")];
+            tensor<bool, []> matrix_ac_15_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_15_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_15_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_15_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_65_perm_0 = const()[name = tensor<string, []>("transpose_65_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_66_perm_0 = const()[name = tensor<string, []>("transpose_66_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 78]> transpose_66 = transpose(perm = transpose_66_perm_0, x = k_29_cast_fp16)[name = tensor<string, []>("transpose_172")];
+            tensor<fp16, [1, 8, 8, 64]> transpose_65 = transpose(perm = transpose_65_perm_0, x = var_1938_cast_fp16)[name = tensor<string, []>("transpose_173")];
+            tensor<fp16, [1, 8, 8, 78]> matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_65, y = transpose_66)[name = tensor<string, []>("matrix_ac_15_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_31_begin_0 = const()[name = tensor<string, []>("matrix_bd_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_31_end_0 = const()[name = tensor<string, []>("matrix_bd_31_end_0"), val = tensor<int32, [4]>([1, 8, 8, 78])];
+            tensor<bool, [4]> matrix_bd_31_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 8, 78]> matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29_cast_fp16)[name = tensor<string, []>("matrix_bd_31_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_1964_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = tensor<string, []>("op_1964_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_29_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_29_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 8, 78]> _inversed_scores_29_cast_fp16 = mul(x = var_1964_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_29_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> scores_31_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_31_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_1970_cast_fp16 = softmax(axis = var_56, x = scores_31_cast_fp16)[name = tensor<string, []>("op_1970_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> input_405_cast_fp16 = select(a = var_40_to_fp16, b = var_1970_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_405_cast_fp16")];
+            tensor<bool, []> x_195_transpose_x_0 = const()[name = tensor<string, []>("x_195_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_195_transpose_y_0 = const()[name = tensor<string, []>("x_195_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 78, 64]> value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_15_cast_fp16)[name = tensor<string, []>("transpose_175")];
+            tensor<fp16, [1, 8, 8, 64]> x_195_cast_fp16 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405_cast_fp16, y = value_17_cast_fp16)[name = tensor<string, []>("x_195_cast_fp16")];
+            tensor<int32, [4]> var_1974_perm_0 = const()[name = tensor<string, []>("op_1974_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1975 = const()[name = tensor<string, []>("op_1975"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 8, 8, 64]> var_1974_cast_fp16 = transpose(perm = var_1974_perm_0, x = x_195_cast_fp16)[name = tensor<string, []>("transpose_171")];
+            tensor<fp16, [1, 8, 512]> input_407_cast_fp16 = reshape(shape = var_1975, x = var_1974_cast_fp16)[name = tensor<string, []>("input_407_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_7_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96373056)))];
+            tensor<fp16, [1, 8, 512]> linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16, x = input_407_cast_fp16)[name = tensor<string, []>("linear_70_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = tensor<string, []>("input_411_cast_fp16")];
+            tensor<int32, [1]> x_199_axes_0 = const()[name = tensor<string, []>("x_199_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96897408)))];
+            tensor<fp16, [512]> encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96898496)))];
+            tensor<fp16, [1, 8, 512]> x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = tensor<string, []>("x_199_cast_fp16")];
+            tensor<int32, [3]> input_413_perm_0 = const()[name = tensor<string, []>("input_413_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_415_pad_type_0 = const()[name = tensor<string, []>("input_415_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_415_strides_0 = const()[name = tensor<string, []>("input_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_415_pad_0 = const()[name = tensor<string, []>("input_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_415_dilations_0 = const()[name = tensor<string, []>("input_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_415_groups_0 = const()[name = tensor<string, []>("input_415_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_7_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96899584)))];
+            tensor<fp16, [1, 512, 8]> input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = tensor<string, []>("transpose_170")];
+            tensor<fp16, [1, 1024, 8]> input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16, x = input_413_cast_fp16)[name = tensor<string, []>("input_415_cast_fp16")];
+            tensor<int32, []> x_201_split_num_splits_0 = const()[name = tensor<string, []>("x_201_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_201_split_axis_0 = const()[name = tensor<string, []>("x_201_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 8]> x_201_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> x_201_split_cast_fp16_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415_cast_fp16)[name = tensor<string, []>("x_201_split_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = tensor<string, []>("x_201_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_201_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> input_417_cast_fp16 = select(a = var_40_to_fp16, b = x_201_cast_fp16, cond = var_551)[name = tensor<string, []>("input_417_cast_fp16")];
+            tensor<bool, []> new_x_31_interleave_0 = const()[name = tensor<string, []>("new_x_31_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 16]> new_x_31_cast_fp16 = concat(axis = var_56, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = tensor<string, []>("new_x_31_cast_fp16")];
+            tensor<int32, [3]> next_cache_15_begin_0 = const()[name = tensor<string, []>("next_cache_15_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_15_end_0 = const()[name = tensor<string, []>("next_cache_15_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> next_cache_15_end_mask_0 = const()[name = tensor<string, []>("next_cache_15_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 12]> next_cache_15_cast_fp16 = slice_by_index(begin = next_cache_15_begin_0, end = next_cache_15_end_0, end_mask = next_cache_15_end_mask_0, x = new_x_31_cast_fp16)[name = tensor<string, []>("next_cache_15_cast_fp16")];
+            tensor<int32, [3]> var_2016_begin_0 = const()[name = tensor<string, []>("op_2016_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
+            tensor<int32, [3]> var_2016_end_0 = const()[name = tensor<string, []>("op_2016_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> var_2016_end_mask_0 = const()[name = tensor<string, []>("op_2016_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_2016_cast_fp16 = slice_by_index(begin = var_2016_begin_0, end = var_2016_end_0, end_mask = var_2016_end_mask_0, x = next_cache_15_cast_fp16)[name = tensor<string, []>("op_2016_cast_fp16")];
+            tensor<string, []> x_203_pad_type_0 = const()[name = tensor<string, []>("x_203_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_203_groups_0 = const()[name = tensor<string, []>("x_203_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_203_strides_0 = const()[name = tensor<string, []>("x_203_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_203_pad_0 = const()[name = tensor<string, []>("x_203_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_203_dilations_0 = const()[name = tensor<string, []>("x_203_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_7_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97948224)))];
+            tensor<fp16, [1, 512, 8]> x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16, x = new_x_31_cast_fp16)[name = tensor<string, []>("x_203_cast_fp16")];
+            tensor<int32, [3]> input_419_perm_0 = const()[name = tensor<string, []>("input_419_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_205_axes_0 = const()[name = tensor<string, []>("x_205_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97957504)))];
+            tensor<fp16, [512]> encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97958592)))];
+            tensor<fp16, [1, 8, 512]> input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = tensor<string, []>("transpose_169")];
+            tensor<fp16, [1, 8, 512]> x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = tensor<string, []>("x_205_cast_fp16")];
+            tensor<int32, [3]> input_421_perm_0 = const()[name = tensor<string, []>("input_421_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 8]> input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = tensor<string, []>("transpose_168")];
+            tensor<fp16, [1, 512, 8]> input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor<string, []>("input_423_cast_fp16")];
+            tensor<string, []> x_207_pad_type_0 = const()[name = tensor<string, []>("x_207_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_207_strides_0 = const()[name = tensor<string, []>("x_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_207_pad_0 = const()[name = tensor<string, []>("x_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_207_dilations_0 = const()[name = tensor<string, []>("x_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_207_groups_0 = const()[name = tensor<string, []>("x_207_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_7_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97959680)))];
+            tensor<fp16, [1, 512, 8]> x_207_cast_fp16 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16, x = input_423_cast_fp16)[name = tensor<string, []>("x_207_cast_fp16")];
+            tensor<int32, [3]> input_425_perm_0 = const()[name = tensor<string, []>("input_425_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 8, 512]> input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = tensor<string, []>("transpose_167")];
+            tensor<fp16, [1, 8, 512]> input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = tensor<string, []>("input_427_cast_fp16")];
+            tensor<int32, [1]> input_429_axes_0 = const()[name = tensor<string, []>("input_429_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98484032)))];
+            tensor<fp16, [512]> encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98485120)))];
+            tensor<fp16, [1, 8, 512]> input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = tensor<string, []>("input_429_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_7_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98486208)))];
+            tensor<fp16, [1, 8, 2048]> linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16, x = input_429_cast_fp16)[name = tensor<string, []>("linear_71_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = tensor<string, []>("input_433_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_7_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100583424)))];
+            tensor<fp16, [1, 8, 512]> linear_72_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16, x = input_433_cast_fp16)[name = tensor<string, []>("linear_72_cast_fp16")];
+            tensor<fp16, []> var_2057_to_fp16 = const()[name = tensor<string, []>("op_2057_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_2058_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_2057_to_fp16)[name = tensor<string, []>("op_2058_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_2058_cast_fp16)[name = tensor<string, []>("input_439_cast_fp16")];
+            tensor<int32, [1]> input_441_axes_0 = const()[name = tensor<string, []>("input_441_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_7_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102680640)))];
+            tensor<fp16, [512]> encoder_layers_7_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102681728)))];
+            tensor<fp16, [1, 8, 512]> input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = tensor<string, []>("input_441_cast_fp16")];
+            tensor<int32, [4]> cache_33_begin_0 = const()[name = tensor<string, []>("cache_33_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_33_end_0 = const()[name = tensor<string, []>("cache_33_end_0"), val = tensor<int32, [4]>([9, 1, 70, 512])];
+            tensor<bool, [4]> cache_33_end_mask_0 = const()[name = tensor<string, []>("cache_33_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_33_squeeze_mask_0 = const()[name = tensor<string, []>("cache_33_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_33_cast_fp16 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_33_cast_fp16")];
+            tensor<int32, [4]> cache_35_begin_0 = const()[name = tensor<string, []>("cache_35_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_35_end_0 = const()[name = tensor<string, []>("cache_35_end_0"), val = tensor<int32, [4]>([9, 1, 512, 8])];
+            tensor<bool, [4]> cache_35_end_mask_0 = const()[name = tensor<string, []>("cache_35_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_35_squeeze_mask_0 = const()[name = tensor<string, []>("cache_35_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_35_cast_fp16 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_35_cast_fp16")];
+            tensor<int32, [1]> input_443_axes_0 = const()[name = tensor<string, []>("input_443_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102682816)))];
+            tensor<fp16, [512]> encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102683904)))];
+            tensor<fp16, [1, 8, 512]> input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = tensor<string, []>("input_443_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_8_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102684992)))];
+            tensor<fp16, [1, 8, 2048]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16, x = input_443_cast_fp16)[name = tensor<string, []>("linear_73_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = tensor<string, []>("input_447_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_8_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104782208)))];
+            tensor<fp16, [1, 8, 512]> linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16, x = input_447_cast_fp16)[name = tensor<string, []>("linear_74_cast_fp16")];
+            tensor<fp16, []> var_2092_to_fp16 = const()[name = tensor<string, []>("op_2092_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_2093_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_2092_to_fp16)[name = tensor<string, []>("op_2093_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_2093_cast_fp16)[name = tensor<string, []>("input_453_cast_fp16")];
+            tensor<int32, [1]> key_17_axes_0 = const()[name = tensor<string, []>("key_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106879424)))];
+            tensor<fp16, [512]> encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106880512)))];
+            tensor<fp16, [1, 8, 512]> key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = tensor<string, []>("key_17_cast_fp16")];
+            tensor<bool, []> input_455_interleave_0 = const()[name = tensor<string, []>("input_455_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 78, 512]> input_455_cast_fp16 = concat(axis = var_65, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = tensor<string, []>("input_455_cast_fp16")];
+            tensor<int32, [3]> var_2115_begin_0 = const()[name = tensor<string, []>("op_2115_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
+            tensor<int32, [3]> var_2115_end_0 = const()[name = tensor<string, []>("op_2115_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_2115_end_mask_0 = const()[name = tensor<string, []>("op_2115_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 66, 512]> var_2115_cast_fp16 = slice_by_index(begin = var_2115_begin_0, end = var_2115_end_0, end_mask = var_2115_end_mask_0, x = cache_33_cast_fp16)[name = tensor<string, []>("op_2115_cast_fp16")];
+            tensor<int32, [3]> var_2118_begin_0 = const()[name = tensor<string, []>("op_2118_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_2118_end_0 = const()[name = tensor<string, []>("op_2118_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
+            tensor<bool, [3]> var_2118_end_mask_0 = const()[name = tensor<string, []>("op_2118_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 4, 512]> var_2118_cast_fp16 = slice_by_index(begin = var_2118_begin_0, end = var_2118_end_0, end_mask = var_2118_end_mask_0, x = key_17_cast_fp16)[name = tensor<string, []>("op_2118_cast_fp16")];
+            tensor<bool, []> var_2121_interleave_0 = const()[name = tensor<string, []>("op_2121_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_2121_cast_fp16 = concat(axis = var_65, interleave = var_2121_interleave_0, values = (var_2115_cast_fp16, var_2118_cast_fp16))[name = tensor<string, []>("op_2121_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_8_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106881600)))];
+            tensor<fp16, [1, 8, 512]> linear_75_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16, x = key_17_cast_fp16)[name = tensor<string, []>("linear_75_cast_fp16")];
+            tensor<int32, [4]> var_2125 = const()[name = tensor<string, []>("op_2125"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 8, 8, 64]> q_49_cast_fp16 = reshape(shape = var_2125, x = linear_75_cast_fp16)[name = tensor<string, []>("q_49_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_8_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107405952)))];
+            tensor<fp16, [1, 78, 512]> linear_76_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16, x = input_455_cast_fp16)[name = tensor<string, []>("linear_76_cast_fp16")];
+            tensor<int32, [4]> var_2129 = const()[name = tensor<string, []>("op_2129"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> k_33_cast_fp16 = reshape(shape = var_2129, x = linear_76_cast_fp16)[name = tensor<string, []>("k_33_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_8_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107930304)))];
+            tensor<fp16, [1, 78, 512]> linear_77_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16, x = input_455_cast_fp16)[name = tensor<string, []>("linear_77_cast_fp16")];
+            tensor<int32, [4]> var_2133 = const()[name = tensor<string, []>("op_2133"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> v_17_cast_fp16 = reshape(shape = var_2133, x = linear_77_cast_fp16)[name = tensor<string, []>("v_17_cast_fp16")];
+            tensor<int32, [4]> value_19_perm_0 = const()[name = tensor<string, []>("value_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108454656)))];
+            tensor<fp16, [1, 8, 8, 64]> var_2145_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2145_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108455744)))];
+            tensor<fp16, [1, 8, 8, 64]> var_2147_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2147_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_17_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_215_transpose_x_0 = const()[name = tensor<string, []>("x_215_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_215_transpose_y_0 = const()[name = tensor<string, []>("x_215_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 155]> var_2149_to_fp16 = const()[name = tensor<string, []>("op_2149_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108456832)))];
+            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2147_cast_fp16)[name = tensor<string, []>("transpose_165")];
+            tensor<fp16, [1, 8, 8, 155]> x_215_cast_fp16 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = var_2149_to_fp16)[name = tensor<string, []>("x_215_cast_fp16")];
+            tensor<int32, [8]> x_217_pad_0 = const()[name = tensor<string, []>("x_217_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_217_mode_0 = const()[name = tensor<string, []>("x_217_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_183_to_fp16 = const()[name = tensor<string, []>("const_183_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 8, 156]> x_217_cast_fp16 = pad(constant_val = const_183_to_fp16, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215_cast_fp16)[name = tensor<string, []>("x_217_cast_fp16")];
+            tensor<int32, [4]> var_2157 = const()[name = tensor<string, []>("op_2157"), val = tensor<int32, [4]>([1, 8, -1, 8])];
+            tensor<fp16, [1, 8, 156, 8]> x_219_cast_fp16 = reshape(shape = var_2157, x = x_217_cast_fp16)[name = tensor<string, []>("x_219_cast_fp16")];
+            tensor<int32, [4]> var_2161_begin_0 = const()[name = tensor<string, []>("op_2161_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2161_end_0 = const()[name = tensor<string, []>("op_2161_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
+            tensor<bool, [4]> var_2161_end_mask_0 = const()[name = tensor<string, []>("op_2161_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 155, 8]> var_2161_cast_fp16 = slice_by_index(begin = var_2161_begin_0, end = var_2161_end_0, end_mask = var_2161_end_mask_0, x = x_219_cast_fp16)[name = tensor<string, []>("op_2161_cast_fp16")];
+            tensor<int32, [4]> var_2162 = const()[name = tensor<string, []>("op_2162"), val = tensor<int32, [4]>([1, 8, 8, 155])];
+            tensor<fp16, [1, 8, 8, 155]> matrix_bd_33_cast_fp16 = reshape(shape = var_2162, x = var_2161_cast_fp16)[name = tensor<string, []>("matrix_bd_33_cast_fp16")];
+            tensor<bool, []> matrix_ac_17_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_17_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_17_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_17_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_67_perm_0 = const()[name = tensor<string, []>("transpose_67_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_68_perm_0 = const()[name = tensor<string, []>("transpose_68_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 78]> transpose_68 = transpose(perm = transpose_68_perm_0, x = k_33_cast_fp16)[name = tensor<string, []>("transpose_163")];
+            tensor<fp16, [1, 8, 8, 64]> transpose_67 = transpose(perm = transpose_67_perm_0, x = var_2145_cast_fp16)[name = tensor<string, []>("transpose_164")];
+            tensor<fp16, [1, 8, 8, 78]> matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_67, y = transpose_68)[name = tensor<string, []>("matrix_ac_17_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_35_begin_0 = const()[name = tensor<string, []>("matrix_bd_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_35_end_0 = const()[name = tensor<string, []>("matrix_bd_35_end_0"), val = tensor<int32, [4]>([1, 8, 8, 78])];
+            tensor<bool, [4]> matrix_bd_35_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 8, 78]> matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33_cast_fp16)[name = tensor<string, []>("matrix_bd_35_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_2171_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = tensor<string, []>("op_2171_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_33_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_33_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 8, 78]> _inversed_scores_33_cast_fp16 = mul(x = var_2171_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_33_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> scores_35_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_35_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_2177_cast_fp16 = softmax(axis = var_56, x = scores_35_cast_fp16)[name = tensor<string, []>("op_2177_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> input_457_cast_fp16 = select(a = var_40_to_fp16, b = var_2177_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_457_cast_fp16")];
+            tensor<bool, []> x_221_transpose_x_0 = const()[name = tensor<string, []>("x_221_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_221_transpose_y_0 = const()[name = tensor<string, []>("x_221_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 78, 64]> value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_17_cast_fp16)[name = tensor<string, []>("transpose_166")];
+            tensor<fp16, [1, 8, 8, 64]> x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457_cast_fp16, y = value_19_cast_fp16)[name = tensor<string, []>("x_221_cast_fp16")];
+            tensor<int32, [4]> var_2181_perm_0 = const()[name = tensor<string, []>("op_2181_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2182 = const()[name = tensor<string, []>("op_2182"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 8, 8, 64]> var_2181_cast_fp16 = transpose(perm = var_2181_perm_0, x = x_221_cast_fp16)[name = tensor<string, []>("transpose_162")];
+            tensor<fp16, [1, 8, 512]> input_459_cast_fp16 = reshape(shape = var_2182, x = var_2181_cast_fp16)[name = tensor<string, []>("input_459_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_8_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108615616)))];
+            tensor<fp16, [1, 8, 512]> linear_79_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16, x = input_459_cast_fp16)[name = tensor<string, []>("linear_79_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = tensor<string, []>("input_463_cast_fp16")];
+            tensor<int32, [1]> x_225_axes_0 = const()[name = tensor<string, []>("x_225_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109139968)))];
+            tensor<fp16, [512]> encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109141056)))];
+            tensor<fp16, [1, 8, 512]> x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = tensor<string, []>("x_225_cast_fp16")];
+            tensor<int32, [3]> input_465_perm_0 = const()[name = tensor<string, []>("input_465_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_467_pad_type_0 = const()[name = tensor<string, []>("input_467_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_467_strides_0 = const()[name = tensor<string, []>("input_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_467_pad_0 = const()[name = tensor<string, []>("input_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_467_dilations_0 = const()[name = tensor<string, []>("input_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_467_groups_0 = const()[name = tensor<string, []>("input_467_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_8_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109142144)))];
+            tensor<fp16, [1, 512, 8]> input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = tensor<string, []>("transpose_161")];
+            tensor<fp16, [1, 1024, 8]> input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16, x = input_465_cast_fp16)[name = tensor<string, []>("input_467_cast_fp16")];
+            tensor<int32, []> x_227_split_num_splits_0 = const()[name = tensor<string, []>("x_227_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_227_split_axis_0 = const()[name = tensor<string, []>("x_227_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 8]> x_227_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> x_227_split_cast_fp16_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467_cast_fp16)[name = tensor<string, []>("x_227_split_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = tensor<string, []>("x_227_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_227_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> input_469_cast_fp16 = select(a = var_40_to_fp16, b = x_227_cast_fp16, cond = var_551)[name = tensor<string, []>("input_469_cast_fp16")];
+            tensor<bool, []> new_x_35_interleave_0 = const()[name = tensor<string, []>("new_x_35_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 16]> new_x_35_cast_fp16 = concat(axis = var_56, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = tensor<string, []>("new_x_35_cast_fp16")];
+            tensor<int32, [3]> next_cache_17_begin_0 = const()[name = tensor<string, []>("next_cache_17_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_17_end_0 = const()[name = tensor<string, []>("next_cache_17_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> next_cache_17_end_mask_0 = const()[name = tensor<string, []>("next_cache_17_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 12]> next_cache_17_cast_fp16 = slice_by_index(begin = next_cache_17_begin_0, end = next_cache_17_end_0, end_mask = next_cache_17_end_mask_0, x = new_x_35_cast_fp16)[name = tensor<string, []>("next_cache_17_cast_fp16")];
+            tensor<int32, [3]> var_2223_begin_0 = const()[name = tensor<string, []>("op_2223_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
+            tensor<int32, [3]> var_2223_end_0 = const()[name = tensor<string, []>("op_2223_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> var_2223_end_mask_0 = const()[name = tensor<string, []>("op_2223_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_2223_cast_fp16 = slice_by_index(begin = var_2223_begin_0, end = var_2223_end_0, end_mask = var_2223_end_mask_0, x = next_cache_17_cast_fp16)[name = tensor<string, []>("op_2223_cast_fp16")];
+            tensor<string, []> x_229_pad_type_0 = const()[name = tensor<string, []>("x_229_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_229_groups_0 = const()[name = tensor<string, []>("x_229_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_229_strides_0 = const()[name = tensor<string, []>("x_229_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_229_pad_0 = const()[name = tensor<string, []>("x_229_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_229_dilations_0 = const()[name = tensor<string, []>("x_229_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_8_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110190784)))];
+            tensor<fp16, [1, 512, 8]> x_229_cast_fp16 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16, x = new_x_35_cast_fp16)[name = tensor<string, []>("x_229_cast_fp16")];
+            tensor<int32, [3]> input_471_perm_0 = const()[name = tensor<string, []>("input_471_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_231_axes_0 = const()[name = tensor<string, []>("x_231_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110200064)))];
+            tensor<fp16, [512]> encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110201152)))];
+            tensor<fp16, [1, 8, 512]> input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = tensor<string, []>("transpose_160")];
+            tensor<fp16, [1, 8, 512]> x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = tensor<string, []>("x_231_cast_fp16")];
+            tensor<int32, [3]> input_473_perm_0 = const()[name = tensor<string, []>("input_473_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 8]> input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = tensor<string, []>("transpose_159")];
+            tensor<fp16, [1, 512, 8]> input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = tensor<string, []>("input_475_cast_fp16")];
+            tensor<string, []> x_233_pad_type_0 = const()[name = tensor<string, []>("x_233_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_233_strides_0 = const()[name = tensor<string, []>("x_233_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_233_pad_0 = const()[name = tensor<string, []>("x_233_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_233_dilations_0 = const()[name = tensor<string, []>("x_233_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_233_groups_0 = const()[name = tensor<string, []>("x_233_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_8_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110202240)))];
+            tensor<fp16, [1, 512, 8]> x_233_cast_fp16 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16, x = input_475_cast_fp16)[name = tensor<string, []>("x_233_cast_fp16")];
+            tensor<int32, [3]> input_477_perm_0 = const()[name = tensor<string, []>("input_477_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 8, 512]> input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = tensor<string, []>("transpose_158")];
+            tensor<fp16, [1, 8, 512]> input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = tensor<string, []>("input_479_cast_fp16")];
+            tensor<int32, [1]> input_481_axes_0 = const()[name = tensor<string, []>("input_481_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110726592)))];
+            tensor<fp16, [512]> encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110727680)))];
+            tensor<fp16, [1, 8, 512]> input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = tensor<string, []>("input_481_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_8_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110728768)))];
+            tensor<fp16, [1, 8, 2048]> linear_80_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16, x = input_481_cast_fp16)[name = tensor<string, []>("linear_80_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = tensor<string, []>("input_485_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_8_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112825984)))];
+            tensor<fp16, [1, 8, 512]> linear_81_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16, x = input_485_cast_fp16)[name = tensor<string, []>("linear_81_cast_fp16")];
+            tensor<fp16, []> var_2264_to_fp16 = const()[name = tensor<string, []>("op_2264_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_2265_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2264_to_fp16)[name = tensor<string, []>("op_2265_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2265_cast_fp16)[name = tensor<string, []>("input_491_cast_fp16")];
+            tensor<int32, [1]> input_493_axes_0 = const()[name = tensor<string, []>("input_493_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_8_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114923200)))];
+            tensor<fp16, [512]> encoder_layers_8_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114924288)))];
+            tensor<fp16, [1, 8, 512]> input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = tensor<string, []>("input_493_cast_fp16")];
+            tensor<int32, [4]> cache_37_begin_0 = const()[name = tensor<string, []>("cache_37_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_37_end_0 = const()[name = tensor<string, []>("cache_37_end_0"), val = tensor<int32, [4]>([10, 1, 70, 512])];
+            tensor<bool, [4]> cache_37_end_mask_0 = const()[name = tensor<string, []>("cache_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_37_squeeze_mask_0 = const()[name = tensor<string, []>("cache_37_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_37_cast_fp16 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_37_cast_fp16")];
+            tensor<int32, [4]> cache_39_begin_0 = const()[name = tensor<string, []>("cache_39_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_39_end_0 = const()[name = tensor<string, []>("cache_39_end_0"), val = tensor<int32, [4]>([10, 1, 512, 8])];
+            tensor<bool, [4]> cache_39_end_mask_0 = const()[name = tensor<string, []>("cache_39_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_39_squeeze_mask_0 = const()[name = tensor<string, []>("cache_39_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_39_cast_fp16 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_39_cast_fp16")];
+            tensor<int32, [1]> input_495_axes_0 = const()[name = tensor<string, []>("input_495_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114925376)))];
+            tensor<fp16, [512]> encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114926464)))];
+            tensor<fp16, [1, 8, 512]> input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = tensor<string, []>("input_495_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_9_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114927552)))];
+            tensor<fp16, [1, 8, 2048]> linear_82_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16, x = input_495_cast_fp16)[name = tensor<string, []>("linear_82_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = tensor<string, []>("input_499_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_9_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117024768)))];
+            tensor<fp16, [1, 8, 512]> linear_83_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16, x = input_499_cast_fp16)[name = tensor<string, []>("linear_83_cast_fp16")];
+            tensor<fp16, []> var_2299_to_fp16 = const()[name = tensor<string, []>("op_2299_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_2300_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2299_to_fp16)[name = tensor<string, []>("op_2300_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2300_cast_fp16)[name = tensor<string, []>("input_505_cast_fp16")];
+            tensor<int32, [1]> key_19_axes_0 = const()[name = tensor<string, []>("key_19_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119121984)))];
+            tensor<fp16, [512]> encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119123072)))];
+            tensor<fp16, [1, 8, 512]> key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = tensor<string, []>("key_19_cast_fp16")];
+            tensor<bool, []> input_507_interleave_0 = const()[name = tensor<string, []>("input_507_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 78, 512]> input_507_cast_fp16 = concat(axis = var_65, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = tensor<string, []>("input_507_cast_fp16")];
+            tensor<int32, [3]> var_2322_begin_0 = const()[name = tensor<string, []>("op_2322_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
+            tensor<int32, [3]> var_2322_end_0 = const()[name = tensor<string, []>("op_2322_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_2322_end_mask_0 = const()[name = tensor<string, []>("op_2322_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 66, 512]> var_2322_cast_fp16 = slice_by_index(begin = var_2322_begin_0, end = var_2322_end_0, end_mask = var_2322_end_mask_0, x = cache_37_cast_fp16)[name = tensor<string, []>("op_2322_cast_fp16")];
+            tensor<int32, [3]> var_2325_begin_0 = const()[name = tensor<string, []>("op_2325_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_2325_end_0 = const()[name = tensor<string, []>("op_2325_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
+            tensor<bool, [3]> var_2325_end_mask_0 = const()[name = tensor<string, []>("op_2325_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 4, 512]> var_2325_cast_fp16 = slice_by_index(begin = var_2325_begin_0, end = var_2325_end_0, end_mask = var_2325_end_mask_0, x = key_19_cast_fp16)[name = tensor<string, []>("op_2325_cast_fp16")];
+            tensor<bool, []> var_2328_interleave_0 = const()[name = tensor<string, []>("op_2328_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_2328_cast_fp16 = concat(axis = var_65, interleave = var_2328_interleave_0, values = (var_2322_cast_fp16, var_2325_cast_fp16))[name = tensor<string, []>("op_2328_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_9_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119124160)))];
+            tensor<fp16, [1, 8, 512]> linear_84_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16, x = key_19_cast_fp16)[name = tensor<string, []>("linear_84_cast_fp16")];
+            tensor<int32, [4]> var_2332 = const()[name = tensor<string, []>("op_2332"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 8, 8, 64]> q_55_cast_fp16 = reshape(shape = var_2332, x = linear_84_cast_fp16)[name = tensor<string, []>("q_55_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_9_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119648512)))];
+            tensor<fp16, [1, 78, 512]> linear_85_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16, x = input_507_cast_fp16)[name = tensor<string, []>("linear_85_cast_fp16")];
+            tensor<int32, [4]> var_2336 = const()[name = tensor<string, []>("op_2336"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> k_37_cast_fp16 = reshape(shape = var_2336, x = linear_85_cast_fp16)[name = tensor<string, []>("k_37_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_9_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120172864)))];
+            tensor<fp16, [1, 78, 512]> linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16, x = input_507_cast_fp16)[name = tensor<string, []>("linear_86_cast_fp16")];
+            tensor<int32, [4]> var_2340 = const()[name = tensor<string, []>("op_2340"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> v_19_cast_fp16 = reshape(shape = var_2340, x = linear_86_cast_fp16)[name = tensor<string, []>("v_19_cast_fp16")];
+            tensor<int32, [4]> value_21_perm_0 = const()[name = tensor<string, []>("value_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120697216)))];
+            tensor<fp16, [1, 8, 8, 64]> var_2352_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2352_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120698304)))];
+            tensor<fp16, [1, 8, 8, 64]> var_2354_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2354_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_19_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_241_transpose_x_0 = const()[name = tensor<string, []>("x_241_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_241_transpose_y_0 = const()[name = tensor<string, []>("x_241_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 155]> var_2356_to_fp16 = const()[name = tensor<string, []>("op_2356_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120699392)))];
+            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2354_cast_fp16)[name = tensor<string, []>("transpose_156")];
+            tensor<fp16, [1, 8, 8, 155]> x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = var_2356_to_fp16)[name = tensor<string, []>("x_241_cast_fp16")];
+            tensor<int32, [8]> x_243_pad_0 = const()[name = tensor<string, []>("x_243_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_243_mode_0 = const()[name = tensor<string, []>("x_243_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_196_to_fp16 = const()[name = tensor<string, []>("const_196_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 8, 156]> x_243_cast_fp16 = pad(constant_val = const_196_to_fp16, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241_cast_fp16)[name = tensor<string, []>("x_243_cast_fp16")];
+            tensor<int32, [4]> var_2364 = const()[name = tensor<string, []>("op_2364"), val = tensor<int32, [4]>([1, 8, -1, 8])];
+            tensor<fp16, [1, 8, 156, 8]> x_245_cast_fp16 = reshape(shape = var_2364, x = x_243_cast_fp16)[name = tensor<string, []>("x_245_cast_fp16")];
+            tensor<int32, [4]> var_2368_begin_0 = const()[name = tensor<string, []>("op_2368_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2368_end_0 = const()[name = tensor<string, []>("op_2368_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
+            tensor<bool, [4]> var_2368_end_mask_0 = const()[name = tensor<string, []>("op_2368_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 155, 8]> var_2368_cast_fp16 = slice_by_index(begin = var_2368_begin_0, end = var_2368_end_0, end_mask = var_2368_end_mask_0, x = x_245_cast_fp16)[name = tensor<string, []>("op_2368_cast_fp16")];
+            tensor<int32, [4]> var_2369 = const()[name = tensor<string, []>("op_2369"), val = tensor<int32, [4]>([1, 8, 8, 155])];
+            tensor<fp16, [1, 8, 8, 155]> matrix_bd_37_cast_fp16 = reshape(shape = var_2369, x = var_2368_cast_fp16)[name = tensor<string, []>("matrix_bd_37_cast_fp16")];
+            tensor<bool, []> matrix_ac_19_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_19_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_19_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_19_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_69_perm_0 = const()[name = tensor<string, []>("transpose_69_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_70_perm_0 = const()[name = tensor<string, []>("transpose_70_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 78]> transpose_70 = transpose(perm = transpose_70_perm_0, x = k_37_cast_fp16)[name = tensor<string, []>("transpose_154")];
+            tensor<fp16, [1, 8, 8, 64]> transpose_69 = transpose(perm = transpose_69_perm_0, x = var_2352_cast_fp16)[name = tensor<string, []>("transpose_155")];
+            tensor<fp16, [1, 8, 8, 78]> matrix_ac_19_cast_fp16 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_69, y = transpose_70)[name = tensor<string, []>("matrix_ac_19_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_39_begin_0 = const()[name = tensor<string, []>("matrix_bd_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_39_end_0 = const()[name = tensor<string, []>("matrix_bd_39_end_0"), val = tensor<int32, [4]>([1, 8, 8, 78])];
+            tensor<bool, [4]> matrix_bd_39_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 8, 78]> matrix_bd_39_cast_fp16 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37_cast_fp16)[name = tensor<string, []>("matrix_bd_39_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_2378_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = tensor<string, []>("op_2378_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_37_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_37_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 8, 78]> _inversed_scores_37_cast_fp16 = mul(x = var_2378_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_37_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> scores_39_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_39_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_2384_cast_fp16 = softmax(axis = var_56, x = scores_39_cast_fp16)[name = tensor<string, []>("op_2384_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> input_509_cast_fp16 = select(a = var_40_to_fp16, b = var_2384_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_509_cast_fp16")];
+            tensor<bool, []> x_247_transpose_x_0 = const()[name = tensor<string, []>("x_247_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_247_transpose_y_0 = const()[name = tensor<string, []>("x_247_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 78, 64]> value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_19_cast_fp16)[name = tensor<string, []>("transpose_157")];
+            tensor<fp16, [1, 8, 8, 64]> x_247_cast_fp16 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509_cast_fp16, y = value_21_cast_fp16)[name = tensor<string, []>("x_247_cast_fp16")];
+            tensor<int32, [4]> var_2388_perm_0 = const()[name = tensor<string, []>("op_2388_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2389 = const()[name = tensor<string, []>("op_2389"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 8, 8, 64]> var_2388_cast_fp16 = transpose(perm = var_2388_perm_0, x = x_247_cast_fp16)[name = tensor<string, []>("transpose_153")];
+            tensor<fp16, [1, 8, 512]> input_511_cast_fp16 = reshape(shape = var_2389, x = var_2388_cast_fp16)[name = tensor<string, []>("input_511_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_9_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120858176)))];
+            tensor<fp16, [1, 8, 512]> linear_88_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16, x = input_511_cast_fp16)[name = tensor<string, []>("linear_88_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = tensor<string, []>("input_515_cast_fp16")];
+            tensor<int32, [1]> x_251_axes_0 = const()[name = tensor<string, []>("x_251_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121382528)))];
+            tensor<fp16, [512]> encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121383616)))];
+            tensor<fp16, [1, 8, 512]> x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = tensor<string, []>("x_251_cast_fp16")];
+            tensor<int32, [3]> input_517_perm_0 = const()[name = tensor<string, []>("input_517_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_519_pad_type_0 = const()[name = tensor<string, []>("input_519_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_519_strides_0 = const()[name = tensor<string, []>("input_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_519_pad_0 = const()[name = tensor<string, []>("input_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_519_dilations_0 = const()[name = tensor<string, []>("input_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_519_groups_0 = const()[name = tensor<string, []>("input_519_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_9_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121384704)))];
+            tensor<fp16, [1, 512, 8]> input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = tensor<string, []>("transpose_152")];
+            tensor<fp16, [1, 1024, 8]> input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16, x = input_517_cast_fp16)[name = tensor<string, []>("input_519_cast_fp16")];
+            tensor<int32, []> x_253_split_num_splits_0 = const()[name = tensor<string, []>("x_253_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_253_split_axis_0 = const()[name = tensor<string, []>("x_253_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 8]> x_253_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> x_253_split_cast_fp16_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519_cast_fp16)[name = tensor<string, []>("x_253_split_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = tensor<string, []>("x_253_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_253_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> input_521_cast_fp16 = select(a = var_40_to_fp16, b = x_253_cast_fp16, cond = var_551)[name = tensor<string, []>("input_521_cast_fp16")];
+            tensor<bool, []> new_x_39_interleave_0 = const()[name = tensor<string, []>("new_x_39_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 16]> new_x_39_cast_fp16 = concat(axis = var_56, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = tensor<string, []>("new_x_39_cast_fp16")];
+            tensor<int32, [3]> next_cache_19_begin_0 = const()[name = tensor<string, []>("next_cache_19_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_19_end_0 = const()[name = tensor<string, []>("next_cache_19_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> next_cache_19_end_mask_0 = const()[name = tensor<string, []>("next_cache_19_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 12]> next_cache_19_cast_fp16 = slice_by_index(begin = next_cache_19_begin_0, end = next_cache_19_end_0, end_mask = next_cache_19_end_mask_0, x = new_x_39_cast_fp16)[name = tensor<string, []>("next_cache_19_cast_fp16")];
+            tensor<int32, [3]> var_2430_begin_0 = const()[name = tensor<string, []>("op_2430_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
+            tensor<int32, [3]> var_2430_end_0 = const()[name = tensor<string, []>("op_2430_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> var_2430_end_mask_0 = const()[name = tensor<string, []>("op_2430_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_2430_cast_fp16 = slice_by_index(begin = var_2430_begin_0, end = var_2430_end_0, end_mask = var_2430_end_mask_0, x = next_cache_19_cast_fp16)[name = tensor<string, []>("op_2430_cast_fp16")];
+            tensor<string, []> x_255_pad_type_0 = const()[name = tensor<string, []>("x_255_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_255_groups_0 = const()[name = tensor<string, []>("x_255_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_255_strides_0 = const()[name = tensor<string, []>("x_255_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_255_pad_0 = const()[name = tensor<string, []>("x_255_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_255_dilations_0 = const()[name = tensor<string, []>("x_255_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_9_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122433344)))];
+            tensor<fp16, [1, 512, 8]> x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16, x = new_x_39_cast_fp16)[name = tensor<string, []>("x_255_cast_fp16")];
+            tensor<int32, [3]> input_523_perm_0 = const()[name = tensor<string, []>("input_523_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_257_axes_0 = const()[name = tensor<string, []>("x_257_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122442624)))];
+            tensor<fp16, [512]> encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122443712)))];
+            tensor<fp16, [1, 8, 512]> input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = tensor<string, []>("transpose_151")];
+            tensor<fp16, [1, 8, 512]> x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = tensor<string, []>("x_257_cast_fp16")];
+            tensor<int32, [3]> input_525_perm_0 = const()[name = tensor<string, []>("input_525_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 8]> input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = tensor<string, []>("transpose_150")];
+            tensor<fp16, [1, 512, 8]> input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = tensor<string, []>("input_527_cast_fp16")];
+            tensor<string, []> x_259_pad_type_0 = const()[name = tensor<string, []>("x_259_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_259_strides_0 = const()[name = tensor<string, []>("x_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_259_pad_0 = const()[name = tensor<string, []>("x_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_259_dilations_0 = const()[name = tensor<string, []>("x_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_259_groups_0 = const()[name = tensor<string, []>("x_259_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_9_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122444800)))];
+            tensor<fp16, [1, 512, 8]> x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16, x = input_527_cast_fp16)[name = tensor<string, []>("x_259_cast_fp16")];
+            tensor<int32, [3]> input_529_perm_0 = const()[name = tensor<string, []>("input_529_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 8, 512]> input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = tensor<string, []>("transpose_149")];
+            tensor<fp16, [1, 8, 512]> input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = tensor<string, []>("input_531_cast_fp16")];
+            tensor<int32, [1]> input_533_axes_0 = const()[name = tensor<string, []>("input_533_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122969152)))];
+            tensor<fp16, [512]> encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122970240)))];
+            tensor<fp16, [1, 8, 512]> input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = tensor<string, []>("input_533_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_9_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122971328)))];
+            tensor<fp16, [1, 8, 2048]> linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16, x = input_533_cast_fp16)[name = tensor<string, []>("linear_89_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = tensor<string, []>("input_537_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_9_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125068544)))];
+            tensor<fp16, [1, 8, 512]> linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16, x = input_537_cast_fp16)[name = tensor<string, []>("linear_90_cast_fp16")];
+            tensor<fp16, []> var_2471_to_fp16 = const()[name = tensor<string, []>("op_2471_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_2472_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2471_to_fp16)[name = tensor<string, []>("op_2472_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2472_cast_fp16)[name = tensor<string, []>("input_543_cast_fp16")];
+            tensor<int32, [1]> input_545_axes_0 = const()[name = tensor<string, []>("input_545_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_9_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127165760)))];
+            tensor<fp16, [512]> encoder_layers_9_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127166848)))];
+            tensor<fp16, [1, 8, 512]> input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = tensor<string, []>("input_545_cast_fp16")];
+            tensor<int32, [4]> cache_41_begin_0 = const()[name = tensor<string, []>("cache_41_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_41_end_0 = const()[name = tensor<string, []>("cache_41_end_0"), val = tensor<int32, [4]>([11, 1, 70, 512])];
+            tensor<bool, [4]> cache_41_end_mask_0 = const()[name = tensor<string, []>("cache_41_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_41_squeeze_mask_0 = const()[name = tensor<string, []>("cache_41_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_41_cast_fp16 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_41_cast_fp16")];
+            tensor<int32, [4]> cache_43_begin_0 = const()[name = tensor<string, []>("cache_43_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_43_end_0 = const()[name = tensor<string, []>("cache_43_end_0"), val = tensor<int32, [4]>([11, 1, 512, 8])];
+            tensor<bool, [4]> cache_43_end_mask_0 = const()[name = tensor<string, []>("cache_43_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_43_squeeze_mask_0 = const()[name = tensor<string, []>("cache_43_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_43_cast_fp16 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_43_cast_fp16")];
+            tensor<int32, [1]> input_547_axes_0 = const()[name = tensor<string, []>("input_547_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127167936)))];
+            tensor<fp16, [512]> encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127169024)))];
+            tensor<fp16, [1, 8, 512]> input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = tensor<string, []>("input_547_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_10_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127170112)))];
+            tensor<fp16, [1, 8, 2048]> linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16, x = input_547_cast_fp16)[name = tensor<string, []>("linear_91_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = tensor<string, []>("input_551_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_10_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129267328)))];
+            tensor<fp16, [1, 8, 512]> linear_92_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16, x = input_551_cast_fp16)[name = tensor<string, []>("linear_92_cast_fp16")];
+            tensor<fp16, []> var_2506_to_fp16 = const()[name = tensor<string, []>("op_2506_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_2507_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2506_to_fp16)[name = tensor<string, []>("op_2507_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2507_cast_fp16)[name = tensor<string, []>("input_557_cast_fp16")];
+            tensor<int32, [1]> key_21_axes_0 = const()[name = tensor<string, []>("key_21_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131364544)))];
+            tensor<fp16, [512]> encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131365632)))];
+            tensor<fp16, [1, 8, 512]> key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = tensor<string, []>("key_21_cast_fp16")];
+            tensor<bool, []> input_559_interleave_0 = const()[name = tensor<string, []>("input_559_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 78, 512]> input_559_cast_fp16 = concat(axis = var_65, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = tensor<string, []>("input_559_cast_fp16")];
+            tensor<int32, [3]> var_2529_begin_0 = const()[name = tensor<string, []>("op_2529_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
+            tensor<int32, [3]> var_2529_end_0 = const()[name = tensor<string, []>("op_2529_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_2529_end_mask_0 = const()[name = tensor<string, []>("op_2529_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 66, 512]> var_2529_cast_fp16 = slice_by_index(begin = var_2529_begin_0, end = var_2529_end_0, end_mask = var_2529_end_mask_0, x = cache_41_cast_fp16)[name = tensor<string, []>("op_2529_cast_fp16")];
+            tensor<int32, [3]> var_2532_begin_0 = const()[name = tensor<string, []>("op_2532_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_2532_end_0 = const()[name = tensor<string, []>("op_2532_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
+            tensor<bool, [3]> var_2532_end_mask_0 = const()[name = tensor<string, []>("op_2532_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 4, 512]> var_2532_cast_fp16 = slice_by_index(begin = var_2532_begin_0, end = var_2532_end_0, end_mask = var_2532_end_mask_0, x = key_21_cast_fp16)[name = tensor<string, []>("op_2532_cast_fp16")];
+            tensor<bool, []> var_2535_interleave_0 = const()[name = tensor<string, []>("op_2535_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_2535_cast_fp16 = concat(axis = var_65, interleave = var_2535_interleave_0, values = (var_2529_cast_fp16, var_2532_cast_fp16))[name = tensor<string, []>("op_2535_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_10_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131366720)))];
+            tensor<fp16, [1, 8, 512]> linear_93_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16, x = key_21_cast_fp16)[name = tensor<string, []>("linear_93_cast_fp16")];
+            tensor<int32, [4]> var_2539 = const()[name = tensor<string, []>("op_2539"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 8, 8, 64]> q_61_cast_fp16 = reshape(shape = var_2539, x = linear_93_cast_fp16)[name = tensor<string, []>("q_61_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_10_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131891072)))];
+            tensor<fp16, [1, 78, 512]> linear_94_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16, x = input_559_cast_fp16)[name = tensor<string, []>("linear_94_cast_fp16")];
+            tensor<int32, [4]> var_2543 = const()[name = tensor<string, []>("op_2543"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> k_41_cast_fp16 = reshape(shape = var_2543, x = linear_94_cast_fp16)[name = tensor<string, []>("k_41_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_10_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132415424)))];
+            tensor<fp16, [1, 78, 512]> linear_95_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16, x = input_559_cast_fp16)[name = tensor<string, []>("linear_95_cast_fp16")];
+            tensor<int32, [4]> var_2547 = const()[name = tensor<string, []>("op_2547"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> v_21_cast_fp16 = reshape(shape = var_2547, x = linear_95_cast_fp16)[name = tensor<string, []>("v_21_cast_fp16")];
+            tensor<int32, [4]> value_23_perm_0 = const()[name = tensor<string, []>("value_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132939776)))];
+            tensor<fp16, [1, 8, 8, 64]> var_2559_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2559_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132940864)))];
+            tensor<fp16, [1, 8, 8, 64]> var_2561_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2561_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_21_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_267_transpose_x_0 = const()[name = tensor<string, []>("x_267_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_267_transpose_y_0 = const()[name = tensor<string, []>("x_267_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 155]> var_2563_to_fp16 = const()[name = tensor<string, []>("op_2563_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132941952)))];
+            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2561_cast_fp16)[name = tensor<string, []>("transpose_147")];
+            tensor<fp16, [1, 8, 8, 155]> x_267_cast_fp16 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = var_2563_to_fp16)[name = tensor<string, []>("x_267_cast_fp16")];
+            tensor<int32, [8]> x_269_pad_0 = const()[name = tensor<string, []>("x_269_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_269_mode_0 = const()[name = tensor<string, []>("x_269_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_209_to_fp16 = const()[name = tensor<string, []>("const_209_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 8, 156]> x_269_cast_fp16 = pad(constant_val = const_209_to_fp16, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267_cast_fp16)[name = tensor<string, []>("x_269_cast_fp16")];
+            tensor<int32, [4]> var_2571 = const()[name = tensor<string, []>("op_2571"), val = tensor<int32, [4]>([1, 8, -1, 8])];
+            tensor<fp16, [1, 8, 156, 8]> x_271_cast_fp16 = reshape(shape = var_2571, x = x_269_cast_fp16)[name = tensor<string, []>("x_271_cast_fp16")];
+            tensor<int32, [4]> var_2575_begin_0 = const()[name = tensor<string, []>("op_2575_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2575_end_0 = const()[name = tensor<string, []>("op_2575_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
+            tensor<bool, [4]> var_2575_end_mask_0 = const()[name = tensor<string, []>("op_2575_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 155, 8]> var_2575_cast_fp16 = slice_by_index(begin = var_2575_begin_0, end = var_2575_end_0, end_mask = var_2575_end_mask_0, x = x_271_cast_fp16)[name = tensor<string, []>("op_2575_cast_fp16")];
+            tensor<int32, [4]> var_2576 = const()[name = tensor<string, []>("op_2576"), val = tensor<int32, [4]>([1, 8, 8, 155])];
+            tensor<fp16, [1, 8, 8, 155]> matrix_bd_41_cast_fp16 = reshape(shape = var_2576, x = var_2575_cast_fp16)[name = tensor<string, []>("matrix_bd_41_cast_fp16")];
+            tensor<bool, []> matrix_ac_21_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_21_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_21_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_21_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_71_perm_0 = const()[name = tensor<string, []>("transpose_71_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_72_perm_0 = const()[name = tensor<string, []>("transpose_72_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 78]> transpose_72 = transpose(perm = transpose_72_perm_0, x = k_41_cast_fp16)[name = tensor<string, []>("transpose_145")];
+            tensor<fp16, [1, 8, 8, 64]> transpose_71 = transpose(perm = transpose_71_perm_0, x = var_2559_cast_fp16)[name = tensor<string, []>("transpose_146")];
+            tensor<fp16, [1, 8, 8, 78]> matrix_ac_21_cast_fp16 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_71, y = transpose_72)[name = tensor<string, []>("matrix_ac_21_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_43_begin_0 = const()[name = tensor<string, []>("matrix_bd_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_43_end_0 = const()[name = tensor<string, []>("matrix_bd_43_end_0"), val = tensor<int32, [4]>([1, 8, 8, 78])];
+            tensor<bool, [4]> matrix_bd_43_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 8, 78]> matrix_bd_43_cast_fp16 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41_cast_fp16)[name = tensor<string, []>("matrix_bd_43_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_2585_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = tensor<string, []>("op_2585_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_41_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_41_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 8, 78]> _inversed_scores_41_cast_fp16 = mul(x = var_2585_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_41_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> scores_43_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_43_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_2591_cast_fp16 = softmax(axis = var_56, x = scores_43_cast_fp16)[name = tensor<string, []>("op_2591_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> input_561_cast_fp16 = select(a = var_40_to_fp16, b = var_2591_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_561_cast_fp16")];
+            tensor<bool, []> x_273_transpose_x_0 = const()[name = tensor<string, []>("x_273_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_273_transpose_y_0 = const()[name = tensor<string, []>("x_273_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 78, 64]> value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_21_cast_fp16)[name = tensor<string, []>("transpose_148")];
+            tensor<fp16, [1, 8, 8, 64]> x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561_cast_fp16, y = value_23_cast_fp16)[name = tensor<string, []>("x_273_cast_fp16")];
+            tensor<int32, [4]> var_2595_perm_0 = const()[name = tensor<string, []>("op_2595_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2596 = const()[name = tensor<string, []>("op_2596"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 8, 8, 64]> var_2595_cast_fp16 = transpose(perm = var_2595_perm_0, x = x_273_cast_fp16)[name = tensor<string, []>("transpose_144")];
+            tensor<fp16, [1, 8, 512]> input_563_cast_fp16 = reshape(shape = var_2596, x = var_2595_cast_fp16)[name = tensor<string, []>("input_563_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_10_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133100736)))];
+            tensor<fp16, [1, 8, 512]> linear_97_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16, x = input_563_cast_fp16)[name = tensor<string, []>("linear_97_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = tensor<string, []>("input_567_cast_fp16")];
+            tensor<int32, [1]> x_277_axes_0 = const()[name = tensor<string, []>("x_277_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133625088)))];
+            tensor<fp16, [512]> encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133626176)))];
+            tensor<fp16, [1, 8, 512]> x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = tensor<string, []>("x_277_cast_fp16")];
+            tensor<int32, [3]> input_569_perm_0 = const()[name = tensor<string, []>("input_569_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_571_pad_type_0 = const()[name = tensor<string, []>("input_571_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_571_strides_0 = const()[name = tensor<string, []>("input_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_571_pad_0 = const()[name = tensor<string, []>("input_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_571_dilations_0 = const()[name = tensor<string, []>("input_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_571_groups_0 = const()[name = tensor<string, []>("input_571_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_10_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133627264)))];
+            tensor<fp16, [1, 512, 8]> input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = tensor<string, []>("transpose_143")];
+            tensor<fp16, [1, 1024, 8]> input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16, x = input_569_cast_fp16)[name = tensor<string, []>("input_571_cast_fp16")];
+            tensor<int32, []> x_279_split_num_splits_0 = const()[name = tensor<string, []>("x_279_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_279_split_axis_0 = const()[name = tensor<string, []>("x_279_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 8]> x_279_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> x_279_split_cast_fp16_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571_cast_fp16)[name = tensor<string, []>("x_279_split_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = tensor<string, []>("x_279_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_279_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> input_573_cast_fp16 = select(a = var_40_to_fp16, b = x_279_cast_fp16, cond = var_551)[name = tensor<string, []>("input_573_cast_fp16")];
+            tensor<bool, []> new_x_43_interleave_0 = const()[name = tensor<string, []>("new_x_43_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 16]> new_x_43_cast_fp16 = concat(axis = var_56, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = tensor<string, []>("new_x_43_cast_fp16")];
+            tensor<int32, [3]> next_cache_21_begin_0 = const()[name = tensor<string, []>("next_cache_21_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_21_end_0 = const()[name = tensor<string, []>("next_cache_21_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> next_cache_21_end_mask_0 = const()[name = tensor<string, []>("next_cache_21_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 12]> next_cache_21_cast_fp16 = slice_by_index(begin = next_cache_21_begin_0, end = next_cache_21_end_0, end_mask = next_cache_21_end_mask_0, x = new_x_43_cast_fp16)[name = tensor<string, []>("next_cache_21_cast_fp16")];
+            tensor<int32, [3]> var_2637_begin_0 = const()[name = tensor<string, []>("op_2637_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
+            tensor<int32, [3]> var_2637_end_0 = const()[name = tensor<string, []>("op_2637_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> var_2637_end_mask_0 = const()[name = tensor<string, []>("op_2637_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_2637_cast_fp16 = slice_by_index(begin = var_2637_begin_0, end = var_2637_end_0, end_mask = var_2637_end_mask_0, x = next_cache_21_cast_fp16)[name = tensor<string, []>("op_2637_cast_fp16")];
+            tensor<string, []> x_281_pad_type_0 = const()[name = tensor<string, []>("x_281_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_281_groups_0 = const()[name = tensor<string, []>("x_281_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_281_strides_0 = const()[name = tensor<string, []>("x_281_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_281_pad_0 = const()[name = tensor<string, []>("x_281_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_281_dilations_0 = const()[name = tensor<string, []>("x_281_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_10_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134675904)))];
+            tensor<fp16, [1, 512, 8]> x_281_cast_fp16 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16, x = new_x_43_cast_fp16)[name = tensor<string, []>("x_281_cast_fp16")];
+            tensor<int32, [3]> input_575_perm_0 = const()[name = tensor<string, []>("input_575_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_283_axes_0 = const()[name = tensor<string, []>("x_283_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134685184)))];
+            tensor<fp16, [512]> encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134686272)))];
+            tensor<fp16, [1, 8, 512]> input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = tensor<string, []>("transpose_142")];
+            tensor<fp16, [1, 8, 512]> x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = tensor<string, []>("x_283_cast_fp16")];
+            tensor<int32, [3]> input_577_perm_0 = const()[name = tensor<string, []>("input_577_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 8]> input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = tensor<string, []>("transpose_141")];
+            tensor<fp16, [1, 512, 8]> input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = tensor<string, []>("input_579_cast_fp16")];
+            tensor<string, []> x_285_pad_type_0 = const()[name = tensor<string, []>("x_285_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_285_strides_0 = const()[name = tensor<string, []>("x_285_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_285_pad_0 = const()[name = tensor<string, []>("x_285_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_285_dilations_0 = const()[name = tensor<string, []>("x_285_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_285_groups_0 = const()[name = tensor<string, []>("x_285_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_10_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134687360)))];
+            tensor<fp16, [1, 512, 8]> x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16, x = input_579_cast_fp16)[name = tensor<string, []>("x_285_cast_fp16")];
+            tensor<int32, [3]> input_581_perm_0 = const()[name = tensor<string, []>("input_581_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 8, 512]> input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = tensor<string, []>("transpose_140")];
+            tensor<fp16, [1, 8, 512]> input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = tensor<string, []>("input_583_cast_fp16")];
+            tensor<int32, [1]> input_585_axes_0 = const()[name = tensor<string, []>("input_585_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135211712)))];
+            tensor<fp16, [512]> encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135212800)))];
+            tensor<fp16, [1, 8, 512]> input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = tensor<string, []>("input_585_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_10_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135213888)))];
+            tensor<fp16, [1, 8, 2048]> linear_98_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16, x = input_585_cast_fp16)[name = tensor<string, []>("linear_98_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = tensor<string, []>("input_589_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_10_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137311104)))];
+            tensor<fp16, [1, 8, 512]> linear_99_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16, x = input_589_cast_fp16)[name = tensor<string, []>("linear_99_cast_fp16")];
+            tensor<fp16, []> var_2678_to_fp16 = const()[name = tensor<string, []>("op_2678_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_2679_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2678_to_fp16)[name = tensor<string, []>("op_2679_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2679_cast_fp16)[name = tensor<string, []>("input_595_cast_fp16")];
+            tensor<int32, [1]> input_597_axes_0 = const()[name = tensor<string, []>("input_597_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_10_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139408320)))];
+            tensor<fp16, [512]> encoder_layers_10_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139409408)))];
+            tensor<fp16, [1, 8, 512]> input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = tensor<string, []>("input_597_cast_fp16")];
+            tensor<int32, [4]> cache_45_begin_0 = const()[name = tensor<string, []>("cache_45_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_45_end_0 = const()[name = tensor<string, []>("cache_45_end_0"), val = tensor<int32, [4]>([12, 1, 70, 512])];
+            tensor<bool, [4]> cache_45_end_mask_0 = const()[name = tensor<string, []>("cache_45_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_45_squeeze_mask_0 = const()[name = tensor<string, []>("cache_45_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_45_cast_fp16 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_45_cast_fp16")];
+            tensor<int32, [4]> cache_47_begin_0 = const()[name = tensor<string, []>("cache_47_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_47_end_0 = const()[name = tensor<string, []>("cache_47_end_0"), val = tensor<int32, [4]>([12, 1, 512, 8])];
+            tensor<bool, [4]> cache_47_end_mask_0 = const()[name = tensor<string, []>("cache_47_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_47_squeeze_mask_0 = const()[name = tensor<string, []>("cache_47_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_47_cast_fp16 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_47_cast_fp16")];
+            tensor<int32, [1]> input_599_axes_0 = const()[name = tensor<string, []>("input_599_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139410496)))];
+            tensor<fp16, [512]> encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139411584)))];
+            tensor<fp16, [1, 8, 512]> input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = tensor<string, []>("input_599_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_11_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139412672)))];
+            tensor<fp16, [1, 8, 2048]> linear_100_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16, x = input_599_cast_fp16)[name = tensor<string, []>("linear_100_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = tensor<string, []>("input_603_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_11_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141509888)))];
+            tensor<fp16, [1, 8, 512]> linear_101_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16, x = input_603_cast_fp16)[name = tensor<string, []>("linear_101_cast_fp16")];
+            tensor<fp16, []> var_2713_to_fp16 = const()[name = tensor<string, []>("op_2713_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_2714_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2713_to_fp16)[name = tensor<string, []>("op_2714_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2714_cast_fp16)[name = tensor<string, []>("input_609_cast_fp16")];
+            tensor<int32, [1]> key_23_axes_0 = const()[name = tensor<string, []>("key_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143607104)))];
+            tensor<fp16, [512]> encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143608192)))];
+            tensor<fp16, [1, 8, 512]> key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = tensor<string, []>("key_23_cast_fp16")];
+            tensor<bool, []> input_611_interleave_0 = const()[name = tensor<string, []>("input_611_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 78, 512]> input_611_cast_fp16 = concat(axis = var_65, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = tensor<string, []>("input_611_cast_fp16")];
+            tensor<int32, [3]> var_2736_begin_0 = const()[name = tensor<string, []>("op_2736_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
+            tensor<int32, [3]> var_2736_end_0 = const()[name = tensor<string, []>("op_2736_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_2736_end_mask_0 = const()[name = tensor<string, []>("op_2736_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 66, 512]> var_2736_cast_fp16 = slice_by_index(begin = var_2736_begin_0, end = var_2736_end_0, end_mask = var_2736_end_mask_0, x = cache_45_cast_fp16)[name = tensor<string, []>("op_2736_cast_fp16")];
+            tensor<int32, [3]> var_2739_begin_0 = const()[name = tensor<string, []>("op_2739_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_2739_end_0 = const()[name = tensor<string, []>("op_2739_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
+            tensor<bool, [3]> var_2739_end_mask_0 = const()[name = tensor<string, []>("op_2739_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 4, 512]> var_2739_cast_fp16 = slice_by_index(begin = var_2739_begin_0, end = var_2739_end_0, end_mask = var_2739_end_mask_0, x = key_23_cast_fp16)[name = tensor<string, []>("op_2739_cast_fp16")];
+            tensor<bool, []> var_2742_interleave_0 = const()[name = tensor<string, []>("op_2742_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_2742_cast_fp16 = concat(axis = var_65, interleave = var_2742_interleave_0, values = (var_2736_cast_fp16, var_2739_cast_fp16))[name = tensor<string, []>("op_2742_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_11_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143609280)))];
+            tensor<fp16, [1, 8, 512]> linear_102_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16, x = key_23_cast_fp16)[name = tensor<string, []>("linear_102_cast_fp16")];
+            tensor<int32, [4]> var_2746 = const()[name = tensor<string, []>("op_2746"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 8, 8, 64]> q_67_cast_fp16 = reshape(shape = var_2746, x = linear_102_cast_fp16)[name = tensor<string, []>("q_67_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_11_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144133632)))];
+            tensor<fp16, [1, 78, 512]> linear_103_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16, x = input_611_cast_fp16)[name = tensor<string, []>("linear_103_cast_fp16")];
+            tensor<int32, [4]> var_2750 = const()[name = tensor<string, []>("op_2750"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> k_45_cast_fp16 = reshape(shape = var_2750, x = linear_103_cast_fp16)[name = tensor<string, []>("k_45_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_11_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144657984)))];
+            tensor<fp16, [1, 78, 512]> linear_104_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16, x = input_611_cast_fp16)[name = tensor<string, []>("linear_104_cast_fp16")];
+            tensor<int32, [4]> var_2754 = const()[name = tensor<string, []>("op_2754"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> v_23_cast_fp16 = reshape(shape = var_2754, x = linear_104_cast_fp16)[name = tensor<string, []>("v_23_cast_fp16")];
+            tensor<int32, [4]> value_25_perm_0 = const()[name = tensor<string, []>("value_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145182336)))];
+            tensor<fp16, [1, 8, 8, 64]> var_2766_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2766_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145183424)))];
+            tensor<fp16, [1, 8, 8, 64]> var_2768_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2768_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_23_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_293_transpose_x_0 = const()[name = tensor<string, []>("x_293_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_293_transpose_y_0 = const()[name = tensor<string, []>("x_293_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 155]> var_2770_to_fp16 = const()[name = tensor<string, []>("op_2770_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145184512)))];
+            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2768_cast_fp16)[name = tensor<string, []>("transpose_138")];
+            tensor<fp16, [1, 8, 8, 155]> x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = var_2770_to_fp16)[name = tensor<string, []>("x_293_cast_fp16")];
+            tensor<int32, [8]> x_295_pad_0 = const()[name = tensor<string, []>("x_295_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_295_mode_0 = const()[name = tensor<string, []>("x_295_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_222_to_fp16 = const()[name = tensor<string, []>("const_222_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 8, 156]> x_295_cast_fp16 = pad(constant_val = const_222_to_fp16, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293_cast_fp16)[name = tensor<string, []>("x_295_cast_fp16")];
+            tensor<int32, [4]> var_2778 = const()[name = tensor<string, []>("op_2778"), val = tensor<int32, [4]>([1, 8, -1, 8])];
+            tensor<fp16, [1, 8, 156, 8]> x_297_cast_fp16 = reshape(shape = var_2778, x = x_295_cast_fp16)[name = tensor<string, []>("x_297_cast_fp16")];
+            tensor<int32, [4]> var_2782_begin_0 = const()[name = tensor<string, []>("op_2782_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2782_end_0 = const()[name = tensor<string, []>("op_2782_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
+            tensor<bool, [4]> var_2782_end_mask_0 = const()[name = tensor<string, []>("op_2782_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 155, 8]> var_2782_cast_fp16 = slice_by_index(begin = var_2782_begin_0, end = var_2782_end_0, end_mask = var_2782_end_mask_0, x = x_297_cast_fp16)[name = tensor<string, []>("op_2782_cast_fp16")];
+            tensor<int32, [4]> var_2783 = const()[name = tensor<string, []>("op_2783"), val = tensor<int32, [4]>([1, 8, 8, 155])];
+            tensor<fp16, [1, 8, 8, 155]> matrix_bd_45_cast_fp16 = reshape(shape = var_2783, x = var_2782_cast_fp16)[name = tensor<string, []>("matrix_bd_45_cast_fp16")];
+            tensor<bool, []> matrix_ac_23_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_23_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_23_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_23_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_73_perm_0 = const()[name = tensor<string, []>("transpose_73_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_74_perm_0 = const()[name = tensor<string, []>("transpose_74_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 78]> transpose_74 = transpose(perm = transpose_74_perm_0, x = k_45_cast_fp16)[name = tensor<string, []>("transpose_136")];
+            tensor<fp16, [1, 8, 8, 64]> transpose_73 = transpose(perm = transpose_73_perm_0, x = var_2766_cast_fp16)[name = tensor<string, []>("transpose_137")];
+            tensor<fp16, [1, 8, 8, 78]> matrix_ac_23_cast_fp16 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_73, y = transpose_74)[name = tensor<string, []>("matrix_ac_23_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_47_begin_0 = const()[name = tensor<string, []>("matrix_bd_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_47_end_0 = const()[name = tensor<string, []>("matrix_bd_47_end_0"), val = tensor<int32, [4]>([1, 8, 8, 78])];
+            tensor<bool, [4]> matrix_bd_47_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 8, 78]> matrix_bd_47_cast_fp16 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45_cast_fp16)[name = tensor<string, []>("matrix_bd_47_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_2792_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = tensor<string, []>("op_2792_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_45_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_45_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 8, 78]> _inversed_scores_45_cast_fp16 = mul(x = var_2792_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_45_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> scores_47_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_47_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_2798_cast_fp16 = softmax(axis = var_56, x = scores_47_cast_fp16)[name = tensor<string, []>("op_2798_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> input_613_cast_fp16 = select(a = var_40_to_fp16, b = var_2798_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_613_cast_fp16")];
+            tensor<bool, []> x_299_transpose_x_0 = const()[name = tensor<string, []>("x_299_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_299_transpose_y_0 = const()[name = tensor<string, []>("x_299_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 78, 64]> value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_23_cast_fp16)[name = tensor<string, []>("transpose_139")];
+            tensor<fp16, [1, 8, 8, 64]> x_299_cast_fp16 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613_cast_fp16, y = value_25_cast_fp16)[name = tensor<string, []>("x_299_cast_fp16")];
+            tensor<int32, [4]> var_2802_perm_0 = const()[name = tensor<string, []>("op_2802_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2803 = const()[name = tensor<string, []>("op_2803"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 8, 8, 64]> var_2802_cast_fp16 = transpose(perm = var_2802_perm_0, x = x_299_cast_fp16)[name = tensor<string, []>("transpose_135")];
+            tensor<fp16, [1, 8, 512]> input_615_cast_fp16 = reshape(shape = var_2803, x = var_2802_cast_fp16)[name = tensor<string, []>("input_615_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_11_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145343296)))];
+            tensor<fp16, [1, 8, 512]> linear_106_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16, x = input_615_cast_fp16)[name = tensor<string, []>("linear_106_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = tensor<string, []>("input_619_cast_fp16")];
+            tensor<int32, [1]> x_303_axes_0 = const()[name = tensor<string, []>("x_303_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145867648)))];
+            tensor<fp16, [512]> encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145868736)))];
+            tensor<fp16, [1, 8, 512]> x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = tensor<string, []>("x_303_cast_fp16")];
+            tensor<int32, [3]> input_621_perm_0 = const()[name = tensor<string, []>("input_621_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_623_pad_type_0 = const()[name = tensor<string, []>("input_623_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_623_strides_0 = const()[name = tensor<string, []>("input_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_623_pad_0 = const()[name = tensor<string, []>("input_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_623_dilations_0 = const()[name = tensor<string, []>("input_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_623_groups_0 = const()[name = tensor<string, []>("input_623_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_11_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145869824)))];
+            tensor<fp16, [1, 512, 8]> input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = tensor<string, []>("transpose_134")];
+            tensor<fp16, [1, 1024, 8]> input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16, x = input_621_cast_fp16)[name = tensor<string, []>("input_623_cast_fp16")];
+            tensor<int32, []> x_305_split_num_splits_0 = const()[name = tensor<string, []>("x_305_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_305_split_axis_0 = const()[name = tensor<string, []>("x_305_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 8]> x_305_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623_cast_fp16)[name = tensor<string, []>("x_305_split_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = tensor<string, []>("x_305_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_305_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> input_625_cast_fp16 = select(a = var_40_to_fp16, b = x_305_cast_fp16, cond = var_551)[name = tensor<string, []>("input_625_cast_fp16")];
+            tensor<bool, []> new_x_47_interleave_0 = const()[name = tensor<string, []>("new_x_47_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 16]> new_x_47_cast_fp16 = concat(axis = var_56, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = tensor<string, []>("new_x_47_cast_fp16")];
+            tensor<int32, [3]> next_cache_23_begin_0 = const()[name = tensor<string, []>("next_cache_23_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_23_end_0 = const()[name = tensor<string, []>("next_cache_23_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> next_cache_23_end_mask_0 = const()[name = tensor<string, []>("next_cache_23_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 12]> next_cache_23_cast_fp16 = slice_by_index(begin = next_cache_23_begin_0, end = next_cache_23_end_0, end_mask = next_cache_23_end_mask_0, x = new_x_47_cast_fp16)[name = tensor<string, []>("next_cache_23_cast_fp16")];
+            tensor<int32, [3]> var_2844_begin_0 = const()[name = tensor<string, []>("op_2844_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
+            tensor<int32, [3]> var_2844_end_0 = const()[name = tensor<string, []>("op_2844_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> var_2844_end_mask_0 = const()[name = tensor<string, []>("op_2844_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_2844_cast_fp16 = slice_by_index(begin = var_2844_begin_0, end = var_2844_end_0, end_mask = var_2844_end_mask_0, x = next_cache_23_cast_fp16)[name = tensor<string, []>("op_2844_cast_fp16")];
+            tensor<string, []> x_307_pad_type_0 = const()[name = tensor<string, []>("x_307_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_307_groups_0 = const()[name = tensor<string, []>("x_307_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_307_strides_0 = const()[name = tensor<string, []>("x_307_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_307_pad_0 = const()[name = tensor<string, []>("x_307_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_307_dilations_0 = const()[name = tensor<string, []>("x_307_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_11_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146918464)))];
+            tensor<fp16, [1, 512, 8]> x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16, x = new_x_47_cast_fp16)[name = tensor<string, []>("x_307_cast_fp16")];
+            tensor<int32, [3]> input_627_perm_0 = const()[name = tensor<string, []>("input_627_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_309_axes_0 = const()[name = tensor<string, []>("x_309_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146927744)))];
+            tensor<fp16, [512]> encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146928832)))];
+            tensor<fp16, [1, 8, 512]> input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = tensor<string, []>("transpose_133")];
+            tensor<fp16, [1, 8, 512]> x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = tensor<string, []>("x_309_cast_fp16")];
+            tensor<int32, [3]> input_629_perm_0 = const()[name = tensor<string, []>("input_629_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 8]> input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = tensor<string, []>("transpose_132")];
+            tensor<fp16, [1, 512, 8]> input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = tensor<string, []>("input_631_cast_fp16")];
+            tensor<string, []> x_311_pad_type_0 = const()[name = tensor<string, []>("x_311_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_311_strides_0 = const()[name = tensor<string, []>("x_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_311_pad_0 = const()[name = tensor<string, []>("x_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_311_dilations_0 = const()[name = tensor<string, []>("x_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_311_groups_0 = const()[name = tensor<string, []>("x_311_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_11_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146929920)))];
+            tensor<fp16, [1, 512, 8]> x_311_cast_fp16 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16, x = input_631_cast_fp16)[name = tensor<string, []>("x_311_cast_fp16")];
+            tensor<int32, [3]> input_633_perm_0 = const()[name = tensor<string, []>("input_633_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 8, 512]> input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = tensor<string, []>("transpose_131")];
+            tensor<fp16, [1, 8, 512]> input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = tensor<string, []>("input_635_cast_fp16")];
+            tensor<int32, [1]> input_637_axes_0 = const()[name = tensor<string, []>("input_637_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147454272)))];
+            tensor<fp16, [512]> encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147455360)))];
+            tensor<fp16, [1, 8, 512]> input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = tensor<string, []>("input_637_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_11_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147456448)))];
+            tensor<fp16, [1, 8, 2048]> linear_107_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16, x = input_637_cast_fp16)[name = tensor<string, []>("linear_107_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = tensor<string, []>("input_641_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_11_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149553664)))];
+            tensor<fp16, [1, 8, 512]> linear_108_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16, x = input_641_cast_fp16)[name = tensor<string, []>("linear_108_cast_fp16")];
+            tensor<fp16, []> var_2885_to_fp16 = const()[name = tensor<string, []>("op_2885_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_2886_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2885_to_fp16)[name = tensor<string, []>("op_2886_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2886_cast_fp16)[name = tensor<string, []>("input_647_cast_fp16")];
+            tensor<int32, [1]> input_649_axes_0 = const()[name = tensor<string, []>("input_649_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_11_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151650880)))];
+            tensor<fp16, [512]> encoder_layers_11_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151651968)))];
+            tensor<fp16, [1, 8, 512]> input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = tensor<string, []>("input_649_cast_fp16")];
+            tensor<int32, [4]> cache_49_begin_0 = const()[name = tensor<string, []>("cache_49_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_49_end_0 = const()[name = tensor<string, []>("cache_49_end_0"), val = tensor<int32, [4]>([13, 1, 70, 512])];
+            tensor<bool, [4]> cache_49_end_mask_0 = const()[name = tensor<string, []>("cache_49_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_49_squeeze_mask_0 = const()[name = tensor<string, []>("cache_49_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_49_cast_fp16 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_49_cast_fp16")];
+            tensor<int32, [4]> cache_51_begin_0 = const()[name = tensor<string, []>("cache_51_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_51_end_0 = const()[name = tensor<string, []>("cache_51_end_0"), val = tensor<int32, [4]>([13, 1, 512, 8])];
+            tensor<bool, [4]> cache_51_end_mask_0 = const()[name = tensor<string, []>("cache_51_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_51_squeeze_mask_0 = const()[name = tensor<string, []>("cache_51_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_51_cast_fp16 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_51_cast_fp16")];
+            tensor<int32, [1]> input_651_axes_0 = const()[name = tensor<string, []>("input_651_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151653056)))];
+            tensor<fp16, [512]> encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151654144)))];
+            tensor<fp16, [1, 8, 512]> input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = tensor<string, []>("input_651_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_12_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151655232)))];
+            tensor<fp16, [1, 8, 2048]> linear_109_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16, x = input_651_cast_fp16)[name = tensor<string, []>("linear_109_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = tensor<string, []>("input_655_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_12_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153752448)))];
+            tensor<fp16, [1, 8, 512]> linear_110_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16, x = input_655_cast_fp16)[name = tensor<string, []>("linear_110_cast_fp16")];
+            tensor<fp16, []> var_2920_to_fp16 = const()[name = tensor<string, []>("op_2920_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_2921_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_2920_to_fp16)[name = tensor<string, []>("op_2921_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_2921_cast_fp16)[name = tensor<string, []>("input_661_cast_fp16")];
+            tensor<int32, [1]> key_25_axes_0 = const()[name = tensor<string, []>("key_25_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155849664)))];
+            tensor<fp16, [512]> encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155850752)))];
+            tensor<fp16, [1, 8, 512]> key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = tensor<string, []>("key_25_cast_fp16")];
+            tensor<bool, []> input_663_interleave_0 = const()[name = tensor<string, []>("input_663_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 78, 512]> input_663_cast_fp16 = concat(axis = var_65, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = tensor<string, []>("input_663_cast_fp16")];
+            tensor<int32, [3]> var_2943_begin_0 = const()[name = tensor<string, []>("op_2943_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
+            tensor<int32, [3]> var_2943_end_0 = const()[name = tensor<string, []>("op_2943_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_2943_end_mask_0 = const()[name = tensor<string, []>("op_2943_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 66, 512]> var_2943_cast_fp16 = slice_by_index(begin = var_2943_begin_0, end = var_2943_end_0, end_mask = var_2943_end_mask_0, x = cache_49_cast_fp16)[name = tensor<string, []>("op_2943_cast_fp16")];
+            tensor<int32, [3]> var_2946_begin_0 = const()[name = tensor<string, []>("op_2946_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_2946_end_0 = const()[name = tensor<string, []>("op_2946_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
+            tensor<bool, [3]> var_2946_end_mask_0 = const()[name = tensor<string, []>("op_2946_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 4, 512]> var_2946_cast_fp16 = slice_by_index(begin = var_2946_begin_0, end = var_2946_end_0, end_mask = var_2946_end_mask_0, x = key_25_cast_fp16)[name = tensor<string, []>("op_2946_cast_fp16")];
+            tensor<bool, []> var_2949_interleave_0 = const()[name = tensor<string, []>("op_2949_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_2949_cast_fp16 = concat(axis = var_65, interleave = var_2949_interleave_0, values = (var_2943_cast_fp16, var_2946_cast_fp16))[name = tensor<string, []>("op_2949_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_12_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155851840)))];
+            tensor<fp16, [1, 8, 512]> linear_111_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16, x = key_25_cast_fp16)[name = tensor<string, []>("linear_111_cast_fp16")];
+            tensor<int32, [4]> var_2953 = const()[name = tensor<string, []>("op_2953"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 8, 8, 64]> q_73_cast_fp16 = reshape(shape = var_2953, x = linear_111_cast_fp16)[name = tensor<string, []>("q_73_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_12_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156376192)))];
+            tensor<fp16, [1, 78, 512]> linear_112_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16, x = input_663_cast_fp16)[name = tensor<string, []>("linear_112_cast_fp16")];
+            tensor<int32, [4]> var_2957 = const()[name = tensor<string, []>("op_2957"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> k_49_cast_fp16 = reshape(shape = var_2957, x = linear_112_cast_fp16)[name = tensor<string, []>("k_49_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_12_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156900544)))];
+            tensor<fp16, [1, 78, 512]> linear_113_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16, x = input_663_cast_fp16)[name = tensor<string, []>("linear_113_cast_fp16")];
+            tensor<int32, [4]> var_2961 = const()[name = tensor<string, []>("op_2961"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> v_25_cast_fp16 = reshape(shape = var_2961, x = linear_113_cast_fp16)[name = tensor<string, []>("v_25_cast_fp16")];
+            tensor<int32, [4]> value_27_perm_0 = const()[name = tensor<string, []>("value_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157424896)))];
+            tensor<fp16, [1, 8, 8, 64]> var_2973_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2973_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157425984)))];
+            tensor<fp16, [1, 8, 8, 64]> var_2975_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2975_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_25_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_319_transpose_x_0 = const()[name = tensor<string, []>("x_319_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_319_transpose_y_0 = const()[name = tensor<string, []>("x_319_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 155]> var_2977_to_fp16 = const()[name = tensor<string, []>("op_2977_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157427072)))];
+            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_2975_cast_fp16)[name = tensor<string, []>("transpose_129")];
+            tensor<fp16, [1, 8, 8, 155]> x_319_cast_fp16 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = var_2977_to_fp16)[name = tensor<string, []>("x_319_cast_fp16")];
+            tensor<int32, [8]> x_321_pad_0 = const()[name = tensor<string, []>("x_321_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_321_mode_0 = const()[name = tensor<string, []>("x_321_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_235_to_fp16 = const()[name = tensor<string, []>("const_235_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 8, 156]> x_321_cast_fp16 = pad(constant_val = const_235_to_fp16, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319_cast_fp16)[name = tensor<string, []>("x_321_cast_fp16")];
+            tensor<int32, [4]> var_2985 = const()[name = tensor<string, []>("op_2985"), val = tensor<int32, [4]>([1, 8, -1, 8])];
+            tensor<fp16, [1, 8, 156, 8]> x_323_cast_fp16 = reshape(shape = var_2985, x = x_321_cast_fp16)[name = tensor<string, []>("x_323_cast_fp16")];
+            tensor<int32, [4]> var_2989_begin_0 = const()[name = tensor<string, []>("op_2989_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2989_end_0 = const()[name = tensor<string, []>("op_2989_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
+            tensor<bool, [4]> var_2989_end_mask_0 = const()[name = tensor<string, []>("op_2989_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 155, 8]> var_2989_cast_fp16 = slice_by_index(begin = var_2989_begin_0, end = var_2989_end_0, end_mask = var_2989_end_mask_0, x = x_323_cast_fp16)[name = tensor<string, []>("op_2989_cast_fp16")];
+            tensor<int32, [4]> var_2990 = const()[name = tensor<string, []>("op_2990"), val = tensor<int32, [4]>([1, 8, 8, 155])];
+            tensor<fp16, [1, 8, 8, 155]> matrix_bd_49_cast_fp16 = reshape(shape = var_2990, x = var_2989_cast_fp16)[name = tensor<string, []>("matrix_bd_49_cast_fp16")];
+            tensor<bool, []> matrix_ac_25_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_25_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_25_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_25_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_75_perm_0 = const()[name = tensor<string, []>("transpose_75_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_76_perm_0 = const()[name = tensor<string, []>("transpose_76_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 78]> transpose_76 = transpose(perm = transpose_76_perm_0, x = k_49_cast_fp16)[name = tensor<string, []>("transpose_127")];
+            tensor<fp16, [1, 8, 8, 64]> transpose_75 = transpose(perm = transpose_75_perm_0, x = var_2973_cast_fp16)[name = tensor<string, []>("transpose_128")];
+            tensor<fp16, [1, 8, 8, 78]> matrix_ac_25_cast_fp16 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_75, y = transpose_76)[name = tensor<string, []>("matrix_ac_25_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_51_begin_0 = const()[name = tensor<string, []>("matrix_bd_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_51_end_0 = const()[name = tensor<string, []>("matrix_bd_51_end_0"), val = tensor<int32, [4]>([1, 8, 8, 78])];
+            tensor<bool, [4]> matrix_bd_51_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 8, 78]> matrix_bd_51_cast_fp16 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49_cast_fp16)[name = tensor<string, []>("matrix_bd_51_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_2999_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = tensor<string, []>("op_2999_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_49_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_49_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 8, 78]> _inversed_scores_49_cast_fp16 = mul(x = var_2999_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_49_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> scores_51_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_51_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_3005_cast_fp16 = softmax(axis = var_56, x = scores_51_cast_fp16)[name = tensor<string, []>("op_3005_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> input_665_cast_fp16 = select(a = var_40_to_fp16, b = var_3005_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_665_cast_fp16")];
+            tensor<bool, []> x_325_transpose_x_0 = const()[name = tensor<string, []>("x_325_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_325_transpose_y_0 = const()[name = tensor<string, []>("x_325_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 78, 64]> value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_25_cast_fp16)[name = tensor<string, []>("transpose_130")];
+            tensor<fp16, [1, 8, 8, 64]> x_325_cast_fp16 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665_cast_fp16, y = value_27_cast_fp16)[name = tensor<string, []>("x_325_cast_fp16")];
+            tensor<int32, [4]> var_3009_perm_0 = const()[name = tensor<string, []>("op_3009_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3010 = const()[name = tensor<string, []>("op_3010"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 8, 8, 64]> var_3009_cast_fp16 = transpose(perm = var_3009_perm_0, x = x_325_cast_fp16)[name = tensor<string, []>("transpose_126")];
+            tensor<fp16, [1, 8, 512]> input_667_cast_fp16 = reshape(shape = var_3010, x = var_3009_cast_fp16)[name = tensor<string, []>("input_667_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_12_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157585856)))];
+            tensor<fp16, [1, 8, 512]> linear_115_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16, x = input_667_cast_fp16)[name = tensor<string, []>("linear_115_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = tensor<string, []>("input_671_cast_fp16")];
+            tensor<int32, [1]> x_329_axes_0 = const()[name = tensor<string, []>("x_329_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158110208)))];
+            tensor<fp16, [512]> encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158111296)))];
+            tensor<fp16, [1, 8, 512]> x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = tensor<string, []>("x_329_cast_fp16")];
+            tensor<int32, [3]> input_673_perm_0 = const()[name = tensor<string, []>("input_673_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_675_pad_type_0 = const()[name = tensor<string, []>("input_675_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_675_strides_0 = const()[name = tensor<string, []>("input_675_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_675_pad_0 = const()[name = tensor<string, []>("input_675_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_675_dilations_0 = const()[name = tensor<string, []>("input_675_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_675_groups_0 = const()[name = tensor<string, []>("input_675_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_12_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158112384)))];
+            tensor<fp16, [1, 512, 8]> input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = tensor<string, []>("transpose_125")];
+            tensor<fp16, [1, 1024, 8]> input_675_cast_fp16 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16, x = input_673_cast_fp16)[name = tensor<string, []>("input_675_cast_fp16")];
+            tensor<int32, []> x_331_split_num_splits_0 = const()[name = tensor<string, []>("x_331_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_331_split_axis_0 = const()[name = tensor<string, []>("x_331_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 8]> x_331_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> x_331_split_cast_fp16_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675_cast_fp16)[name = tensor<string, []>("x_331_split_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = tensor<string, []>("x_331_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_331_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> input_677_cast_fp16 = select(a = var_40_to_fp16, b = x_331_cast_fp16, cond = var_551)[name = tensor<string, []>("input_677_cast_fp16")];
+            tensor<bool, []> new_x_51_interleave_0 = const()[name = tensor<string, []>("new_x_51_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 16]> new_x_51_cast_fp16 = concat(axis = var_56, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = tensor<string, []>("new_x_51_cast_fp16")];
+            tensor<int32, [3]> next_cache_25_begin_0 = const()[name = tensor<string, []>("next_cache_25_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_25_end_0 = const()[name = tensor<string, []>("next_cache_25_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> next_cache_25_end_mask_0 = const()[name = tensor<string, []>("next_cache_25_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 12]> next_cache_25_cast_fp16 = slice_by_index(begin = next_cache_25_begin_0, end = next_cache_25_end_0, end_mask = next_cache_25_end_mask_0, x = new_x_51_cast_fp16)[name = tensor<string, []>("next_cache_25_cast_fp16")];
+            tensor<int32, [3]> var_3051_begin_0 = const()[name = tensor<string, []>("op_3051_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
+            tensor<int32, [3]> var_3051_end_0 = const()[name = tensor<string, []>("op_3051_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> var_3051_end_mask_0 = const()[name = tensor<string, []>("op_3051_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_3051_cast_fp16 = slice_by_index(begin = var_3051_begin_0, end = var_3051_end_0, end_mask = var_3051_end_mask_0, x = next_cache_25_cast_fp16)[name = tensor<string, []>("op_3051_cast_fp16")];
+            tensor<string, []> x_333_pad_type_0 = const()[name = tensor<string, []>("x_333_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_333_groups_0 = const()[name = tensor<string, []>("x_333_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_333_strides_0 = const()[name = tensor<string, []>("x_333_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_333_pad_0 = const()[name = tensor<string, []>("x_333_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_333_dilations_0 = const()[name = tensor<string, []>("x_333_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_12_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159161024)))];
+            tensor<fp16, [1, 512, 8]> x_333_cast_fp16 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16, x = new_x_51_cast_fp16)[name = tensor<string, []>("x_333_cast_fp16")];
+            tensor<int32, [3]> input_679_perm_0 = const()[name = tensor<string, []>("input_679_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_335_axes_0 = const()[name = tensor<string, []>("x_335_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159170304)))];
+            tensor<fp16, [512]> encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159171392)))];
+            tensor<fp16, [1, 8, 512]> input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = tensor<string, []>("transpose_124")];
+            tensor<fp16, [1, 8, 512]> x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = tensor<string, []>("x_335_cast_fp16")];
+            tensor<int32, [3]> input_681_perm_0 = const()[name = tensor<string, []>("input_681_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 8]> input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = tensor<string, []>("transpose_123")];
+            tensor<fp16, [1, 512, 8]> input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = tensor<string, []>("input_683_cast_fp16")];
+            tensor<string, []> x_337_pad_type_0 = const()[name = tensor<string, []>("x_337_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_337_strides_0 = const()[name = tensor<string, []>("x_337_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_337_pad_0 = const()[name = tensor<string, []>("x_337_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_337_dilations_0 = const()[name = tensor<string, []>("x_337_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_337_groups_0 = const()[name = tensor<string, []>("x_337_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_12_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159172480)))];
+            tensor<fp16, [1, 512, 8]> x_337_cast_fp16 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16, x = input_683_cast_fp16)[name = tensor<string, []>("x_337_cast_fp16")];
+            tensor<int32, [3]> input_685_perm_0 = const()[name = tensor<string, []>("input_685_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 8, 512]> input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = tensor<string, []>("transpose_122")];
+            tensor<fp16, [1, 8, 512]> input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = tensor<string, []>("input_687_cast_fp16")];
+            tensor<int32, [1]> input_689_axes_0 = const()[name = tensor<string, []>("input_689_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159696832)))];
+            tensor<fp16, [512]> encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159697920)))];
+            tensor<fp16, [1, 8, 512]> input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = tensor<string, []>("input_689_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_12_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159699008)))];
+            tensor<fp16, [1, 8, 2048]> linear_116_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16, x = input_689_cast_fp16)[name = tensor<string, []>("linear_116_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = tensor<string, []>("input_693_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_12_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161796224)))];
+            tensor<fp16, [1, 8, 512]> linear_117_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16, x = input_693_cast_fp16)[name = tensor<string, []>("linear_117_cast_fp16")];
+            tensor<fp16, []> var_3092_to_fp16 = const()[name = tensor<string, []>("op_3092_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_3093_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_3092_to_fp16)[name = tensor<string, []>("op_3093_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_3093_cast_fp16)[name = tensor<string, []>("input_699_cast_fp16")];
+            tensor<int32, [1]> input_701_axes_0 = const()[name = tensor<string, []>("input_701_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_12_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163893440)))];
+            tensor<fp16, [512]> encoder_layers_12_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163894528)))];
+            tensor<fp16, [1, 8, 512]> input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = tensor<string, []>("input_701_cast_fp16")];
+            tensor<int32, [4]> cache_53_begin_0 = const()[name = tensor<string, []>("cache_53_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_53_end_0 = const()[name = tensor<string, []>("cache_53_end_0"), val = tensor<int32, [4]>([14, 1, 70, 512])];
+            tensor<bool, [4]> cache_53_end_mask_0 = const()[name = tensor<string, []>("cache_53_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_53_squeeze_mask_0 = const()[name = tensor<string, []>("cache_53_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_53_cast_fp16 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_53_cast_fp16")];
+            tensor<int32, [4]> cache_55_begin_0 = const()[name = tensor<string, []>("cache_55_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_55_end_0 = const()[name = tensor<string, []>("cache_55_end_0"), val = tensor<int32, [4]>([14, 1, 512, 8])];
+            tensor<bool, [4]> cache_55_end_mask_0 = const()[name = tensor<string, []>("cache_55_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_55_squeeze_mask_0 = const()[name = tensor<string, []>("cache_55_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_55_cast_fp16 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_55_cast_fp16")];
+            tensor<int32, [1]> input_703_axes_0 = const()[name = tensor<string, []>("input_703_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163895616)))];
+            tensor<fp16, [512]> encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163896704)))];
+            tensor<fp16, [1, 8, 512]> input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = tensor<string, []>("input_703_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_13_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163897792)))];
+            tensor<fp16, [1, 8, 2048]> linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16, x = input_703_cast_fp16)[name = tensor<string, []>("linear_118_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = tensor<string, []>("input_707_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_13_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165995008)))];
+            tensor<fp16, [1, 8, 512]> linear_119_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16, x = input_707_cast_fp16)[name = tensor<string, []>("linear_119_cast_fp16")];
+            tensor<fp16, []> var_3127_to_fp16 = const()[name = tensor<string, []>("op_3127_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_3128_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_3127_to_fp16)[name = tensor<string, []>("op_3128_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_3128_cast_fp16)[name = tensor<string, []>("input_713_cast_fp16")];
+            tensor<int32, [1]> key_27_axes_0 = const()[name = tensor<string, []>("key_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168092224)))];
+            tensor<fp16, [512]> encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168093312)))];
+            tensor<fp16, [1, 8, 512]> key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = tensor<string, []>("key_27_cast_fp16")];
+            tensor<bool, []> input_715_interleave_0 = const()[name = tensor<string, []>("input_715_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 78, 512]> input_715_cast_fp16 = concat(axis = var_65, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = tensor<string, []>("input_715_cast_fp16")];
+            tensor<int32, [3]> var_3150_begin_0 = const()[name = tensor<string, []>("op_3150_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
+            tensor<int32, [3]> var_3150_end_0 = const()[name = tensor<string, []>("op_3150_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_3150_end_mask_0 = const()[name = tensor<string, []>("op_3150_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 66, 512]> var_3150_cast_fp16 = slice_by_index(begin = var_3150_begin_0, end = var_3150_end_0, end_mask = var_3150_end_mask_0, x = cache_53_cast_fp16)[name = tensor<string, []>("op_3150_cast_fp16")];
+            tensor<int32, [3]> var_3153_begin_0 = const()[name = tensor<string, []>("op_3153_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_3153_end_0 = const()[name = tensor<string, []>("op_3153_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
+            tensor<bool, [3]> var_3153_end_mask_0 = const()[name = tensor<string, []>("op_3153_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 4, 512]> var_3153_cast_fp16 = slice_by_index(begin = var_3153_begin_0, end = var_3153_end_0, end_mask = var_3153_end_mask_0, x = key_27_cast_fp16)[name = tensor<string, []>("op_3153_cast_fp16")];
+            tensor<bool, []> var_3156_interleave_0 = const()[name = tensor<string, []>("op_3156_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_3156_cast_fp16 = concat(axis = var_65, interleave = var_3156_interleave_0, values = (var_3150_cast_fp16, var_3153_cast_fp16))[name = tensor<string, []>("op_3156_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_13_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168094400)))];
+            tensor<fp16, [1, 8, 512]> linear_120_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16, x = key_27_cast_fp16)[name = tensor<string, []>("linear_120_cast_fp16")];
+            tensor<int32, [4]> var_3160 = const()[name = tensor<string, []>("op_3160"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 8, 8, 64]> q_79_cast_fp16 = reshape(shape = var_3160, x = linear_120_cast_fp16)[name = tensor<string, []>("q_79_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_13_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168618752)))];
+            tensor<fp16, [1, 78, 512]> linear_121_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16, x = input_715_cast_fp16)[name = tensor<string, []>("linear_121_cast_fp16")];
+            tensor<int32, [4]> var_3164 = const()[name = tensor<string, []>("op_3164"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> k_53_cast_fp16 = reshape(shape = var_3164, x = linear_121_cast_fp16)[name = tensor<string, []>("k_53_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_13_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169143104)))];
+            tensor<fp16, [1, 78, 512]> linear_122_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16, x = input_715_cast_fp16)[name = tensor<string, []>("linear_122_cast_fp16")];
+            tensor<int32, [4]> var_3168 = const()[name = tensor<string, []>("op_3168"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> v_27_cast_fp16 = reshape(shape = var_3168, x = linear_122_cast_fp16)[name = tensor<string, []>("v_27_cast_fp16")];
+            tensor<int32, [4]> value_29_perm_0 = const()[name = tensor<string, []>("value_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169667456)))];
+            tensor<fp16, [1, 8, 8, 64]> var_3180_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3180_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169668544)))];
+            tensor<fp16, [1, 8, 8, 64]> var_3182_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3182_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_27_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_345_transpose_x_0 = const()[name = tensor<string, []>("x_345_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_345_transpose_y_0 = const()[name = tensor<string, []>("x_345_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 155]> var_3184_to_fp16 = const()[name = tensor<string, []>("op_3184_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169669632)))];
+            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3182_cast_fp16)[name = tensor<string, []>("transpose_120")];
+            tensor<fp16, [1, 8, 8, 155]> x_345_cast_fp16 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = var_3184_to_fp16)[name = tensor<string, []>("x_345_cast_fp16")];
+            tensor<int32, [8]> x_347_pad_0 = const()[name = tensor<string, []>("x_347_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_347_mode_0 = const()[name = tensor<string, []>("x_347_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_248_to_fp16 = const()[name = tensor<string, []>("const_248_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 8, 156]> x_347_cast_fp16 = pad(constant_val = const_248_to_fp16, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345_cast_fp16)[name = tensor<string, []>("x_347_cast_fp16")];
+            tensor<int32, [4]> var_3192 = const()[name = tensor<string, []>("op_3192"), val = tensor<int32, [4]>([1, 8, -1, 8])];
+            tensor<fp16, [1, 8, 156, 8]> x_349_cast_fp16 = reshape(shape = var_3192, x = x_347_cast_fp16)[name = tensor<string, []>("x_349_cast_fp16")];
+            tensor<int32, [4]> var_3196_begin_0 = const()[name = tensor<string, []>("op_3196_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3196_end_0 = const()[name = tensor<string, []>("op_3196_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
+            tensor<bool, [4]> var_3196_end_mask_0 = const()[name = tensor<string, []>("op_3196_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 155, 8]> var_3196_cast_fp16 = slice_by_index(begin = var_3196_begin_0, end = var_3196_end_0, end_mask = var_3196_end_mask_0, x = x_349_cast_fp16)[name = tensor<string, []>("op_3196_cast_fp16")];
+            tensor<int32, [4]> var_3197 = const()[name = tensor<string, []>("op_3197"), val = tensor<int32, [4]>([1, 8, 8, 155])];
+            tensor<fp16, [1, 8, 8, 155]> matrix_bd_53_cast_fp16 = reshape(shape = var_3197, x = var_3196_cast_fp16)[name = tensor<string, []>("matrix_bd_53_cast_fp16")];
+            tensor<bool, []> matrix_ac_27_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_27_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_27_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_27_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_77_perm_0 = const()[name = tensor<string, []>("transpose_77_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_78_perm_0 = const()[name = tensor<string, []>("transpose_78_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 78]> transpose_78 = transpose(perm = transpose_78_perm_0, x = k_53_cast_fp16)[name = tensor<string, []>("transpose_118")];
+            tensor<fp16, [1, 8, 8, 64]> transpose_77 = transpose(perm = transpose_77_perm_0, x = var_3180_cast_fp16)[name = tensor<string, []>("transpose_119")];
+            tensor<fp16, [1, 8, 8, 78]> matrix_ac_27_cast_fp16 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_77, y = transpose_78)[name = tensor<string, []>("matrix_ac_27_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_55_begin_0 = const()[name = tensor<string, []>("matrix_bd_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_55_end_0 = const()[name = tensor<string, []>("matrix_bd_55_end_0"), val = tensor<int32, [4]>([1, 8, 8, 78])];
+            tensor<bool, [4]> matrix_bd_55_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 8, 78]> matrix_bd_55_cast_fp16 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53_cast_fp16)[name = tensor<string, []>("matrix_bd_55_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_3206_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = tensor<string, []>("op_3206_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_53_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_53_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 8, 78]> _inversed_scores_53_cast_fp16 = mul(x = var_3206_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_53_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> scores_55_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_55_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_3212_cast_fp16 = softmax(axis = var_56, x = scores_55_cast_fp16)[name = tensor<string, []>("op_3212_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> input_717_cast_fp16 = select(a = var_40_to_fp16, b = var_3212_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_717_cast_fp16")];
+            tensor<bool, []> x_351_transpose_x_0 = const()[name = tensor<string, []>("x_351_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_351_transpose_y_0 = const()[name = tensor<string, []>("x_351_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 78, 64]> value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_27_cast_fp16)[name = tensor<string, []>("transpose_121")];
+            tensor<fp16, [1, 8, 8, 64]> x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717_cast_fp16, y = value_29_cast_fp16)[name = tensor<string, []>("x_351_cast_fp16")];
+            tensor<int32, [4]> var_3216_perm_0 = const()[name = tensor<string, []>("op_3216_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3217 = const()[name = tensor<string, []>("op_3217"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 8, 8, 64]> var_3216_cast_fp16 = transpose(perm = var_3216_perm_0, x = x_351_cast_fp16)[name = tensor<string, []>("transpose_117")];
+            tensor<fp16, [1, 8, 512]> input_719_cast_fp16 = reshape(shape = var_3217, x = var_3216_cast_fp16)[name = tensor<string, []>("input_719_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_13_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169828416)))];
+            tensor<fp16, [1, 8, 512]> linear_124_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16, x = input_719_cast_fp16)[name = tensor<string, []>("linear_124_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = tensor<string, []>("input_723_cast_fp16")];
+            tensor<int32, [1]> x_355_axes_0 = const()[name = tensor<string, []>("x_355_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170352768)))];
+            tensor<fp16, [512]> encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170353856)))];
+            tensor<fp16, [1, 8, 512]> x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = tensor<string, []>("x_355_cast_fp16")];
+            tensor<int32, [3]> input_725_perm_0 = const()[name = tensor<string, []>("input_725_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_727_pad_type_0 = const()[name = tensor<string, []>("input_727_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_727_strides_0 = const()[name = tensor<string, []>("input_727_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_727_pad_0 = const()[name = tensor<string, []>("input_727_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_727_dilations_0 = const()[name = tensor<string, []>("input_727_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_727_groups_0 = const()[name = tensor<string, []>("input_727_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_13_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170354944)))];
+            tensor<fp16, [1, 512, 8]> input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = tensor<string, []>("transpose_116")];
+            tensor<fp16, [1, 1024, 8]> input_727_cast_fp16 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16, x = input_725_cast_fp16)[name = tensor<string, []>("input_727_cast_fp16")];
+            tensor<int32, []> x_357_split_num_splits_0 = const()[name = tensor<string, []>("x_357_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_357_split_axis_0 = const()[name = tensor<string, []>("x_357_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 8]> x_357_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727_cast_fp16)[name = tensor<string, []>("x_357_split_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = tensor<string, []>("x_357_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_357_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> input_729_cast_fp16 = select(a = var_40_to_fp16, b = x_357_cast_fp16, cond = var_551)[name = tensor<string, []>("input_729_cast_fp16")];
+            tensor<bool, []> new_x_55_interleave_0 = const()[name = tensor<string, []>("new_x_55_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 16]> new_x_55_cast_fp16 = concat(axis = var_56, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = tensor<string, []>("new_x_55_cast_fp16")];
+            tensor<int32, [3]> next_cache_27_begin_0 = const()[name = tensor<string, []>("next_cache_27_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_27_end_0 = const()[name = tensor<string, []>("next_cache_27_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> next_cache_27_end_mask_0 = const()[name = tensor<string, []>("next_cache_27_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 12]> next_cache_27_cast_fp16 = slice_by_index(begin = next_cache_27_begin_0, end = next_cache_27_end_0, end_mask = next_cache_27_end_mask_0, x = new_x_55_cast_fp16)[name = tensor<string, []>("next_cache_27_cast_fp16")];
+            tensor<int32, [3]> var_3258_begin_0 = const()[name = tensor<string, []>("op_3258_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
+            tensor<int32, [3]> var_3258_end_0 = const()[name = tensor<string, []>("op_3258_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> var_3258_end_mask_0 = const()[name = tensor<string, []>("op_3258_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_3258_cast_fp16 = slice_by_index(begin = var_3258_begin_0, end = var_3258_end_0, end_mask = var_3258_end_mask_0, x = next_cache_27_cast_fp16)[name = tensor<string, []>("op_3258_cast_fp16")];
+            tensor<string, []> x_359_pad_type_0 = const()[name = tensor<string, []>("x_359_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_359_groups_0 = const()[name = tensor<string, []>("x_359_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_359_strides_0 = const()[name = tensor<string, []>("x_359_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_359_pad_0 = const()[name = tensor<string, []>("x_359_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_359_dilations_0 = const()[name = tensor<string, []>("x_359_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_13_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171403584)))];
+            tensor<fp16, [1, 512, 8]> x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16, x = new_x_55_cast_fp16)[name = tensor<string, []>("x_359_cast_fp16")];
+            tensor<int32, [3]> input_731_perm_0 = const()[name = tensor<string, []>("input_731_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_361_axes_0 = const()[name = tensor<string, []>("x_361_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171412864)))];
+            tensor<fp16, [512]> encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171413952)))];
+            tensor<fp16, [1, 8, 512]> input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = tensor<string, []>("transpose_115")];
+            tensor<fp16, [1, 8, 512]> x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = tensor<string, []>("x_361_cast_fp16")];
+            tensor<int32, [3]> input_733_perm_0 = const()[name = tensor<string, []>("input_733_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 8]> input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = tensor<string, []>("transpose_114")];
+            tensor<fp16, [1, 512, 8]> input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = tensor<string, []>("input_735_cast_fp16")];
+            tensor<string, []> x_363_pad_type_0 = const()[name = tensor<string, []>("x_363_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_363_strides_0 = const()[name = tensor<string, []>("x_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_363_pad_0 = const()[name = tensor<string, []>("x_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_363_dilations_0 = const()[name = tensor<string, []>("x_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_363_groups_0 = const()[name = tensor<string, []>("x_363_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_13_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171415040)))];
+            tensor<fp16, [1, 512, 8]> x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16, x = input_735_cast_fp16)[name = tensor<string, []>("x_363_cast_fp16")];
+            tensor<int32, [3]> input_737_perm_0 = const()[name = tensor<string, []>("input_737_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 8, 512]> input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = tensor<string, []>("transpose_113")];
+            tensor<fp16, [1, 8, 512]> input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = tensor<string, []>("input_739_cast_fp16")];
+            tensor<int32, [1]> input_741_axes_0 = const()[name = tensor<string, []>("input_741_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171939392)))];
+            tensor<fp16, [512]> encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171940480)))];
+            tensor<fp16, [1, 8, 512]> input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = tensor<string, []>("input_741_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_13_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171941568)))];
+            tensor<fp16, [1, 8, 2048]> linear_125_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16, x = input_741_cast_fp16)[name = tensor<string, []>("linear_125_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = tensor<string, []>("input_745_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_13_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174038784)))];
+            tensor<fp16, [1, 8, 512]> linear_126_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16, x = input_745_cast_fp16)[name = tensor<string, []>("linear_126_cast_fp16")];
+            tensor<fp16, []> var_3299_to_fp16 = const()[name = tensor<string, []>("op_3299_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_3300_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3299_to_fp16)[name = tensor<string, []>("op_3300_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3300_cast_fp16)[name = tensor<string, []>("input_751_cast_fp16")];
+            tensor<int32, [1]> input_753_axes_0 = const()[name = tensor<string, []>("input_753_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_13_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176136000)))];
+            tensor<fp16, [512]> encoder_layers_13_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176137088)))];
+            tensor<fp16, [1, 8, 512]> input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = tensor<string, []>("input_753_cast_fp16")];
+            tensor<int32, [4]> cache_57_begin_0 = const()[name = tensor<string, []>("cache_57_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_57_end_0 = const()[name = tensor<string, []>("cache_57_end_0"), val = tensor<int32, [4]>([15, 1, 70, 512])];
+            tensor<bool, [4]> cache_57_end_mask_0 = const()[name = tensor<string, []>("cache_57_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_57_squeeze_mask_0 = const()[name = tensor<string, []>("cache_57_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_57_cast_fp16 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_57_cast_fp16")];
+            tensor<int32, [4]> cache_59_begin_0 = const()[name = tensor<string, []>("cache_59_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_59_end_0 = const()[name = tensor<string, []>("cache_59_end_0"), val = tensor<int32, [4]>([15, 1, 512, 8])];
+            tensor<bool, [4]> cache_59_end_mask_0 = const()[name = tensor<string, []>("cache_59_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_59_squeeze_mask_0 = const()[name = tensor<string, []>("cache_59_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_59_cast_fp16 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_59_cast_fp16")];
+            tensor<int32, [1]> input_755_axes_0 = const()[name = tensor<string, []>("input_755_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176138176)))];
+            tensor<fp16, [512]> encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176139264)))];
+            tensor<fp16, [1, 8, 512]> input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = tensor<string, []>("input_755_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_14_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176140352)))];
+            tensor<fp16, [1, 8, 2048]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16, x = input_755_cast_fp16)[name = tensor<string, []>("linear_127_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = tensor<string, []>("input_759_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_14_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178237568)))];
+            tensor<fp16, [1, 8, 512]> linear_128_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16, x = input_759_cast_fp16)[name = tensor<string, []>("linear_128_cast_fp16")];
+            tensor<fp16, []> var_3334_to_fp16 = const()[name = tensor<string, []>("op_3334_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_3335_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3334_to_fp16)[name = tensor<string, []>("op_3335_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3335_cast_fp16)[name = tensor<string, []>("input_765_cast_fp16")];
+            tensor<int32, [1]> key_29_axes_0 = const()[name = tensor<string, []>("key_29_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180334784)))];
+            tensor<fp16, [512]> encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180335872)))];
+            tensor<fp16, [1, 8, 512]> key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = tensor<string, []>("key_29_cast_fp16")];
+            tensor<bool, []> input_767_interleave_0 = const()[name = tensor<string, []>("input_767_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 78, 512]> input_767_cast_fp16 = concat(axis = var_65, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = tensor<string, []>("input_767_cast_fp16")];
+            tensor<int32, [3]> var_3357_begin_0 = const()[name = tensor<string, []>("op_3357_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
+            tensor<int32, [3]> var_3357_end_0 = const()[name = tensor<string, []>("op_3357_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_3357_end_mask_0 = const()[name = tensor<string, []>("op_3357_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 66, 512]> var_3357_cast_fp16 = slice_by_index(begin = var_3357_begin_0, end = var_3357_end_0, end_mask = var_3357_end_mask_0, x = cache_57_cast_fp16)[name = tensor<string, []>("op_3357_cast_fp16")];
+            tensor<int32, [3]> var_3360_begin_0 = const()[name = tensor<string, []>("op_3360_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_3360_end_0 = const()[name = tensor<string, []>("op_3360_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
+            tensor<bool, [3]> var_3360_end_mask_0 = const()[name = tensor<string, []>("op_3360_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 4, 512]> var_3360_cast_fp16 = slice_by_index(begin = var_3360_begin_0, end = var_3360_end_0, end_mask = var_3360_end_mask_0, x = key_29_cast_fp16)[name = tensor<string, []>("op_3360_cast_fp16")];
+            tensor<bool, []> var_3363_interleave_0 = const()[name = tensor<string, []>("op_3363_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_3363_cast_fp16 = concat(axis = var_65, interleave = var_3363_interleave_0, values = (var_3357_cast_fp16, var_3360_cast_fp16))[name = tensor<string, []>("op_3363_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_14_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180336960)))];
+            tensor<fp16, [1, 8, 512]> linear_129_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16, x = key_29_cast_fp16)[name = tensor<string, []>("linear_129_cast_fp16")];
+            tensor<int32, [4]> var_3367 = const()[name = tensor<string, []>("op_3367"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 8, 8, 64]> q_85_cast_fp16 = reshape(shape = var_3367, x = linear_129_cast_fp16)[name = tensor<string, []>("q_85_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_14_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180861312)))];
+            tensor<fp16, [1, 78, 512]> linear_130_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16, x = input_767_cast_fp16)[name = tensor<string, []>("linear_130_cast_fp16")];
+            tensor<int32, [4]> var_3371 = const()[name = tensor<string, []>("op_3371"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> k_57_cast_fp16 = reshape(shape = var_3371, x = linear_130_cast_fp16)[name = tensor<string, []>("k_57_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_14_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181385664)))];
+            tensor<fp16, [1, 78, 512]> linear_131_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16, x = input_767_cast_fp16)[name = tensor<string, []>("linear_131_cast_fp16")];
+            tensor<int32, [4]> var_3375 = const()[name = tensor<string, []>("op_3375"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> v_29_cast_fp16 = reshape(shape = var_3375, x = linear_131_cast_fp16)[name = tensor<string, []>("v_29_cast_fp16")];
+            tensor<int32, [4]> value_31_perm_0 = const()[name = tensor<string, []>("value_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181910016)))];
+            tensor<fp16, [1, 8, 8, 64]> var_3387_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3387_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181911104)))];
+            tensor<fp16, [1, 8, 8, 64]> var_3389_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3389_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_29_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_371_transpose_x_0 = const()[name = tensor<string, []>("x_371_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_371_transpose_y_0 = const()[name = tensor<string, []>("x_371_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 155]> var_3391_to_fp16 = const()[name = tensor<string, []>("op_3391_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181912192)))];
+            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3389_cast_fp16)[name = tensor<string, []>("transpose_111")];
+            tensor<fp16, [1, 8, 8, 155]> x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = var_3391_to_fp16)[name = tensor<string, []>("x_371_cast_fp16")];
+            tensor<int32, [8]> x_373_pad_0 = const()[name = tensor<string, []>("x_373_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_373_mode_0 = const()[name = tensor<string, []>("x_373_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_261_to_fp16 = const()[name = tensor<string, []>("const_261_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 8, 156]> x_373_cast_fp16 = pad(constant_val = const_261_to_fp16, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371_cast_fp16)[name = tensor<string, []>("x_373_cast_fp16")];
+            tensor<int32, [4]> var_3399 = const()[name = tensor<string, []>("op_3399"), val = tensor<int32, [4]>([1, 8, -1, 8])];
+            tensor<fp16, [1, 8, 156, 8]> x_375_cast_fp16 = reshape(shape = var_3399, x = x_373_cast_fp16)[name = tensor<string, []>("x_375_cast_fp16")];
+            tensor<int32, [4]> var_3403_begin_0 = const()[name = tensor<string, []>("op_3403_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3403_end_0 = const()[name = tensor<string, []>("op_3403_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
+            tensor<bool, [4]> var_3403_end_mask_0 = const()[name = tensor<string, []>("op_3403_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 155, 8]> var_3403_cast_fp16 = slice_by_index(begin = var_3403_begin_0, end = var_3403_end_0, end_mask = var_3403_end_mask_0, x = x_375_cast_fp16)[name = tensor<string, []>("op_3403_cast_fp16")];
+            tensor<int32, [4]> var_3404 = const()[name = tensor<string, []>("op_3404"), val = tensor<int32, [4]>([1, 8, 8, 155])];
+            tensor<fp16, [1, 8, 8, 155]> matrix_bd_57_cast_fp16 = reshape(shape = var_3404, x = var_3403_cast_fp16)[name = tensor<string, []>("matrix_bd_57_cast_fp16")];
+            tensor<bool, []> matrix_ac_29_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_29_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_29_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_29_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_79_perm_0 = const()[name = tensor<string, []>("transpose_79_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_80_perm_0 = const()[name = tensor<string, []>("transpose_80_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 78]> transpose_80 = transpose(perm = transpose_80_perm_0, x = k_57_cast_fp16)[name = tensor<string, []>("transpose_109")];
+            tensor<fp16, [1, 8, 8, 64]> transpose_79 = transpose(perm = transpose_79_perm_0, x = var_3387_cast_fp16)[name = tensor<string, []>("transpose_110")];
+            tensor<fp16, [1, 8, 8, 78]> matrix_ac_29_cast_fp16 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_79, y = transpose_80)[name = tensor<string, []>("matrix_ac_29_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_59_begin_0 = const()[name = tensor<string, []>("matrix_bd_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_59_end_0 = const()[name = tensor<string, []>("matrix_bd_59_end_0"), val = tensor<int32, [4]>([1, 8, 8, 78])];
+            tensor<bool, [4]> matrix_bd_59_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 8, 78]> matrix_bd_59_cast_fp16 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57_cast_fp16)[name = tensor<string, []>("matrix_bd_59_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_3413_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = tensor<string, []>("op_3413_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_57_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_57_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 8, 78]> _inversed_scores_57_cast_fp16 = mul(x = var_3413_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_57_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> scores_59_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_59_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_3419_cast_fp16 = softmax(axis = var_56, x = scores_59_cast_fp16)[name = tensor<string, []>("op_3419_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> input_769_cast_fp16 = select(a = var_40_to_fp16, b = var_3419_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_769_cast_fp16")];
+            tensor<bool, []> x_377_transpose_x_0 = const()[name = tensor<string, []>("x_377_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_377_transpose_y_0 = const()[name = tensor<string, []>("x_377_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 78, 64]> value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_29_cast_fp16)[name = tensor<string, []>("transpose_112")];
+            tensor<fp16, [1, 8, 8, 64]> x_377_cast_fp16 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769_cast_fp16, y = value_31_cast_fp16)[name = tensor<string, []>("x_377_cast_fp16")];
+            tensor<int32, [4]> var_3423_perm_0 = const()[name = tensor<string, []>("op_3423_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3424 = const()[name = tensor<string, []>("op_3424"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 8, 8, 64]> var_3423_cast_fp16 = transpose(perm = var_3423_perm_0, x = x_377_cast_fp16)[name = tensor<string, []>("transpose_108")];
+            tensor<fp16, [1, 8, 512]> input_771_cast_fp16 = reshape(shape = var_3424, x = var_3423_cast_fp16)[name = tensor<string, []>("input_771_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_14_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182070976)))];
+            tensor<fp16, [1, 8, 512]> linear_133_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16, x = input_771_cast_fp16)[name = tensor<string, []>("linear_133_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = tensor<string, []>("input_775_cast_fp16")];
+            tensor<int32, [1]> x_381_axes_0 = const()[name = tensor<string, []>("x_381_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182595328)))];
+            tensor<fp16, [512]> encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182596416)))];
+            tensor<fp16, [1, 8, 512]> x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = tensor<string, []>("x_381_cast_fp16")];
+            tensor<int32, [3]> input_777_perm_0 = const()[name = tensor<string, []>("input_777_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_779_pad_type_0 = const()[name = tensor<string, []>("input_779_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_779_strides_0 = const()[name = tensor<string, []>("input_779_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_779_pad_0 = const()[name = tensor<string, []>("input_779_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_779_dilations_0 = const()[name = tensor<string, []>("input_779_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_779_groups_0 = const()[name = tensor<string, []>("input_779_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_14_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182597504)))];
+            tensor<fp16, [1, 512, 8]> input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = tensor<string, []>("transpose_107")];
+            tensor<fp16, [1, 1024, 8]> input_779_cast_fp16 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16, x = input_777_cast_fp16)[name = tensor<string, []>("input_779_cast_fp16")];
+            tensor<int32, []> x_383_split_num_splits_0 = const()[name = tensor<string, []>("x_383_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_383_split_axis_0 = const()[name = tensor<string, []>("x_383_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 8]> x_383_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> x_383_split_cast_fp16_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779_cast_fp16)[name = tensor<string, []>("x_383_split_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = tensor<string, []>("x_383_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_383_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> input_781_cast_fp16 = select(a = var_40_to_fp16, b = x_383_cast_fp16, cond = var_551)[name = tensor<string, []>("input_781_cast_fp16")];
+            tensor<bool, []> new_x_59_interleave_0 = const()[name = tensor<string, []>("new_x_59_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 16]> new_x_59_cast_fp16 = concat(axis = var_56, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = tensor<string, []>("new_x_59_cast_fp16")];
+            tensor<int32, [3]> next_cache_29_begin_0 = const()[name = tensor<string, []>("next_cache_29_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_29_end_0 = const()[name = tensor<string, []>("next_cache_29_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> next_cache_29_end_mask_0 = const()[name = tensor<string, []>("next_cache_29_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 12]> next_cache_29_cast_fp16 = slice_by_index(begin = next_cache_29_begin_0, end = next_cache_29_end_0, end_mask = next_cache_29_end_mask_0, x = new_x_59_cast_fp16)[name = tensor<string, []>("next_cache_29_cast_fp16")];
+            tensor<int32, [3]> var_3465_begin_0 = const()[name = tensor<string, []>("op_3465_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
+            tensor<int32, [3]> var_3465_end_0 = const()[name = tensor<string, []>("op_3465_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> var_3465_end_mask_0 = const()[name = tensor<string, []>("op_3465_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_3465_cast_fp16 = slice_by_index(begin = var_3465_begin_0, end = var_3465_end_0, end_mask = var_3465_end_mask_0, x = next_cache_29_cast_fp16)[name = tensor<string, []>("op_3465_cast_fp16")];
+            tensor<string, []> x_385_pad_type_0 = const()[name = tensor<string, []>("x_385_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_385_groups_0 = const()[name = tensor<string, []>("x_385_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_385_strides_0 = const()[name = tensor<string, []>("x_385_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_385_pad_0 = const()[name = tensor<string, []>("x_385_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_385_dilations_0 = const()[name = tensor<string, []>("x_385_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_14_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183646144)))];
+            tensor<fp16, [1, 512, 8]> x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16, x = new_x_59_cast_fp16)[name = tensor<string, []>("x_385_cast_fp16")];
+            tensor<int32, [3]> input_783_perm_0 = const()[name = tensor<string, []>("input_783_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_387_axes_0 = const()[name = tensor<string, []>("x_387_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183655424)))];
+            tensor<fp16, [512]> encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183656512)))];
+            tensor<fp16, [1, 8, 512]> input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = tensor<string, []>("transpose_106")];
+            tensor<fp16, [1, 8, 512]> x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = tensor<string, []>("x_387_cast_fp16")];
+            tensor<int32, [3]> input_785_perm_0 = const()[name = tensor<string, []>("input_785_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 8]> input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = tensor<string, []>("transpose_105")];
+            tensor<fp16, [1, 512, 8]> input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = tensor<string, []>("input_787_cast_fp16")];
+            tensor<string, []> x_389_pad_type_0 = const()[name = tensor<string, []>("x_389_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_389_strides_0 = const()[name = tensor<string, []>("x_389_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_389_pad_0 = const()[name = tensor<string, []>("x_389_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_389_dilations_0 = const()[name = tensor<string, []>("x_389_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_389_groups_0 = const()[name = tensor<string, []>("x_389_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_14_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183657600)))];
+            tensor<fp16, [1, 512, 8]> x_389_cast_fp16 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16, x = input_787_cast_fp16)[name = tensor<string, []>("x_389_cast_fp16")];
+            tensor<int32, [3]> input_789_perm_0 = const()[name = tensor<string, []>("input_789_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 8, 512]> input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = tensor<string, []>("transpose_104")];
+            tensor<fp16, [1, 8, 512]> input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = tensor<string, []>("input_791_cast_fp16")];
+            tensor<int32, [1]> input_793_axes_0 = const()[name = tensor<string, []>("input_793_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184181952)))];
+            tensor<fp16, [512]> encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184183040)))];
+            tensor<fp16, [1, 8, 512]> input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = tensor<string, []>("input_793_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_14_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184184128)))];
+            tensor<fp16, [1, 8, 2048]> linear_134_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16, x = input_793_cast_fp16)[name = tensor<string, []>("linear_134_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = tensor<string, []>("input_797_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_14_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186281344)))];
+            tensor<fp16, [1, 8, 512]> linear_135_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16, x = input_797_cast_fp16)[name = tensor<string, []>("linear_135_cast_fp16")];
+            tensor<fp16, []> var_3506_to_fp16 = const()[name = tensor<string, []>("op_3506_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_3507_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3506_to_fp16)[name = tensor<string, []>("op_3507_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3507_cast_fp16)[name = tensor<string, []>("input_803_cast_fp16")];
+            tensor<int32, [1]> input_805_axes_0 = const()[name = tensor<string, []>("input_805_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_14_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188378560)))];
+            tensor<fp16, [512]> encoder_layers_14_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188379648)))];
+            tensor<fp16, [1, 8, 512]> input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = tensor<string, []>("input_805_cast_fp16")];
+            tensor<int32, [4]> cache_61_begin_0 = const()[name = tensor<string, []>("cache_61_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_61_end_0 = const()[name = tensor<string, []>("cache_61_end_0"), val = tensor<int32, [4]>([16, 1, 70, 512])];
+            tensor<bool, [4]> cache_61_end_mask_0 = const()[name = tensor<string, []>("cache_61_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_61_squeeze_mask_0 = const()[name = tensor<string, []>("cache_61_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_61_cast_fp16 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_61_cast_fp16")];
+            tensor<int32, [4]> cache_63_begin_0 = const()[name = tensor<string, []>("cache_63_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_63_end_0 = const()[name = tensor<string, []>("cache_63_end_0"), val = tensor<int32, [4]>([16, 1, 512, 8])];
+            tensor<bool, [4]> cache_63_end_mask_0 = const()[name = tensor<string, []>("cache_63_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_63_squeeze_mask_0 = const()[name = tensor<string, []>("cache_63_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_63_cast_fp16 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_63_cast_fp16")];
+            tensor<int32, [1]> input_807_axes_0 = const()[name = tensor<string, []>("input_807_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188380736)))];
+            tensor<fp16, [512]> encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188381824)))];
+            tensor<fp16, [1, 8, 512]> input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = tensor<string, []>("input_807_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_15_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188382912)))];
+            tensor<fp16, [1, 8, 2048]> linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16, x = input_807_cast_fp16)[name = tensor<string, []>("linear_136_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = tensor<string, []>("input_811_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_15_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190480128)))];
+            tensor<fp16, [1, 8, 512]> linear_137_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16, x = input_811_cast_fp16)[name = tensor<string, []>("linear_137_cast_fp16")];
+            tensor<fp16, []> var_3541_to_fp16 = const()[name = tensor<string, []>("op_3541_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_3542_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3541_to_fp16)[name = tensor<string, []>("op_3542_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3542_cast_fp16)[name = tensor<string, []>("input_817_cast_fp16")];
+            tensor<int32, [1]> key_31_axes_0 = const()[name = tensor<string, []>("key_31_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192577344)))];
+            tensor<fp16, [512]> encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192578432)))];
+            tensor<fp16, [1, 8, 512]> key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = tensor<string, []>("key_31_cast_fp16")];
+            tensor<bool, []> input_819_interleave_0 = const()[name = tensor<string, []>("input_819_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 78, 512]> input_819_cast_fp16 = concat(axis = var_65, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = tensor<string, []>("input_819_cast_fp16")];
+            tensor<int32, [3]> var_3564_begin_0 = const()[name = tensor<string, []>("op_3564_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
+            tensor<int32, [3]> var_3564_end_0 = const()[name = tensor<string, []>("op_3564_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_3564_end_mask_0 = const()[name = tensor<string, []>("op_3564_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 66, 512]> var_3564_cast_fp16 = slice_by_index(begin = var_3564_begin_0, end = var_3564_end_0, end_mask = var_3564_end_mask_0, x = cache_61_cast_fp16)[name = tensor<string, []>("op_3564_cast_fp16")];
+            tensor<int32, [3]> var_3567_begin_0 = const()[name = tensor<string, []>("op_3567_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_3567_end_0 = const()[name = tensor<string, []>("op_3567_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
+            tensor<bool, [3]> var_3567_end_mask_0 = const()[name = tensor<string, []>("op_3567_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 4, 512]> var_3567_cast_fp16 = slice_by_index(begin = var_3567_begin_0, end = var_3567_end_0, end_mask = var_3567_end_mask_0, x = key_31_cast_fp16)[name = tensor<string, []>("op_3567_cast_fp16")];
+            tensor<bool, []> var_3570_interleave_0 = const()[name = tensor<string, []>("op_3570_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> var_3570_cast_fp16 = concat(axis = var_65, interleave = var_3570_interleave_0, values = (var_3564_cast_fp16, var_3567_cast_fp16))[name = tensor<string, []>("op_3570_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_15_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192579520)))];
+            tensor<fp16, [1, 8, 512]> linear_138_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16, x = key_31_cast_fp16)[name = tensor<string, []>("linear_138_cast_fp16")];
+            tensor<int32, [4]> var_3574 = const()[name = tensor<string, []>("op_3574"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 8, 8, 64]> q_91_cast_fp16 = reshape(shape = var_3574, x = linear_138_cast_fp16)[name = tensor<string, []>("q_91_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_15_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193103872)))];
+            tensor<fp16, [1, 78, 512]> linear_139_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16, x = input_819_cast_fp16)[name = tensor<string, []>("linear_139_cast_fp16")];
+            tensor<int32, [4]> var_3578 = const()[name = tensor<string, []>("op_3578"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> k_61_cast_fp16 = reshape(shape = var_3578, x = linear_139_cast_fp16)[name = tensor<string, []>("k_61_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_15_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193628224)))];
+            tensor<fp16, [1, 78, 512]> linear_140_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16, x = input_819_cast_fp16)[name = tensor<string, []>("linear_140_cast_fp16")];
+            tensor<int32, [4]> var_3582 = const()[name = tensor<string, []>("op_3582"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> v_31_cast_fp16 = reshape(shape = var_3582, x = linear_140_cast_fp16)[name = tensor<string, []>("v_31_cast_fp16")];
+            tensor<int32, [4]> value_33_perm_0 = const()[name = tensor<string, []>("value_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194152576)))];
+            tensor<fp16, [1, 8, 8, 64]> var_3594_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3594_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194153664)))];
+            tensor<fp16, [1, 8, 8, 64]> var_3596_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3596_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_31_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_397_transpose_x_0 = const()[name = tensor<string, []>("x_397_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_397_transpose_y_0 = const()[name = tensor<string, []>("x_397_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 155]> var_3598_to_fp16 = const()[name = tensor<string, []>("op_3598_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194154752)))];
+            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3596_cast_fp16)[name = tensor<string, []>("transpose_102")];
+            tensor<fp16, [1, 8, 8, 155]> x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = var_3598_to_fp16)[name = tensor<string, []>("x_397_cast_fp16")];
+            tensor<int32, [8]> x_399_pad_0 = const()[name = tensor<string, []>("x_399_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_399_mode_0 = const()[name = tensor<string, []>("x_399_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_274_to_fp16 = const()[name = tensor<string, []>("const_274_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 8, 156]> x_399_cast_fp16 = pad(constant_val = const_274_to_fp16, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397_cast_fp16)[name = tensor<string, []>("x_399_cast_fp16")];
+            tensor<int32, [4]> var_3606 = const()[name = tensor<string, []>("op_3606"), val = tensor<int32, [4]>([1, 8, -1, 8])];
+            tensor<fp16, [1, 8, 156, 8]> x_401_cast_fp16 = reshape(shape = var_3606, x = x_399_cast_fp16)[name = tensor<string, []>("x_401_cast_fp16")];
+            tensor<int32, [4]> var_3610_begin_0 = const()[name = tensor<string, []>("op_3610_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3610_end_0 = const()[name = tensor<string, []>("op_3610_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
+            tensor<bool, [4]> var_3610_end_mask_0 = const()[name = tensor<string, []>("op_3610_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 155, 8]> var_3610_cast_fp16 = slice_by_index(begin = var_3610_begin_0, end = var_3610_end_0, end_mask = var_3610_end_mask_0, x = x_401_cast_fp16)[name = tensor<string, []>("op_3610_cast_fp16")];
+            tensor<int32, [4]> var_3611 = const()[name = tensor<string, []>("op_3611"), val = tensor<int32, [4]>([1, 8, 8, 155])];
+            tensor<fp16, [1, 8, 8, 155]> matrix_bd_61_cast_fp16 = reshape(shape = var_3611, x = var_3610_cast_fp16)[name = tensor<string, []>("matrix_bd_61_cast_fp16")];
+            tensor<bool, []> matrix_ac_31_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_31_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_31_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_31_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_81_perm_0 = const()[name = tensor<string, []>("transpose_81_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_82_perm_0 = const()[name = tensor<string, []>("transpose_82_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 78]> transpose_82 = transpose(perm = transpose_82_perm_0, x = k_61_cast_fp16)[name = tensor<string, []>("transpose_100")];
+            tensor<fp16, [1, 8, 8, 64]> transpose_81 = transpose(perm = transpose_81_perm_0, x = var_3594_cast_fp16)[name = tensor<string, []>("transpose_101")];
+            tensor<fp16, [1, 8, 8, 78]> matrix_ac_31_cast_fp16 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_81, y = transpose_82)[name = tensor<string, []>("matrix_ac_31_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_63_begin_0 = const()[name = tensor<string, []>("matrix_bd_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_63_end_0 = const()[name = tensor<string, []>("matrix_bd_63_end_0"), val = tensor<int32, [4]>([1, 8, 8, 78])];
+            tensor<bool, [4]> matrix_bd_63_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 8, 78]> matrix_bd_63_cast_fp16 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61_cast_fp16)[name = tensor<string, []>("matrix_bd_63_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_3620_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = tensor<string, []>("op_3620_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_61_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_61_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 8, 78]> _inversed_scores_61_cast_fp16 = mul(x = var_3620_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_61_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> scores_63_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_63_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_3626_cast_fp16 = softmax(axis = var_56, x = scores_63_cast_fp16)[name = tensor<string, []>("op_3626_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> input_821_cast_fp16 = select(a = var_40_to_fp16, b = var_3626_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_821_cast_fp16")];
+            tensor<bool, []> x_403_transpose_x_0 = const()[name = tensor<string, []>("x_403_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_403_transpose_y_0 = const()[name = tensor<string, []>("x_403_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 78, 64]> value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_31_cast_fp16)[name = tensor<string, []>("transpose_103")];
+            tensor<fp16, [1, 8, 8, 64]> x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821_cast_fp16, y = value_33_cast_fp16)[name = tensor<string, []>("x_403_cast_fp16")];
+            tensor<int32, [4]> var_3630_perm_0 = const()[name = tensor<string, []>("op_3630_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3631 = const()[name = tensor<string, []>("op_3631"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 8, 8, 64]> var_3630_cast_fp16 = transpose(perm = var_3630_perm_0, x = x_403_cast_fp16)[name = tensor<string, []>("transpose_99")];
+            tensor<fp16, [1, 8, 512]> input_823_cast_fp16 = reshape(shape = var_3631, x = var_3630_cast_fp16)[name = tensor<string, []>("input_823_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_15_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194313536)))];
+            tensor<fp16, [1, 8, 512]> linear_142_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16, x = input_823_cast_fp16)[name = tensor<string, []>("linear_142_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = tensor<string, []>("input_827_cast_fp16")];
+            tensor<int32, [1]> x_407_axes_0 = const()[name = tensor<string, []>("x_407_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194837888)))];
+            tensor<fp16, [512]> encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194838976)))];
+            tensor<fp16, [1, 8, 512]> x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = tensor<string, []>("x_407_cast_fp16")];
+            tensor<int32, [3]> input_829_perm_0 = const()[name = tensor<string, []>("input_829_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_831_pad_type_0 = const()[name = tensor<string, []>("input_831_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_831_strides_0 = const()[name = tensor<string, []>("input_831_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_831_pad_0 = const()[name = tensor<string, []>("input_831_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_831_dilations_0 = const()[name = tensor<string, []>("input_831_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_831_groups_0 = const()[name = tensor<string, []>("input_831_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_15_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194840064)))];
+            tensor<fp16, [1, 512, 8]> input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = tensor<string, []>("transpose_98")];
+            tensor<fp16, [1, 1024, 8]> input_831_cast_fp16 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16, x = input_829_cast_fp16)[name = tensor<string, []>("input_831_cast_fp16")];
+            tensor<int32, []> x_409_split_num_splits_0 = const()[name = tensor<string, []>("x_409_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_409_split_axis_0 = const()[name = tensor<string, []>("x_409_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 8]> x_409_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> x_409_split_cast_fp16_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831_cast_fp16)[name = tensor<string, []>("x_409_split_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = tensor<string, []>("x_409_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_409_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> input_833_cast_fp16 = select(a = var_40_to_fp16, b = x_409_cast_fp16, cond = var_551)[name = tensor<string, []>("input_833_cast_fp16")];
+            tensor<bool, []> new_x_63_interleave_0 = const()[name = tensor<string, []>("new_x_63_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 16]> new_x_63_cast_fp16 = concat(axis = var_56, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = tensor<string, []>("new_x_63_cast_fp16")];
+            tensor<int32, [3]> next_cache_31_begin_0 = const()[name = tensor<string, []>("next_cache_31_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_31_end_0 = const()[name = tensor<string, []>("next_cache_31_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> next_cache_31_end_mask_0 = const()[name = tensor<string, []>("next_cache_31_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 12]> next_cache_31_cast_fp16 = slice_by_index(begin = next_cache_31_begin_0, end = next_cache_31_end_0, end_mask = next_cache_31_end_mask_0, x = new_x_63_cast_fp16)[name = tensor<string, []>("next_cache_31_cast_fp16")];
+            tensor<int32, [3]> var_3672_begin_0 = const()[name = tensor<string, []>("op_3672_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
+            tensor<int32, [3]> var_3672_end_0 = const()[name = tensor<string, []>("op_3672_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> var_3672_end_mask_0 = const()[name = tensor<string, []>("op_3672_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> var_3672_cast_fp16 = slice_by_index(begin = var_3672_begin_0, end = var_3672_end_0, end_mask = var_3672_end_mask_0, x = next_cache_31_cast_fp16)[name = tensor<string, []>("op_3672_cast_fp16")];
+            tensor<string, []> x_411_pad_type_0 = const()[name = tensor<string, []>("x_411_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_411_groups_0 = const()[name = tensor<string, []>("x_411_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_411_strides_0 = const()[name = tensor<string, []>("x_411_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_411_pad_0 = const()[name = tensor<string, []>("x_411_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_411_dilations_0 = const()[name = tensor<string, []>("x_411_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_15_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195888704)))];
+            tensor<fp16, [1, 512, 8]> x_411_cast_fp16 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16, x = new_x_63_cast_fp16)[name = tensor<string, []>("x_411_cast_fp16")];
+            tensor<int32, [3]> input_835_perm_0 = const()[name = tensor<string, []>("input_835_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_413_axes_0 = const()[name = tensor<string, []>("x_413_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195897984)))];
+            tensor<fp16, [512]> encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195899072)))];
+            tensor<fp16, [1, 8, 512]> input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = tensor<string, []>("transpose_97")];
+            tensor<fp16, [1, 8, 512]> x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = tensor<string, []>("x_413_cast_fp16")];
+            tensor<int32, [3]> input_837_perm_0 = const()[name = tensor<string, []>("input_837_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 8]> input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = tensor<string, []>("transpose_96")];
+            tensor<fp16, [1, 512, 8]> input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = tensor<string, []>("input_839_cast_fp16")];
+            tensor<string, []> x_415_pad_type_0 = const()[name = tensor<string, []>("x_415_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_415_strides_0 = const()[name = tensor<string, []>("x_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_415_pad_0 = const()[name = tensor<string, []>("x_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_415_dilations_0 = const()[name = tensor<string, []>("x_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_415_groups_0 = const()[name = tensor<string, []>("x_415_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_15_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195900160)))];
+            tensor<fp16, [1, 512, 8]> x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16, x = input_839_cast_fp16)[name = tensor<string, []>("x_415_cast_fp16")];
+            tensor<int32, [3]> input_841_perm_0 = const()[name = tensor<string, []>("input_841_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 8, 512]> input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = tensor<string, []>("transpose_95")];
+            tensor<fp16, [1, 8, 512]> input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = tensor<string, []>("input_843_cast_fp16")];
+            tensor<int32, [1]> input_845_axes_0 = const()[name = tensor<string, []>("input_845_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196424512)))];
+            tensor<fp16, [512]> encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196425600)))];
+            tensor<fp16, [1, 8, 512]> input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = tensor<string, []>("input_845_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_15_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196426688)))];
+            tensor<fp16, [1, 8, 2048]> linear_143_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16, x = input_845_cast_fp16)[name = tensor<string, []>("linear_143_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = tensor<string, []>("input_849_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_15_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198523904)))];
+            tensor<fp16, [1, 8, 512]> linear_144_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16, x = input_849_cast_fp16)[name = tensor<string, []>("linear_144_cast_fp16")];
+            tensor<fp16, []> var_3713_to_fp16 = const()[name = tensor<string, []>("op_3713_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_3714_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3713_to_fp16)[name = tensor<string, []>("op_3714_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3714_cast_fp16)[name = tensor<string, []>("input_855_cast_fp16")];
+            tensor<int32, [1]> input_857_axes_0 = const()[name = tensor<string, []>("input_857_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_15_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200621120)))];
+            tensor<fp16, [512]> encoder_layers_15_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200622208)))];
+            tensor<fp16, [1, 8, 512]> input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = tensor<string, []>("input_857_cast_fp16")];
+            tensor<int32, [4]> cache_65_begin_0 = const()[name = tensor<string, []>("cache_65_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_65_end_0 = const()[name = tensor<string, []>("cache_65_end_0"), val = tensor<int32, [4]>([17, 1, 70, 512])];
+            tensor<bool, [4]> cache_65_end_mask_0 = const()[name = tensor<string, []>("cache_65_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_65_squeeze_mask_0 = const()[name = tensor<string, []>("cache_65_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 70, 512]> cache_65_cast_fp16 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = tensor<string, []>("cache_65_cast_fp16")];
+            tensor<int32, [4]> cache_begin_0 = const()[name = tensor<string, []>("cache_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_end_0 = const()[name = tensor<string, []>("cache_end_0"), val = tensor<int32, [4]>([17, 1, 512, 8])];
+            tensor<bool, [4]> cache_end_mask_0 = const()[name = tensor<string, []>("cache_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_squeeze_mask_0 = const()[name = tensor<string, []>("cache_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 512, 8]> cache_cast_fp16 = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = cache_last_time_to_fp16)[name = tensor<string, []>("cache_cast_fp16")];
+            tensor<int32, [1]> input_859_axes_0 = const()[name = tensor<string, []>("input_859_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200623296)))];
+            tensor<fp16, [512]> encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200624384)))];
+            tensor<fp16, [1, 8, 512]> input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = tensor<string, []>("input_859_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_16_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200625472)))];
+            tensor<fp16, [1, 8, 2048]> linear_145_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16, x = input_859_cast_fp16)[name = tensor<string, []>("linear_145_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = tensor<string, []>("input_863_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_16_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202722688)))];
+            tensor<fp16, [1, 8, 512]> linear_146_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16, x = input_863_cast_fp16)[name = tensor<string, []>("linear_146_cast_fp16")];
+            tensor<fp16, []> var_3748_to_fp16 = const()[name = tensor<string, []>("op_3748_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_3749_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3748_to_fp16)[name = tensor<string, []>("op_3749_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3749_cast_fp16)[name = tensor<string, []>("input_869_cast_fp16")];
+            tensor<int32, [1]> key_axes_0 = const()[name = tensor<string, []>("key_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(204819904)))];
+            tensor<fp16, [512]> encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(204820992)))];
+            tensor<fp16, [1, 8, 512]> key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = tensor<string, []>("key_cast_fp16")];
+            tensor<bool, []> input_871_interleave_0 = const()[name = tensor<string, []>("input_871_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 78, 512]> input_871_cast_fp16 = concat(axis = var_65, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_cast_fp16))[name = tensor<string, []>("input_871_cast_fp16")];
+            tensor<int32, [3]> var_3771_begin_0 = const()[name = tensor<string, []>("op_3771_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
+            tensor<int32, [3]> var_3771_end_0 = const()[name = tensor<string, []>("op_3771_end_0"), val = tensor<int32, [3]>([1, 70, 512])];
+            tensor<bool, [3]> var_3771_end_mask_0 = const()[name = tensor<string, []>("op_3771_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 66, 512]> var_3771_cast_fp16 = slice_by_index(begin = var_3771_begin_0, end = var_3771_end_0, end_mask = var_3771_end_mask_0, x = cache_65_cast_fp16)[name = tensor<string, []>("op_3771_cast_fp16")];
+            tensor<int32, [3]> var_3774_begin_0 = const()[name = tensor<string, []>("op_3774_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_3774_end_0 = const()[name = tensor<string, []>("op_3774_end_0"), val = tensor<int32, [3]>([1, 4, 512])];
+            tensor<bool, [3]> var_3774_end_mask_0 = const()[name = tensor<string, []>("op_3774_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 4, 512]> var_3774_cast_fp16 = slice_by_index(begin = var_3774_begin_0, end = var_3774_end_0, end_mask = var_3774_end_mask_0, x = key_cast_fp16)[name = tensor<string, []>("op_3774_cast_fp16")];
+            tensor<bool, []> cache_last_channel_cur_interleave_0 = const()[name = tensor<string, []>("cache_last_channel_cur_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 70, 512]> cache_last_channel_cur_cast_fp16 = concat(axis = var_65, interleave = cache_last_channel_cur_interleave_0, values = (var_3771_cast_fp16, var_3774_cast_fp16))[name = tensor<string, []>("cache_last_channel_cur_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_16_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(204822080)))];
+            tensor<fp16, [1, 8, 512]> linear_147_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16, x = key_cast_fp16)[name = tensor<string, []>("linear_147_cast_fp16")];
+            tensor<int32, [4]> var_3781 = const()[name = tensor<string, []>("op_3781"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 8, 8, 64]> q_97_cast_fp16 = reshape(shape = var_3781, x = linear_147_cast_fp16)[name = tensor<string, []>("q_97_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_16_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205346432)))];
+            tensor<fp16, [1, 78, 512]> linear_148_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16, x = input_871_cast_fp16)[name = tensor<string, []>("linear_148_cast_fp16")];
+            tensor<int32, [4]> var_3785 = const()[name = tensor<string, []>("op_3785"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> k_65_cast_fp16 = reshape(shape = var_3785, x = linear_148_cast_fp16)[name = tensor<string, []>("k_65_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_16_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205870784)))];
+            tensor<fp16, [1, 78, 512]> linear_149_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16, x = input_871_cast_fp16)[name = tensor<string, []>("linear_149_cast_fp16")];
+            tensor<int32, [4]> var_3789 = const()[name = tensor<string, []>("op_3789"), val = tensor<int32, [4]>([1, -1, 8, 64])];
+            tensor<fp16, [1, 78, 8, 64]> v_cast_fp16 = reshape(shape = var_3789, x = linear_149_cast_fp16)[name = tensor<string, []>("v_cast_fp16")];
+            tensor<int32, [4]> value_perm_0 = const()[name = tensor<string, []>("value_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [8, 64]> encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206395136)))];
+            tensor<fp16, [1, 8, 8, 64]> var_3801_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3801_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206396224)))];
+            tensor<fp16, [1, 8, 8, 64]> var_3803_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3803_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> x_423_transpose_x_0 = const()[name = tensor<string, []>("x_423_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_423_transpose_y_0 = const()[name = tensor<string, []>("x_423_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 64, 155]> var_3805_to_fp16 = const()[name = tensor<string, []>("op_3805_to_fp16"), val = tensor<fp16, [1, 8, 64, 155]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206397312)))];
+            tensor<fp16, [1, 8, 8, 64]> q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_3803_cast_fp16)[name = tensor<string, []>("transpose_93")];
+            tensor<fp16, [1, 8, 8, 155]> x_423_cast_fp16 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_cast_fp16, y = var_3805_to_fp16)[name = tensor<string, []>("x_423_cast_fp16")];
+            tensor<int32, [8]> x_425_pad_0 = const()[name = tensor<string, []>("x_425_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_425_mode_0 = const()[name = tensor<string, []>("x_425_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_287_to_fp16 = const()[name = tensor<string, []>("const_287_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 8, 156]> x_425_cast_fp16 = pad(constant_val = const_287_to_fp16, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423_cast_fp16)[name = tensor<string, []>("x_425_cast_fp16")];
+            tensor<int32, [4]> var_3813 = const()[name = tensor<string, []>("op_3813"), val = tensor<int32, [4]>([1, 8, -1, 8])];
+            tensor<fp16, [1, 8, 156, 8]> x_427_cast_fp16 = reshape(shape = var_3813, x = x_425_cast_fp16)[name = tensor<string, []>("x_427_cast_fp16")];
+            tensor<int32, [4]> var_3817_begin_0 = const()[name = tensor<string, []>("op_3817_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3817_end_0 = const()[name = tensor<string, []>("op_3817_end_0"), val = tensor<int32, [4]>([1, 8, 156, 8])];
+            tensor<bool, [4]> var_3817_end_mask_0 = const()[name = tensor<string, []>("op_3817_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 155, 8]> var_3817_cast_fp16 = slice_by_index(begin = var_3817_begin_0, end = var_3817_end_0, end_mask = var_3817_end_mask_0, x = x_427_cast_fp16)[name = tensor<string, []>("op_3817_cast_fp16")];
+            tensor<int32, [4]> var_3818 = const()[name = tensor<string, []>("op_3818"), val = tensor<int32, [4]>([1, 8, 8, 155])];
+            tensor<fp16, [1, 8, 8, 155]> matrix_bd_65_cast_fp16 = reshape(shape = var_3818, x = var_3817_cast_fp16)[name = tensor<string, []>("matrix_bd_65_cast_fp16")];
+            tensor<bool, []> matrix_ac_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_83_perm_0 = const()[name = tensor<string, []>("transpose_83_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_84_perm_0 = const()[name = tensor<string, []>("transpose_84_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 78]> transpose_84 = transpose(perm = transpose_84_perm_0, x = k_65_cast_fp16)[name = tensor<string, []>("transpose_91")];
+            tensor<fp16, [1, 8, 8, 64]> transpose_83 = transpose(perm = transpose_83_perm_0, x = var_3801_cast_fp16)[name = tensor<string, []>("transpose_92")];
+            tensor<fp16, [1, 8, 8, 78]> matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_83, y = transpose_84)[name = tensor<string, []>("matrix_ac_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_begin_0 = const()[name = tensor<string, []>("matrix_bd_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_end_0 = const()[name = tensor<string, []>("matrix_bd_end_0"), val = tensor<int32, [4]>([1, 8, 8, 78])];
+            tensor<bool, [4]> matrix_bd_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 8, 78]> matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_65_cast_fp16)[name = tensor<string, []>("matrix_bd_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_3827_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = tensor<string, []>("op_3827_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_65_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_65_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 8, 78]> _inversed_scores_65_cast_fp16 = mul(x = var_3827_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_65_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> scores_cast_fp16 = select(a = var_41_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> var_3833_cast_fp16 = softmax(axis = var_56, x = scores_cast_fp16)[name = tensor<string, []>("op_3833_cast_fp16")];
+            tensor<fp16, [1, 8, 8, 78]> input_873_cast_fp16 = select(a = var_40_to_fp16, b = var_3833_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_873_cast_fp16")];
+            tensor<bool, []> x_429_transpose_x_0 = const()[name = tensor<string, []>("x_429_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_429_transpose_y_0 = const()[name = tensor<string, []>("x_429_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 78, 64]> value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = tensor<string, []>("transpose_94")];
+            tensor<fp16, [1, 8, 8, 64]> x_429_cast_fp16 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873_cast_fp16, y = value_cast_fp16)[name = tensor<string, []>("x_429_cast_fp16")];
+            tensor<int32, [4]> var_3837_perm_0 = const()[name = tensor<string, []>("op_3837_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3838 = const()[name = tensor<string, []>("op_3838"), val = tensor<int32, [3]>([1, -1, 512])];
+            tensor<fp16, [1, 8, 8, 64]> var_3837_cast_fp16 = transpose(perm = var_3837_perm_0, x = x_429_cast_fp16)[name = tensor<string, []>("transpose_90")];
+            tensor<fp16, [1, 8, 512]> input_875_cast_fp16 = reshape(shape = var_3838, x = var_3837_cast_fp16)[name = tensor<string, []>("input_875_cast_fp16")];
+            tensor<fp16, [512, 512]> encoder_layers_16_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206556096)))];
+            tensor<fp16, [1, 8, 512]> linear_151_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16, x = input_875_cast_fp16)[name = tensor<string, []>("linear_151_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = tensor<string, []>("input_879_cast_fp16")];
+            tensor<int32, [1]> x_433_axes_0 = const()[name = tensor<string, []>("x_433_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207080448)))];
+            tensor<fp16, [512]> encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207081536)))];
+            tensor<fp16, [1, 8, 512]> x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = tensor<string, []>("x_433_cast_fp16")];
+            tensor<int32, [3]> input_881_perm_0 = const()[name = tensor<string, []>("input_881_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_883_pad_type_0 = const()[name = tensor<string, []>("input_883_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_883_strides_0 = const()[name = tensor<string, []>("input_883_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_883_pad_0 = const()[name = tensor<string, []>("input_883_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_883_dilations_0 = const()[name = tensor<string, []>("input_883_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_883_groups_0 = const()[name = tensor<string, []>("input_883_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 512, 1]> encoder_layers_16_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207082624)))];
+            tensor<fp16, [1, 512, 8]> input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = tensor<string, []>("transpose_89")];
+            tensor<fp16, [1, 1024, 8]> input_883_cast_fp16 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16, x = input_881_cast_fp16)[name = tensor<string, []>("input_883_cast_fp16")];
+            tensor<int32, []> x_435_split_num_splits_0 = const()[name = tensor<string, []>("x_435_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_435_split_axis_0 = const()[name = tensor<string, []>("x_435_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 8]> x_435_split_cast_fp16_0, tensor<fp16, [1, 512, 8]> x_435_split_cast_fp16_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883_cast_fp16)[name = tensor<string, []>("x_435_split_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = tensor<string, []>("x_435_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_435_cast_fp16")];
+            tensor<fp16, [1, 512, 8]> input_885_cast_fp16 = select(a = var_40_to_fp16, b = x_435_cast_fp16, cond = var_551)[name = tensor<string, []>("input_885_cast_fp16")];
+            tensor<bool, []> new_x_interleave_0 = const()[name = tensor<string, []>("new_x_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512, 16]> new_x_cast_fp16 = concat(axis = var_56, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_885_cast_fp16))[name = tensor<string, []>("new_x_cast_fp16")];
+            tensor<int32, [3]> next_cache_begin_0 = const()[name = tensor<string, []>("next_cache_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> next_cache_end_0 = const()[name = tensor<string, []>("next_cache_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> next_cache_end_mask_0 = const()[name = tensor<string, []>("next_cache_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 12]> next_cache_cast_fp16 = slice_by_index(begin = next_cache_begin_0, end = next_cache_end_0, end_mask = next_cache_end_mask_0, x = new_x_cast_fp16)[name = tensor<string, []>("next_cache_cast_fp16")];
+            tensor<int32, [3]> cache_last_time_cur_begin_0 = const()[name = tensor<string, []>("cache_last_time_cur_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
+            tensor<int32, [3]> cache_last_time_cur_end_0 = const()[name = tensor<string, []>("cache_last_time_cur_end_0"), val = tensor<int32, [3]>([1, 512, 12])];
+            tensor<bool, [3]> cache_last_time_cur_end_mask_0 = const()[name = tensor<string, []>("cache_last_time_cur_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 8]> cache_last_time_cur_cast_fp16 = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = next_cache_cast_fp16)[name = tensor<string, []>("cache_last_time_cur_cast_fp16")];
+            tensor<string, []> x_437_pad_type_0 = const()[name = tensor<string, []>("x_437_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_437_groups_0 = const()[name = tensor<string, []>("x_437_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> x_437_strides_0 = const()[name = tensor<string, []>("x_437_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_437_pad_0 = const()[name = tensor<string, []>("x_437_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_437_dilations_0 = const()[name = tensor<string, []>("x_437_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 9]> encoder_layers_16_conv_depthwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_depthwise_conv_weight_to_fp16"), val = tensor<fp16, [512, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208131264)))];
+            tensor<fp16, [1, 512, 8]> x_437_cast_fp16 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16, x = new_x_cast_fp16)[name = tensor<string, []>("x_437_cast_fp16")];
+            tensor<int32, [3]> input_887_perm_0 = const()[name = tensor<string, []>("input_887_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_439_axes_0 = const()[name = tensor<string, []>("x_439_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208140544)))];
+            tensor<fp16, [512]> encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208141632)))];
+            tensor<fp16, [1, 8, 512]> input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = tensor<string, []>("transpose_88")];
+            tensor<fp16, [1, 8, 512]> x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = tensor<string, []>("x_439_cast_fp16")];
+            tensor<int32, [3]> input_889_perm_0 = const()[name = tensor<string, []>("input_889_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 512, 8]> input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = tensor<string, []>("transpose_87")];
+            tensor<fp16, [1, 512, 8]> input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = tensor<string, []>("input_891_cast_fp16")];
+            tensor<string, []> x_441_pad_type_0 = const()[name = tensor<string, []>("x_441_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_441_strides_0 = const()[name = tensor<string, []>("x_441_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_441_pad_0 = const()[name = tensor<string, []>("x_441_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_441_dilations_0 = const()[name = tensor<string, []>("x_441_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_441_groups_0 = const()[name = tensor<string, []>("x_441_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 1]> encoder_layers_16_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208142720)))];
+            tensor<fp16, [1, 512, 8]> x_441_cast_fp16 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16, x = input_891_cast_fp16)[name = tensor<string, []>("x_441_cast_fp16")];
+            tensor<int32, [3]> input_893_perm_0 = const()[name = tensor<string, []>("input_893_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 8, 512]> input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = tensor<string, []>("transpose_86")];
+            tensor<fp16, [1, 8, 512]> input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = tensor<string, []>("input_895_cast_fp16")];
+            tensor<int32, [1]> input_897_axes_0 = const()[name = tensor<string, []>("input_897_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208667072)))];
+            tensor<fp16, [512]> encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208668160)))];
+            tensor<fp16, [1, 8, 512]> input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = tensor<string, []>("input_897_cast_fp16")];
+            tensor<fp16, [2048, 512]> encoder_layers_16_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208669248)))];
+            tensor<fp16, [1, 8, 2048]> linear_152_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16, x = input_897_cast_fp16)[name = tensor<string, []>("linear_152_cast_fp16")];
+            tensor<fp16, [1, 8, 2048]> input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = tensor<string, []>("input_901_cast_fp16")];
+            tensor<fp16, [512, 2048]> encoder_layers_16_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210766464)))];
+            tensor<fp16, [1, 8, 512]> linear_153_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16, x = input_901_cast_fp16)[name = tensor<string, []>("linear_153_cast_fp16")];
+            tensor<fp16, []> var_3920_to_fp16 = const()[name = tensor<string, []>("op_3920_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 8, 512]> var_3921_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_3920_to_fp16)[name = tensor<string, []>("op_3921_cast_fp16")];
+            tensor<fp16, [1, 8, 512]> input_cast_fp16 = add(x = input_895_cast_fp16, y = var_3921_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [1]> audio_signal_axes_0 = const()[name = tensor<string, []>("audio_signal_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> encoder_layers_16_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212863680)))];
+            tensor<fp16, [512]> encoder_layers_16_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212864768)))];
+            tensor<fp16, [1, 8, 512]> audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_38_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("audio_signal_cast_fp16")];
+            tensor<int32, [3]> obj_1_perm_0 = const()[name = tensor<string, []>("obj_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> obj_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_1_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, []> obj_5_axis_0 = const()[name = tensor<string, []>("obj_5_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [17, 1, 70, 512]> obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_465_cast_fp16, var_672_cast_fp16, var_879_cast_fp16, var_1086_cast_fp16, var_1293_cast_fp16, var_1500_cast_fp16, var_1707_cast_fp16, var_1914_cast_fp16, var_2121_cast_fp16, var_2328_cast_fp16, var_2535_cast_fp16, var_2742_cast_fp16, var_2949_cast_fp16, var_3156_cast_fp16, var_3363_cast_fp16, var_3570_cast_fp16, cache_last_channel_cur_cast_fp16))[name = tensor<string, []>("obj_5_cast_fp16")];
+            tensor<int32, []> obj_7_axis_0 = const()[name = tensor<string, []>("obj_7_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [17, 1, 512, 8]> obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_567_cast_fp16, var_774_cast_fp16, var_981_cast_fp16, var_1188_cast_fp16, var_1395_cast_fp16, var_1602_cast_fp16, var_1809_cast_fp16, var_2016_cast_fp16, var_2223_cast_fp16, var_2430_cast_fp16, var_2637_cast_fp16, var_2844_cast_fp16, var_3051_cast_fp16, var_3258_cast_fp16, var_3465_cast_fp16, var_3672_cast_fp16, cache_last_time_cur_cast_fp16))[name = tensor<string, []>("obj_7_cast_fp16")];
+            tensor<string, []> obj_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_7_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> var_3937 = add(x = cache_last_channel_len, y = cache_keep_size)[name = tensor<string, []>("op_3937")];
+            tensor<string, []> var_3937_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3937_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> const_293_to_fp16 = const()[name = tensor<string, []>("const_293_to_fp16"), val = tensor<fp16, []>(-inf)];
+            tensor<fp16, []> var_46_promoted_to_fp16 = const()[name = tensor<string, []>("op_46_promoted_to_fp16"), val = tensor<fp16, []>(0x1.18p+6)];
+            tensor<fp16, [1]> var_3937_to_fp16 = cast(dtype = var_3937_promoted_to_fp16_dtype_0, x = var_3937)[name = tensor<string, []>("cast_184")];
+            tensor<fp16, [1]> clip_1_cast_fp16 = clip(alpha = const_293_to_fp16, beta = var_46_promoted_to_fp16, x = var_3937_to_fp16)[name = tensor<string, []>("clip_1_cast_fp16")];
+            tensor<string, []> var_3964_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3964_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<string, []> cast_179_dtype_0 = const()[name = tensor<string, []>("cast_179_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<string, []> cast_180_dtype_0 = const()[name = tensor<string, []>("cast_180_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1]> new_cache_last_channel_len = cast(dtype = cast_180_dtype_0, x = clip_1_cast_fp16)[name = tensor<string, []>("cast_181")];
+            tensor<int32, [1]> encoded_length = cast(dtype = cast_179_dtype_0, x = clip_0_cast_fp16)[name = tensor<string, []>("cast_182")];
+            tensor<fp32, [17, 1, 70, 512]> new_cache_last_channel = cast(dtype = var_3964_cast_fp16_to_fp32_dtype_0, x = obj_5_cast_fp16)[name = tensor<string, []>("cast_183")];
+            tensor<fp32, [17, 1, 512, 8]> new_cache_last_time = cast(dtype = obj_7_cast_fp16_to_fp32_dtype_0, x = obj_7_cast_fp16)[name = tensor<string, []>("cast_185")];
+            tensor<fp16, [1, 512, 8]> obj_1_cast_fp16 = transpose(perm = obj_1_perm_0, x = audio_signal_cast_fp16)[name = tensor<string, []>("transpose_85")];
+            tensor<fp32, [1, 512, 8]> encoded_output = cast(dtype = obj_1_cast_fp16_to_fp32_dtype_0, x = obj_1_cast_fp16)[name = tensor<string, []>("cast_186")];
+            tensor<fp32, [1, 128, 9]> new_pre_cache = cast(dtype = var_28_cast_fp16_to_fp32_dtype_0, x = var_28_cast_fp16)[name = tensor<string, []>("cast_198")];
+        } -> (encoded_output, encoded_length, new_pre_cache, new_cache_last_channel, new_cache_last_time, new_cache_last_channel_len);
+}
\ No newline at end of file
diff --git a/320ms/streaming_encoder.mlmodelc/weights/weight.bin b/320ms/streaming_encoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..15654dd70b3be49baf2409d7d99786b2bbfd35da
--- /dev/null
+++ b/320ms/streaming_encoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff553fcb4f4560cfce1e558e16bb9d856e4b4f6a4c08f800d5c126fce4196822
+size 212865856
diff --git a/320ms/streaming_encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/320ms/streaming_encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..da8d55a342f8863cbab69d42b663e39fc14294aa
--- /dev/null
+++ b/320ms/streaming_encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2579ec91db1c064c91973e33c01a226d1243eaae3ae204ccc81951ad350b2698
+size 539070
diff --git a/320ms/streaming_encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/320ms/streaming_encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..15654dd70b3be49baf2409d7d99786b2bbfd35da
--- /dev/null
+++ b/320ms/streaming_encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff553fcb4f4560cfce1e558e16bb9d856e4b4f6a4c08f800d5c126fce4196822
+size 212865856
diff --git a/320ms/streaming_encoder.mlpackage/Manifest.json b/320ms/streaming_encoder.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..365b912e230f6f5a69bcb16eda6dd90e0f3d59e4
--- /dev/null
+++ b/320ms/streaming_encoder.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "446DB82B-CCCF-45A7-8BC1-6A0B513ADDAA": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "56C7CF0D-51B6-4E0A-ACAE-6A71D089B1B7": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "446DB82B-CCCF-45A7-8BC1-6A0B513ADDAA"
+}
diff --git a/320ms/streaming_encoder_metadata.json b/320ms/streaming_encoder_metadata.json
new file mode 100644
index 0000000000000000000000000000000000000000..87b2d329d7e571863559397e2c855250dc5019b0
--- /dev/null
+++ b/320ms/streaming_encoder_metadata.json
@@ -0,0 +1,12 @@
+{
+  "model_id": "nvidia/parakeet_realtime_eou_120m-v1",
+  "chunk_ms": 320,
+  "mel_frames": 64,
+  "pre_cache_size": 9,
+  "valid_out_len": 4,
+  "samples_per_chunk": 10240,
+  "hidden_dim": 512,
+  "num_layers": 17,
+  "cache_channel_size": 70,
+  "cache_time_size": 8
+}
\ No newline at end of file
diff --git a/320ms/vocab.json b/320ms/vocab.json
new file mode 100644
index 0000000000000000000000000000000000000000..462081190dbb331a0dde13e99d5cfa99b302f0a7
--- /dev/null
+++ b/320ms/vocab.json
@@ -0,0 +1,1028 @@
+{
+  "0": "<unk>",
+  "1": "▁t",
+  "2": "▁th",
+  "3": "▁a",
+  "4": "▁i",
+  "5": "▁the",
+  "6": "▁s",
+  "7": "re",
+  "8": "▁w",
+  "9": "▁o",
+  "10": "in",
+  "11": "at",
+  "12": "er",
+  "13": "nd",
+  "14": "ou",
+  "15": "▁c",
+  "16": "▁b",
+  "17": "▁h",
+  "18": "en",
+  "19": "on",
+  "20": "▁m",
+  "21": "▁f",
+  "22": "ing",
+  "23": "▁p",
+  "24": "▁to",
+  "25": "▁and",
+  "26": "▁d",
+  "27": "an",
+  "28": "or",
+  "29": "es",
+  "30": "▁y",
+  "31": "▁l",
+  "32": "▁of",
+  "33": "ll",
+  "34": "▁in",
+  "35": "ed",
+  "36": "it",
+  "37": "▁g",
+  "38": "is",
+  "39": "▁you",
+  "40": "▁n",
+  "41": "ar",
+  "42": "om",
+  "43": "as",
+  "44": "ve",
+  "45": "▁e",
+  "46": "ic",
+  "47": "▁it",
+  "48": "al",
+  "49": "us",
+  "50": "▁wh",
+  "51": "▁we",
+  "52": "▁be",
+  "53": "ion",
+  "54": "ow",
+  "55": "le",
+  "56": "▁is",
+  "57": "et",
+  "58": "ent",
+  "59": "ot",
+  "60": "ut",
+  "61": "▁re",
+  "62": "▁on",
+  "63": "ay",
+  "64": "▁ha",
+  "65": "ig",
+  "66": "▁so",
+  "67": "ct",
+  "68": "▁he",
+  "69": "▁for",
+  "70": "ver",
+  "71": "ke",
+  "72": "ro",
+  "73": "▁st",
+  "74": "id",
+  "75": "▁go",
+  "76": "all",
+  "77": "se",
+  "78": "ly",
+  "79": "▁u",
+  "80": "ch",
+  "81": "st",
+  "82": "ld",
+  "83": "▁k",
+  "84": "ce",
+  "85": "ur",
+  "86": "▁li",
+  "87": "am",
+  "88": "▁r",
+  "89": "ht",
+  "90": "▁j",
+  "91": "ith",
+  "92": "▁se",
+  "93": "ir",
+  "94": "▁as",
+  "95": "▁an",
+  "96": "im",
+  "97": "▁do",
+  "98": "ad",
+  "99": "▁was",
+  "100": "ight",
+  "101": "th",
+  "102": "▁are",
+  "103": "▁but",
+  "104": "▁sh",
+  "105": "ust",
+  "106": "ally",
+  "107": "▁not",
+  "108": "▁or",
+  "109": "▁com",
+  "110": "▁can",
+  "111": "▁me",
+  "112": "op",
+  "113": "▁mo",
+  "114": "▁at",
+  "115": "ill",
+  "116": "▁ch",
+  "117": "▁ne",
+  "118": "ant",
+  "119": "▁de",
+  "120": "▁kn",
+  "121": "▁one",
+  "122": "il",
+  "123": "ol",
+  "124": "▁con",
+  "125": "ter",
+  "126": "▁ab",
+  "127": "▁fr",
+  "128": "ere",
+  "129": "ck",
+  "130": "▁al",
+  "131": "▁all",
+  "132": "qu",
+  "133": "▁pro",
+  "134": "▁som",
+  "135": "ould",
+  "136": "▁tw",
+  "137": "ul",
+  "138": "ra",
+  "139": "od",
+  "140": "ers",
+  "141": "▁su",
+  "142": "ive",
+  "143": "▁v",
+  "144": "use",
+  "145": "ate",
+  "146": "ge",
+  "147": "if",
+  "148": "▁ex",
+  "149": "ess",
+  "150": "pp",
+  "151": "▁lo",
+  "152": "out",
+  "153": "▁if",
+  "154": "est",
+  "155": "ain",
+  "156": "ist",
+  "157": "and",
+  "158": "ea",
+  "159": "very",
+  "160": "art",
+  "161": "▁wor",
+  "162": "▁my",
+  "163": "ab",
+  "164": "ment",
+  "165": "▁bec",
+  "166": "un",
+  "167": "ity",
+  "168": "ri",
+  "169": "pe",
+  "170": "ions",
+  "171": "▁by",
+  "172": "ok",
+  "173": "our",
+  "174": "ort",
+  "175": "ind",
+  "176": "ink",
+  "177": "nt",
+  "178": "▁up",
+  "179": "um",
+  "180": "▁don",
+  "181": "▁get",
+  "182": "red",
+  "183": "▁out",
+  "184": "el",
+  "185": "ause",
+  "186": "res",
+  "187": "▁ma",
+  "188": "ich",
+  "189": "▁us",
+  "190": "rou",
+  "191": "▁int",
+  "192": "em",
+  "193": "os",
+  "194": "ies",
+  "195": "ie",
+  "196": "▁pl",
+  "197": "▁tr",
+  "198": "ven",
+  "199": "ous",
+  "200": "▁le",
+  "201": "▁two",
+  "202": "ard",
+  "203": "ine",
+  "204": "▁co",
+  "205": "een",
+  "206": "▁now",
+  "207": "ty",
+  "208": "her",
+  "209": "ack",
+  "210": "▁pe",
+  "211": "ame",
+  "212": "▁how",
+  "213": "▁who",
+  "214": "▁see",
+  "215": "▁tim",
+  "216": "ect",
+  "217": "ast",
+  "218": "▁our",
+  "219": "ci",
+  "220": "ree",
+  "221": "ople",
+  "222": "gh",
+  "223": "▁no",
+  "224": "▁had",
+  "225": "▁man",
+  "226": "▁qu",
+  "227": "▁en",
+  "228": "ide",
+  "229": "ure",
+  "230": "ud",
+  "231": "so",
+  "232": "▁his",
+  "233": "▁sa",
+  "234": "▁sp",
+  "235": "▁say",
+  "236": "ose",
+  "237": "ther",
+  "238": "▁act",
+  "239": "▁ta",
+  "240": "▁cl",
+  "241": "ings",
+  "242": "pt",
+  "243": "king",
+  "244": "▁any",
+  "245": "▁has",
+  "246": "▁un",
+  "247": "iv",
+  "248": "▁im",
+  "249": "▁ag",
+  "250": "▁te",
+  "251": "▁fe",
+  "252": "one",
+  "253": "per",
+  "254": "ong",
+  "255": "▁po",
+  "256": "▁ad",
+  "257": "ff",
+  "258": "ore",
+  "259": "itt",
+  "260": "ans",
+  "261": "iz",
+  "262": "eah",
+  "263": "reat",
+  "264": "act",
+  "265": "own",
+  "266": "hing",
+  "267": "enty",
+  "268": "age",
+  "269": "ber",
+  "270": "ice",
+  "271": "▁am",
+  "272": "ple",
+  "273": "are",
+  "274": "▁per",
+  "275": "und",
+  "276": "ite",
+  "277": "ix",
+  "278": "pl",
+  "279": "▁way",
+  "280": "▁did",
+  "281": "▁pr",
+  "282": "▁got",
+  "283": "ars",
+  "284": "▁she",
+  "285": "▁let",
+  "286": "ag",
+  "287": "▁ac",
+  "288": "int",
+  "289": "▁ar",
+  "290": "ry",
+  "291": "ign",
+  "292": "ish",
+  "293": "▁fir",
+  "294": "ace",
+  "295": "ble",
+  "296": "og",
+  "297": "ue",
+  "298": "▁ye",
+  "299": "ap",
+  "300": "iff",
+  "301": "▁ro",
+  "302": "▁her",
+  "303": "nder",
+  "304": "▁ok",
+  "305": "▁res",
+  "306": "▁gu",
+  "307": "ence",
+  "308": "▁may",
+  "309": "ated",
+  "310": "ip",
+  "311": "▁bo",
+  "312": "▁him",
+  "313": "way",
+  "314": "ac",
+  "315": "ical",
+  "316": "ass",
+  "317": "ase",
+  "318": "▁dis",
+  "319": "able",
+  "320": "ick",
+  "321": "▁app",
+  "322": "ance",
+  "323": "▁pre",
+  "324": "▁six",
+  "325": "▁off",
+  "326": "▁new",
+  "327": "ia",
+  "328": "orm",
+  "329": "ank",
+  "330": "▁lot",
+  "331": "ach",
+  "332": "▁fo",
+  "333": "inet",
+  "334": "ire",
+  "335": "ary",
+  "336": "ult",
+  "337": "▁tal",
+  "338": "▁mu",
+  "339": "▁bl",
+  "340": "ount",
+  "341": "sel",
+  "342": "vel",
+  "343": "▁br",
+  "344": "▁imp",
+  "345": "ep",
+  "346": "cess",
+  "347": "ord",
+  "348": "▁sc",
+  "349": "▁inc",
+  "350": "ound",
+  "351": "ang",
+  "352": "be",
+  "353": "ress",
+  "354": "uct",
+  "355": "▁ind",
+  "356": "▁af",
+  "357": "ving",
+  "358": "▁oh",
+  "359": "▁bet",
+  "360": "▁use",
+  "361": "ome",
+  "362": "ens",
+  "363": "ys",
+  "364": "▁bu",
+  "365": "co",
+  "366": "ory",
+  "367": "ater",
+  "368": "ild",
+  "369": "ght",
+  "370": "ial",
+  "371": "▁day",
+  "372": "ning",
+  "373": "na",
+  "374": "ile",
+  "375": "▁spe",
+  "376": "▁mar",
+  "377": "ody",
+  "378": "ough",
+  "379": "ade",
+  "380": "vers",
+  "381": "xt",
+  "382": "▁fl",
+  "383": "▁ke",
+  "384": "ian",
+  "385": "▁sy",
+  "386": "▁put",
+  "387": "fore",
+  "388": "ub",
+  "389": "▁ph",
+  "390": "fe",
+  "391": "▁em",
+  "392": "▁ser",
+  "393": "form",
+  "394": "ting",
+  "395": "te",
+  "396": "av",
+  "397": "ious",
+  "398": "▁rec",
+  "399": "ks",
+  "400": "▁gr",
+  "401": "ces",
+  "402": "wn",
+  "403": "ors",
+  "404": "▁jo",
+  "405": "ents",
+  "406": "▁des",
+  "407": "▁try",
+  "408": "▁equ",
+  "409": "▁z",
+  "410": "▁rem",
+  "411": "▁str",
+  "412": "self",
+  "413": "▁bit",
+  "414": "ph",
+  "415": "ved",
+  "416": "▁why",
+  "417": "▁bas",
+  "418": "▁hel",
+  "419": "▁rel",
+  "420": "ath",
+  "421": "ject",
+  "422": "ail",
+  "423": "▁la",
+  "424": "ual",
+  "425": "▁god",
+  "426": "▁nat",
+  "427": "erm",
+  "428": "day",
+  "429": "▁id",
+  "430": "ft",
+  "431": "▁wr",
+  "432": "▁min",
+  "433": "ates",
+  "434": "▁gen",
+  "435": "tain",
+  "436": "▁ob",
+  "437": "ull",
+  "438": "ict",
+  "439": "▁tra",
+  "440": "▁end",
+  "441": "▁hig",
+  "442": "▁fif",
+  "443": "oth",
+  "444": "tern",
+  "445": "▁its",
+  "446": "vent",
+  "447": "▁sm",
+  "448": "ons",
+  "449": "▁add",
+  "450": "iss",
+  "451": "▁bel",
+  "452": "ful",
+  "453": "get",
+  "454": "▁ele",
+  "455": "▁rep",
+  "456": "ak",
+  "457": "▁ho",
+  "458": "▁pos",
+  "459": "▁num",
+  "460": "ange",
+  "461": "ves",
+  "462": "ific",
+  "463": "urn",
+  "464": "ise",
+  "465": "▁cr",
+  "466": "▁um",
+  "467": "ward",
+  "468": "▁reg",
+  "469": "ady",
+  "470": "ower",
+  "471": "uc",
+  "472": "▁dec",
+  "473": "lic",
+  "474": "▁set",
+  "475": "▁gon",
+  "476": "▁op",
+  "477": "▁ear",
+  "478": "▁sub",
+  "479": "▁sl",
+  "480": "les",
+  "481": "stem",
+  "482": "cial",
+  "483": "olog",
+  "484": "atch",
+  "485": "ily",
+  "486": "body",
+  "487": "nds",
+  "488": "ular",
+  "489": "ren",
+  "490": "▁own",
+  "491": "▁too",
+  "492": "cent",
+  "493": "ible",
+  "494": "pect",
+  "495": "ered",
+  "496": "ways",
+  "497": "teen",
+  "498": "▁uh",
+  "499": "▁big",
+  "500": "▁mod",
+  "501": "▁att",
+  "502": "▁car",
+  "503": "gr",
+  "504": "▁acc",
+  "505": "ied",
+  "506": "mun",
+  "507": "ib",
+  "508": "▁mon",
+  "509": "▁sch",
+  "510": "▁pol",
+  "511": "▁dat",
+  "512": "▁fin",
+  "513": "▁sim",
+  "514": "▁inv",
+  "515": "▁def",
+  "516": "ked",
+  "517": "▁ent",
+  "518": "▁yes",
+  "519": "ows",
+  "520": "ics",
+  "521": "ited",
+  "522": "ute",
+  "523": "ism",
+  "524": "ps",
+  "525": "▁ed",
+  "526": "▁el",
+  "527": "ably",
+  "528": "ppen",
+  "529": "als",
+  "530": "▁ten",
+  "531": "ract",
+  "532": "ss",
+  "533": "▁ass",
+  "534": "▁met",
+  "535": "gan",
+  "536": "▁eng",
+  "537": "▁stu",
+  "538": "ween",
+  "539": "arch",
+  "540": "▁gl",
+  "541": "▁cor",
+  "542": "▁dr",
+  "543": "vern",
+  "544": "▁ty",
+  "545": "▁run",
+  "546": "hip",
+  "547": "cus",
+  "548": "cond",
+  "549": "▁ins",
+  "550": "irty",
+  "551": "▁pub",
+  "552": "lud",
+  "553": "llow",
+  "554": "▁cou",
+  "555": "ew",
+  "556": "iew",
+  "557": "▁sur",
+  "558": "ero",
+  "559": "ood",
+  "560": "ness",
+  "561": "▁fun",
+  "562": "▁eff",
+  "563": "cept",
+  "564": "▁ca",
+  "565": "▁exp",
+  "566": "duct",
+  "567": "▁sw",
+  "568": "ize",
+  "569": "ope",
+  "570": "▁par",
+  "571": "kes",
+  "572": "cy",
+  "573": "▁ev",
+  "574": "▁ref",
+  "575": "ell",
+  "576": "▁bus",
+  "577": "ug",
+  "578": "rib",
+  "579": "▁cur",
+  "580": "mo",
+  "581": "ock",
+  "582": "ures",
+  "583": "air",
+  "584": "▁war",
+  "585": "str",
+  "586": "▁med",
+  "587": "▁wa",
+  "588": "▁val",
+  "589": "▁sin",
+  "590": "blem",
+  "591": "▁fam",
+  "592": "li",
+  "593": "▁far",
+  "594": "▁cle",
+  "595": "▁col",
+  "596": "mon",
+  "597": "▁gra",
+  "598": "led",
+  "599": "ense",
+  "600": "tin",
+  "601": "ues",
+  "602": "its",
+  "603": "▁mem",
+  "604": "▁inf",
+  "605": "▁eas",
+  "606": "ideo",
+  "607": "▁top",
+  "608": "io",
+  "609": "pan",
+  "610": "▁hum",
+  "611": "▁old",
+  "612": "ead",
+  "613": "▁ord",
+  "614": "ric",
+  "615": "ants",
+  "616": "oy",
+  "617": "esn",
+  "618": "uck",
+  "619": "ason",
+  "620": "ced",
+  "621": "ool",
+  "622": "rat",
+  "623": "ouse",
+  "624": "▁lar",
+  "625": "▁art",
+  "626": "▁wee",
+  "627": "▁cer",
+  "628": "ized",
+  "629": "▁mat",
+  "630": "con",
+  "631": "erg",
+  "632": "land",
+  "633": "ines",
+  "634": "▁chr",
+  "635": "▁aut",
+  "636": "▁lea",
+  "637": "▁sou",
+  "638": "oney",
+  "639": "tty",
+  "640": "▁ple",
+  "641": "ulat",
+  "642": "oks",
+  "643": "▁few",
+  "644": "▁sol",
+  "645": "▁che",
+  "646": "chn",
+  "647": "ird",
+  "648": "▁bre",
+  "649": "▁dur",
+  "650": "▁wom",
+  "651": "me",
+  "652": "izat",
+  "653": "eric",
+  "654": "ote",
+  "655": "▁uni",
+  "656": "eren",
+  "657": "arn",
+  "658": "ross",
+  "659": "ices",
+  "660": "ten",
+  "661": "eral",
+  "662": "ever",
+  "663": "ieve",
+  "664": "lish",
+  "665": "ash",
+  "666": "▁opp",
+  "667": "alth",
+  "668": "ger",
+  "669": "▁sk",
+  "670": "▁red",
+  "671": "peri",
+  "672": "▁det",
+  "673": "▁ext",
+  "674": "ner",
+  "675": "ah",
+  "676": "▁var",
+  "677": "▁loc",
+  "678": "gram",
+  "679": "ists",
+  "680": "ives",
+  "681": "▁es",
+  "682": "▁nor",
+  "683": "tro",
+  "684": "ale",
+  "685": "▁iss",
+  "686": "▁pri",
+  "687": "gin",
+  "688": "az",
+  "689": "oc",
+  "690": "▁pop",
+  "691": "ern",
+  "692": "▁sit",
+  "693": "ket",
+  "694": "▁pa",
+  "695": "▁law",
+  "696": "ages",
+  "697": "br",
+  "698": "▁cam",
+  "699": "▁mom",
+  "700": "osed",
+  "701": "▁bro",
+  "702": "ne",
+  "703": "bs",
+  "704": "▁cre",
+  "705": "erat",
+  "706": "▁sec",
+  "707": "▁cap",
+  "708": "▁vis",
+  "709": "▁pat",
+  "710": "ield",
+  "711": "iet",
+  "712": "▁tri",
+  "713": "up",
+  "714": "▁bra",
+  "715": "ts",
+  "716": "▁mot",
+  "717": "▁unt",
+  "718": "put",
+  "719": "bo",
+  "720": "ork",
+  "721": "mer",
+  "722": "ital",
+  "723": "▁air",
+  "724": "ined",
+  "725": "▁beh",
+  "726": "▁adv",
+  "727": "▁ret",
+  "728": "imes",
+  "729": "▁tea",
+  "730": "ural",
+  "731": "sid",
+  "732": "ters",
+  "733": "▁pur",
+  "734": "▁sci",
+  "735": "bers",
+  "736": "ient",
+  "737": "ier",
+  "738": "cc",
+  "739": "sw",
+  "740": "▁av",
+  "741": "reen",
+  "742": "ode",
+  "743": "ont",
+  "744": "▁dra",
+  "745": "ann",
+  "746": "nect",
+  "747": "▁x",
+  "748": "▁eu",
+  "749": "ton",
+  "750": "inat",
+  "751": "ene",
+  "752": "ared",
+  "753": "els",
+  "754": "▁mor",
+  "755": "▁rat",
+  "756": "cri",
+  "757": "▁men",
+  "758": "▁ah",
+  "759": "ames",
+  "760": "▁arm",
+  "761": "eak",
+  "762": "▁pay",
+  "763": "▁hal",
+  "764": "ins",
+  "765": "ilit",
+  "766": "stit",
+  "767": "▁ra",
+  "768": "▁leg",
+  "769": "cl",
+  "770": "pr",
+  "771": "▁wal",
+  "772": "▁bad",
+  "773": "▁ge",
+  "774": "roup",
+  "775": "▁mus",
+  "776": "man",
+  "777": "▁gi",
+  "778": "eds",
+  "779": "▁aw",
+  "780": "po",
+  "781": "ark",
+  "782": "row",
+  "783": "▁dep",
+  "784": "ully",
+  "785": "ral",
+  "786": "lect",
+  "787": "pend",
+  "788": "▁sev",
+  "789": "ime",
+  "790": "gest",
+  "791": "here",
+  "792": "▁yet",
+  "793": "ted",
+  "794": "▁rev",
+  "795": "ds",
+  "796": "▁ask",
+  "797": "less",
+  "798": "▁di",
+  "799": "ets",
+  "800": "line",
+  "801": "▁aff",
+  "802": "ired",
+  "803": "▁est",
+  "804": "ken",
+  "805": "vid",
+  "806": "most",
+  "807": "ivid",
+  "808": "unch",
+  "809": "par",
+  "810": "med",
+  "811": "rop",
+  "812": "ased",
+  "813": "eone",
+  "814": "▁ve",
+  "815": "▁abs",
+  "816": "ergy",
+  "817": "ret",
+  "818": "▁saw",
+  "819": "▁ey",
+  "820": "▁cal",
+  "821": "uat",
+  "822": "▁mid",
+  "823": "vat",
+  "824": "ream",
+  "825": "vice",
+  "826": "ians",
+  "827": "rent",
+  "828": "ctor",
+  "829": "err",
+  "830": "ush",
+  "831": "ases",
+  "832": "▁suc",
+  "833": "erms",
+  "834": "ave",
+  "835": "angu",
+  "836": "ries",
+  "837": "▁wo",
+  "838": "arts",
+  "839": "▁fil",
+  "840": "▁fat",
+  "841": "▁cho",
+  "842": "orts",
+  "843": "▁fre",
+  "844": "ee",
+  "845": "ught",
+  "846": "eng",
+  "847": "ump",
+  "848": "▁bar",
+  "849": "ying",
+  "850": "ane",
+  "851": "▁tem",
+  "852": "anks",
+  "853": "ury",
+  "854": "iat",
+  "855": "mit",
+  "856": "trol",
+  "857": "▁net",
+  "858": "▁maj",
+  "859": "▁cra",
+  "860": "ling",
+  "861": "▁fig",
+  "862": "orn",
+  "863": "icat",
+  "864": "pany",
+  "865": "▁occ",
+  "866": "ott",
+  "867": "ands",
+  "868": "▁exc",
+  "869": "▁mr",
+  "870": "ency",
+  "871": "rope",
+  "872": "itch",
+  "873": "▁lit",
+  "874": "abil",
+  "875": "not",
+  "876": "ma",
+  "877": "▁typ",
+  "878": "▁opt",
+  "879": "ob",
+  "880": "ser",
+  "881": "ety",
+  "882": "ms",
+  "883": "peci",
+  "884": "aces",
+  "885": "aut",
+  "886": "▁hon",
+  "887": "cuss",
+  "888": "▁sal",
+  "889": "▁sor",
+  "890": "att",
+  "891": "▁lab",
+  "892": "▁har",
+  "893": "urch",
+  "894": "nded",
+  "895": "uce",
+  "896": "ids",
+  "897": "▁hy",
+  "898": "▁fut",
+  "899": "▁ste",
+  "900": "ours",
+  "901": "ems",
+  "902": "utes",
+  "903": "ng",
+  "904": "ta",
+  "905": "▁won",
+  "906": "▁fa",
+  "907": "▁env",
+  "908": "ards",
+  "909": "▁job",
+  "910": "ium",
+  "911": "▁dot",
+  "912": "▁obv",
+  "913": "ina",
+  "914": "side",
+  "915": "elve",
+  "916": "cu",
+  "917": "▁jes",
+  "918": "▁pot",
+  "919": "▁pie",
+  "920": "▁tre",
+  "921": "▁hey",
+  "922": "▁mag",
+  "923": "ron",
+  "924": "▁key",
+  "925": "swer",
+  "926": "▁win",
+  "927": "ucat",
+  "928": "work",
+  "929": "ides",
+  "930": "▁low",
+  "931": "▁vol",
+  "932": "▁oth",
+  "933": "atic",
+  "934": "lf",
+  "935": "ads",
+  "936": "inds",
+  "937": "com",
+  "938": "ths",
+  "939": "▁ver",
+  "940": "ised",
+  "941": "lo",
+  "942": "▁squ",
+  "943": "▁cut",
+  "944": "oked",
+  "945": "irit",
+  "946": "ateg",
+  "947": "ppy",
+  "948": "mitt",
+  "949": "come",
+  "950": "hn",
+  "951": "igin",
+  "952": "mand",
+  "953": "▁dam",
+  "954": "ho",
+  "955": "▁da",
+  "956": "▁fur",
+  "957": "iron",
+  "958": "ilar",
+  "959": "▁fac",
+  "960": "▁neg",
+  "961": "▁ago",
+  "962": "ged",
+  "963": "miss",
+  "964": "enth",
+  "965": "▁dou",
+  "966": "▁hit",
+  "967": "▁guy",
+  "968": "▁bi",
+  "969": "ove",
+  "970": "fess",
+  "971": "ples",
+  "972": "owed",
+  "973": "ured",
+  "974": "▁ris",
+  "975": "ints",
+  "976": "rew",
+  "977": "▁sum",
+  "978": "▁hu",
+  "979": "ploy",
+  "980": "ude",
+  "981": "ried",
+  "982": "▁cir",
+  "983": "▁dev",
+  "984": "ear",
+  "985": "▁tot",
+  "986": "▁ann",
+  "987": "duc",
+  "988": "ik",
+  "989": "pon",
+  "990": "sted",
+  "991": "▁ide",
+  "992": "▁'",
+  "993": "ipp",
+  "994": "▁eat",
+  "995": "▁dom",
+  "996": "▁",
+  "997": "e",
+  "998": "t",
+  "999": "o",
+  "1000": "a",
+  "1001": "i",
+  "1002": "n",
+  "1003": "s",
+  "1004": "r",
+  "1005": "h",
+  "1006": "l",
+  "1007": "d",
+  "1008": "u",
+  "1009": "c",
+  "1010": "m",
+  "1011": "y",
+  "1012": "g",
+  "1013": "w",
+  "1014": "f",
+  "1015": "p",
+  "1016": "b",
+  "1017": "v",
+  "1018": "k",
+  "1019": "'",
+  "1020": "j",
+  "1021": "x",
+  "1022": "q",
+  "1023": "z",
+  "1024": "<EOU>",
+  "1025": "<EOB>"
+}
\ No newline at end of file
diff --git a/README.md b/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..c581f781ee4e6d4379c815b0aecb326ffabd858d
--- /dev/null
+++ b/README.md
@@ -0,0 +1,117 @@
+---
+license: other
+license_name: nvidia-open-model-license
+license_link: >-
+  https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/
+language:
+- en
+metrics:
+- wer
+library_name: nemo
+tags:
+- speech-recognition
+- FastConformer
+- end-of-utterance
+- voice agent
+pipeline_tag: automatic-speech-recognition
+base_model:
+- nvidia/parakeet_realtime_eou_120m-v1
+base_model_relation: finetune
+---
+
+
+# Parakeet Realtime EOU 120M — CoreML
+
+CoreML conversion of [nvidia/parakeet-realtime-eou-120m-v1](https://huggingface.co/nvidia/parakee
+t-realtime-eou-120m-v1) for streaming speech recognition with end-of-utterance detection on Apple
+ Silicon.
+
+Used by [FluidAudio](https://github.com/FluidInference/FluidAudio) for real-time transcription.
+
+## Models
+
+The RNNT pipeline is split into three CoreML models, exported at two chunk sizes:
+
+| Model | Description |
+|-------|-------------|
+| `streaming_encoder.mlmodelc` | FastConformer encoder with loopback state caching |
+| `decoder.mlmodelc` | 1-layer LSTM decoder (640 hidden units) |
+| `joint_decision.mlmodelc` | Joint network for token prediction + EOU detection |
+
+### Chunk Size Variants
+
+| Variant | Latency | WER (test-clean) | RTFx (M2) |
+|---------|---------|-------------------|------------|
+| `160ms/` | 160ms | 8.29% | 4.78x |
+| `320ms/` | 320ms | 4.87% | 12.48x |
+
+Benchmarked on LibriSpeech test-clean (2620 files, 5.40h audio) on Apple M2.
+
+## Usage with FluidAudio
+
+```swift
+import FluidAudio
+
+let manager = StreamingEouAsrManager()
+await manager.initialize()
+
+// Transcribe with EOU detection
+await manager.startStreaming(
+    eouCallback: { transcript in
+        print("Utterance complete: \(transcript)")
+    },
+    partialCallback: { partial in
+        print("Partial: \(partial)")
+    }
+)
+
+// Feed audio chunks as they arrive
+await manager.feedAudio(samples)
+```
+
+CLI
+
+# Transcribe a file
+swift run fluidaudio parakeet-eou --input audio.wav
+
+# Benchmark
+swift run -c release fluidaudio parakeet-eou --benchmark --chunk-size 320
+
+Architecture
+
+120M parameter RNNT (Recurrent Neural Network Transducer) with:
+- Encoder: 17-layer FastConformer with cache-aware streaming
+- Decoder: 1-layer LSTM, 640 hidden size
+- Joint: Linear projection with 1027 output classes (1024 tokens + EOU token + SOS + blank)
+- EOU token: ID 1024 signals end-of-utterance
+
+Streaming State
+
+The encoder maintains loopback state between chunks:
+┌─────────────────────┬──────────────────┬───────────────────────┐
+│        State        │      Shape       │      Description      │
+├─────────────────────┼──────────────────┼───────────────────────┤
+│ preCache            │ [1, 128, N]      │ Mel-level context     │
+├─────────────────────┼──────────────────┼───────────────────────┤
+│ cacheLastChannel    │ [17, 1, 70, 512] │ Conformer layer cache │
+├─────────────────────┼──────────────────┼───────────────────────┤
+│ cacheLastTime       │ [17, 1, 512, 8]  │ Temporal cache        │
+├─────────────────────┼──────────────────┼───────────────────────┤
+│ cacheLastChannelLen │ [1]              │ Cache length tracking │
+└─────────────────────┴──────────────────┴───────────────────────┘
+Export
+
+Converted from PyTorch using coremltools. To re-export:
+
+python3 Scripts/ParakeetEOU/Conversion/convert_split_encoder.py \
+  --output-dir Models/ParakeetEOU \
+  --model-id nvidia/parakeet-realtime-eou-120m-v1
+
+License
+
+NVIDIA Open Model License — see
+https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/.
+
+Original model: https://huggingface.co/nvidia/parakeet-realtime-eou-120m-v1
+
+---
\ No newline at end of file
diff --git a/config.json b/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..9e26dfeeb6e641a33dae4961196235bdb965b21b
--- /dev/null
+++ b/config.json
@@ -0,0 +1 @@
+{}
\ No newline at end of file
diff --git a/convert_parakeet_eou.py b/convert_parakeet_eou.py
new file mode 100644
index 0000000000000000000000000000000000000000..9c865f90c8132840b5755dcd62fcfac15e3f0352
--- /dev/null
+++ b/convert_parakeet_eou.py
@@ -0,0 +1,740 @@
+#!/usr/bin/env python3
+"""CLI for exporting Parakeet Realtime EOU 120M components to CoreML.
+
+This model is a cache-aware streaming FastConformer-RNNT model optimized for
+low-latency speech recognition with end-of-utterance detection.
+
+Key differences from Parakeet TDT v3:
+- Smaller model (120M vs 600M params)
+- No duration outputs (standard RNNT, not TDT)
+- Cache-aware streaming encoder (17 layers, attention context [70,1])
+- Special <EOU> token for end-of-utterance detection
+- Optimized for 80-160ms latency
+
+Reference: https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1
+"""
+from __future__ import annotations
+
+import json
+from dataclasses import asdict
+from pathlib import Path
+from typing import Dict, Optional, Tuple
+
+import coremltools as ct
+import numpy as np
+import soundfile as sf
+import torch
+import typer
+
+import nemo.collections.asr as nemo_asr
+
+from individual_components import (
+    DecoderWrapper,
+    EncoderWrapper,
+    ExportSettings,
+    JointWrapper,
+    JointDecisionWrapper,
+    JointDecisionSingleStep,
+    PreprocessorWrapper,
+    MelEncoderWrapper,
+    _coreml_convert,
+)
+
+def apply_stft_patch():
+    # Monkey patch coremltools.stft to handle extra arguments from newer torch versions
+    try:
+        import coremltools.converters.mil.frontend.torch.ops as torch_ops
+        _original_stft = torch_ops.stft
+
+        def patched_stft(context, node):
+            if len(node.inputs) > 8:
+                node.inputs = node.inputs[:8]
+            return _original_stft(context, node)
+
+        torch_ops.stft = patched_stft
+        if "stft" in torch_ops._TORCH_OPS_REGISTRY:
+            torch_ops._TORCH_OPS_REGISTRY["stft"] = patched_stft
+        print("Monkey patched coremltools.stft for compatibility.")
+    except Exception as e:
+        print(f"Warning: Could not monkey patch stft: {e}")
+
+DEFAULT_MODEL_ID = "nvidia/parakeet_realtime_eou_120m-v1"
+AUTHOR = "Fluid Inference"
+
+
+def _compute_length(seconds: float, sample_rate: int) -> int:
+    return int(round(seconds * sample_rate))
+
+
+def _prepare_audio(
+    validation_audio: Optional[Path],
+    sample_rate: int,
+    max_samples: int,
+    seed: Optional[int],
+) -> torch.Tensor:
+    if validation_audio is None:
+        if seed is not None:
+            torch.manual_seed(seed)
+        audio = torch.randn(1, max_samples, dtype=torch.float32)
+        return audio
+
+    data, sr = sf.read(str(validation_audio), dtype="float32")
+    if sr != sample_rate:
+        raise typer.BadParameter(
+            f"Validation audio sample rate {sr} does not match model rate {sample_rate}"
+        )
+
+    if data.ndim > 1:
+        data = data[:, 0]
+
+    if data.size == 0:
+        raise typer.BadParameter("Validation audio is empty")
+
+    if data.size < max_samples:
+        pad_width = max_samples - data.size
+        data = np.pad(data, (0, pad_width))
+    elif data.size > max_samples:
+        data = data[:max_samples]
+
+    audio = torch.from_numpy(data).unsqueeze(0).to(dtype=torch.float32)
+    return audio
+
+
+def _save_mlpackage(model: ct.models.MLModel, path: Path, description: str) -> None:
+    try:
+        model.minimum_deployment_target = ct.target.iOS17
+    except Exception:
+        pass
+    model.short_description = description
+    model.author = AUTHOR
+    path.parent.mkdir(parents=True, exist_ok=True)
+    model.save(str(path))
+
+
+def _tensor_shape(tensor: torch.Tensor) -> Tuple[int, ...]:
+    return tuple(int(dim) for dim in tensor.shape)
+
+
+def _parse_compute_units(name: str) -> ct.ComputeUnit:
+    """Parse a human-friendly compute units string into ct.ComputeUnit."""
+    normalized = str(name).strip().upper()
+    mapping = {
+        "ALL": ct.ComputeUnit.ALL,
+        "CPU_ONLY": ct.ComputeUnit.CPU_ONLY,
+        "CPU_AND_GPU": ct.ComputeUnit.CPU_AND_GPU,
+        "CPU_AND_NE": ct.ComputeUnit.CPU_AND_NE,
+        "CPU_AND_NEURALENGINE": ct.ComputeUnit.CPU_AND_NE,
+    }
+    if normalized not in mapping:
+        raise typer.BadParameter(
+            f"Unknown compute units '{name}'. Choose from: " + ", ".join(mapping.keys())
+        )
+    return mapping[normalized]
+
+
+def _parse_compute_precision(name: Optional[str]) -> Optional[ct.precision]:
+    """Parse compute precision string into ct.precision or None."""
+    if name is None:
+        return None
+    normalized = str(name).strip().upper()
+    if normalized == "":
+        return None
+    mapping = {
+        "FLOAT32": ct.precision.FLOAT32,
+        "FLOAT16": ct.precision.FLOAT16,
+    }
+    if normalized not in mapping:
+        raise typer.BadParameter(
+            f"Unknown compute precision '{name}'. Choose from: "
+            + ", ".join(mapping.keys())
+        )
+    return mapping[normalized]
+
+
+app = typer.Typer(add_completion=False, pretty_exceptions_show_locals=False)
+
+
+@app.command()
+def convert(
+    nemo_path: Optional[Path] = typer.Option(
+        None,
+        "--nemo-path",
+        exists=True,
+        resolve_path=True,
+        help="Path to parakeet_realtime_eou_120m-v1.nemo checkpoint (skip to auto-download)",
+    ),
+    model_id: str = typer.Option(
+        DEFAULT_MODEL_ID,
+        "--model-id",
+        help="Model identifier to download when --nemo-path is omitted",
+    ),
+    output_dir: Path = typer.Option(
+        Path("parakeet_eou_coreml"),
+        help="Directory where mlpackages and metadata will be written",
+    ),
+    preprocessor_cu: str = typer.Option(
+        "CPU_ONLY",
+        "--preprocessor-cu",
+        help="Compute units for preprocessor (default CPU_ONLY)",
+    ),
+    mel_encoder_cu: str = typer.Option(
+        "CPU_ONLY",
+        "--mel-encoder-cu",
+        help="Compute units for fused mel+encoder (default CPU_ONLY)",
+    ),
+    compute_precision: Optional[str] = typer.Option(
+        None,
+        "--compute-precision",
+        help="Export precision: FLOAT32 (default) or FLOAT16 to shrink non-quantized weights.",
+    ),
+    max_audio_seconds: float = typer.Option(
+        15.0,
+        "--max-audio-seconds",
+        help="Maximum audio duration in seconds for the fixed window export",
+    ),
+    validation_audio: Optional[Path] = typer.Option(
+        None,
+        "--validation-audio",
+        exists=True,
+        resolve_path=True,
+        help="Path to a 16kHz WAV file for tracing (uses random if not provided)",
+    ),
+) -> None:
+    """Export all Parakeet Realtime EOU sub-modules to CoreML.
+
+    This exports the cache-aware streaming FastConformer-RNNT model for
+    low-latency speech recognition with end-of-utterance detection.
+    """
+    export_settings = ExportSettings(
+        output_dir=output_dir,
+        compute_units=ct.ComputeUnit.CPU_ONLY,
+        deployment_target=ct.target.iOS17,
+        compute_precision=_parse_compute_precision(compute_precision),
+        max_audio_seconds=max_audio_seconds,
+        max_symbol_steps=1,
+    )
+
+    typer.echo("Export configuration:")
+    typer.echo(asdict(export_settings))
+
+    output_dir.mkdir(parents=True, exist_ok=True)
+    pre_cu = _parse_compute_units(preprocessor_cu)
+    melenc_cu = _parse_compute_units(mel_encoder_cu)
+
+    if nemo_path is not None:
+        typer.echo(f"Loading NeMo model from {nemo_path}…")
+        # Try loading as generic ASRModel first, then specific class
+        try:
+            asr_model = nemo_asr.models.ASRModel.restore_from(
+                str(nemo_path), map_location="cpu"
+            )
+        except Exception:
+            # Fallback to EncDecRNNTBPEModel
+            asr_model = nemo_asr.models.EncDecRNNTBPEModel.restore_from(
+                str(nemo_path), map_location="cpu"
+            )
+        checkpoint_meta = {
+            "type": "file",
+            "path": str(nemo_path),
+        }
+    else:
+        typer.echo(f"Downloading NeMo model via {model_id}…")
+        # Use ASRModel.from_pretrained as recommended for this model
+        try:
+            asr_model = nemo_asr.models.ASRModel.from_pretrained(
+                model_id, map_location="cpu"
+            )
+        except Exception:
+            # Fallback to EncDecRNNTBPEModel
+            asr_model = nemo_asr.models.EncDecRNNTBPEModel.from_pretrained(
+                model_id, map_location="cpu"
+            )
+        checkpoint_meta = {
+            "type": "pretrained",
+            "model_id": model_id,
+        }
+    asr_model.eval()
+
+    # Print model info
+    typer.echo(f"Model class: {type(asr_model).__name__}")
+    typer.echo(f"Encoder class: {type(asr_model.encoder).__name__}")
+
+    sample_rate = int(asr_model.cfg.preprocessor.sample_rate)
+    max_samples = _compute_length(export_settings.max_audio_seconds, sample_rate)
+
+    # Prepare audio for tracing
+    if validation_audio is not None:
+        typer.echo(f"Using validation audio: {validation_audio}")
+        audio_tensor = _prepare_audio(validation_audio, sample_rate, max_samples, seed=None)
+    else:
+        typer.echo("Using random audio for tracing (seed=42)")
+        audio_tensor = _prepare_audio(None, sample_rate, max_samples, seed=42)
+
+    audio_length = torch.tensor([max_samples], dtype=torch.int32)
+
+    preprocessor = PreprocessorWrapper(asr_model.preprocessor.eval())
+    encoder = EncoderWrapper(asr_model.encoder.eval())
+    decoder = DecoderWrapper(asr_model.decoder.eval())
+    joint = JointWrapper(asr_model.joint.eval())
+
+    decoder_export_flag = getattr(asr_model.decoder, "_rnnt_export", False)
+    asr_model.decoder._rnnt_export = True
+
+    try:
+        with torch.no_grad():
+            mel_ref, mel_length_ref = preprocessor(audio_tensor, audio_length)
+            mel_length_ref = mel_length_ref.to(dtype=torch.int32)
+            encoder_ref, encoder_length_ref, frame_times_ref = encoder(
+                mel_ref, mel_length_ref
+            )
+            encoder_length_ref = encoder_length_ref.to(dtype=torch.int32)
+
+            # Clone tensors to drop inference flags
+            mel_ref = mel_ref.clone().detach()
+            mel_length_ref = mel_length_ref.clone().detach()
+            encoder_ref = encoder_ref.clone().detach()
+            encoder_length_ref = encoder_length_ref.clone().detach()
+            frame_times_ref = frame_times_ref.clone().detach()
+
+        vocab_size = int(asr_model.tokenizer.vocab_size)
+        decoder_hidden = int(asr_model.decoder.pred_hidden)
+        decoder_layers = int(asr_model.decoder.pred_rnn_layers)
+
+        # Check if model has extra outputs (TDT-style duration)
+        num_extra = getattr(asr_model.joint, "num_extra_outputs", 0)
+        typer.echo(f"Vocab size: {vocab_size}, num_extra_outputs: {num_extra}")
+
+        targets = torch.full(
+            (1, export_settings.max_symbol_steps),
+            fill_value=asr_model.decoder.blank_idx,
+            dtype=torch.int32,
+        )
+        target_lengths = torch.tensor(
+            [export_settings.max_symbol_steps], dtype=torch.int32
+        )
+        zero_state = torch.zeros(
+            decoder_layers,
+            1,
+            decoder_hidden,
+            dtype=torch.float32,
+        )
+
+        with torch.no_grad():
+            decoder_ref, h_ref, c_ref = decoder(
+                targets, target_lengths, zero_state, zero_state
+            )
+            joint_ref = joint(encoder_ref, decoder_ref)
+
+        decoder_ref = decoder_ref.clone()
+        h_ref = h_ref.clone()
+        c_ref = c_ref.clone()
+        joint_ref = joint_ref.clone()
+
+        typer.echo(f"Encoder output shape: {encoder_ref.shape}")
+        typer.echo(f"Decoder output shape: {decoder_ref.shape}")
+        typer.echo(f"Joint output shape: {joint_ref.shape}")
+
+        # === Export Preprocessor ===
+        typer.echo("Tracing and converting preprocessor…")
+        preprocessor = preprocessor.cpu()
+        audio_tensor = audio_tensor.cpu()
+        audio_length = audio_length.cpu()
+        traced_preprocessor = torch.jit.trace(
+            preprocessor, (audio_tensor, audio_length), strict=False
+        )
+        traced_preprocessor.eval()
+        preprocessor_inputs = [
+            ct.TensorType(
+                name="audio_signal",
+                shape=(1, ct.RangeDim(1, max_samples)),
+                dtype=np.float32,
+            ),
+            ct.TensorType(name="audio_length", shape=(1,), dtype=np.int32),
+        ]
+        preprocessor_outputs = [
+            ct.TensorType(name="mel", dtype=np.float32),
+            ct.TensorType(name="mel_length", dtype=np.int32),
+        ]
+        preprocessor_model = _coreml_convert(
+            traced_preprocessor,
+            preprocessor_inputs,
+            preprocessor_outputs,
+            export_settings,
+            compute_units_override=pre_cu,
+        )
+        preprocessor_path = output_dir / "parakeet_eou_preprocessor.mlpackage"
+        _save_mlpackage(
+            preprocessor_model,
+            preprocessor_path,
+            f"Parakeet EOU preprocessor ({max_audio_seconds}s window)",
+        )
+
+        # === Export Encoder ===
+        typer.echo("Tracing and converting encoder…")
+        traced_encoder = torch.jit.trace(
+            encoder, (mel_ref, mel_length_ref), strict=False
+        )
+        traced_encoder.eval()
+        encoder_inputs = [
+            ct.TensorType(
+                name="mel", shape=_tensor_shape(mel_ref), dtype=np.float32
+            ),
+            ct.TensorType(name="mel_length", shape=(1,), dtype=np.int32),
+        ]
+        encoder_outputs = [
+            ct.TensorType(name="encoder", dtype=np.float32),
+            ct.TensorType(name="encoder_length", dtype=np.int32),
+            ct.TensorType(name="frame_times", dtype=np.float32),
+        ]
+        encoder_model = _coreml_convert(
+            traced_encoder,
+            encoder_inputs,
+            encoder_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        encoder_path = output_dir / "parakeet_eou_encoder.mlpackage"
+        _save_mlpackage(
+            encoder_model,
+            encoder_path,
+            f"Parakeet EOU encoder ({max_audio_seconds}s window)",
+        )
+
+        # === Export Fused Mel+Encoder ===
+        typer.echo("Tracing and converting fused mel+encoder…")
+        mel_encoder = MelEncoderWrapper(preprocessor, encoder)
+        traced_mel_encoder = torch.jit.trace(
+            mel_encoder, (audio_tensor, audio_length), strict=False
+        )
+        traced_mel_encoder.eval()
+        mel_encoder_inputs = [
+            ct.TensorType(
+                name="audio_signal", shape=(1, max_samples), dtype=np.float32
+            ),
+            ct.TensorType(name="audio_length", shape=(1,), dtype=np.int32),
+        ]
+        mel_encoder_outputs = [
+            ct.TensorType(name="encoder", dtype=np.float32),
+            ct.TensorType(name="encoder_length", dtype=np.int32),
+            ct.TensorType(name="frame_times", dtype=np.float32),
+        ]
+        mel_encoder_model = _coreml_convert(
+            traced_mel_encoder,
+            mel_encoder_inputs,
+            mel_encoder_outputs,
+            export_settings,
+            compute_units_override=melenc_cu,
+        )
+        mel_encoder_path = output_dir / "parakeet_eou_mel_encoder.mlpackage"
+        _save_mlpackage(
+            mel_encoder_model,
+            mel_encoder_path,
+            f"Parakeet EOU fused Mel+Encoder ({max_audio_seconds}s window)",
+        )
+
+        # === Export Decoder ===
+        typer.echo("Tracing and converting decoder…")
+        traced_decoder = torch.jit.trace(
+            decoder,
+            (targets, target_lengths, zero_state, zero_state),
+            strict=False,
+        )
+        traced_decoder.eval()
+        decoder_inputs = [
+            ct.TensorType(
+                name="targets", shape=_tensor_shape(targets), dtype=np.int32
+            ),
+            ct.TensorType(name="target_length", shape=(1,), dtype=np.int32),
+            ct.TensorType(
+                name="h_in", shape=_tensor_shape(zero_state), dtype=np.float32
+            ),
+            ct.TensorType(
+                name="c_in", shape=_tensor_shape(zero_state), dtype=np.float32
+            ),
+        ]
+        decoder_outputs = [
+            ct.TensorType(name="decoder", dtype=np.float32),
+            ct.TensorType(name="h_out", dtype=np.float32),
+            ct.TensorType(name="c_out", dtype=np.float32),
+        ]
+        decoder_model = _coreml_convert(
+            traced_decoder,
+            decoder_inputs,
+            decoder_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        decoder_path = output_dir / "parakeet_eou_decoder.mlpackage"
+        _save_mlpackage(
+            decoder_model,
+            decoder_path,
+            "Parakeet EOU decoder (RNNT prediction network)",
+        )
+
+        # === Export Joint ===
+        typer.echo("Tracing and converting joint…")
+        traced_joint = torch.jit.trace(
+            joint,
+            (encoder_ref, decoder_ref),
+            strict=False,
+        )
+        traced_joint.eval()
+        joint_inputs = [
+            ct.TensorType(
+                name="encoder", shape=_tensor_shape(encoder_ref), dtype=np.float32
+            ),
+            ct.TensorType(
+                name="decoder", shape=_tensor_shape(decoder_ref), dtype=np.float32
+            ),
+        ]
+        joint_outputs = [
+            ct.TensorType(name="logits", dtype=np.float32),
+        ]
+        joint_model = _coreml_convert(
+            traced_joint,
+            joint_inputs,
+            joint_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        joint_path = output_dir / "parakeet_eou_joint.mlpackage"
+        _save_mlpackage(
+            joint_model,
+            joint_path,
+            "Parakeet EOU joint network (RNNT)",
+        )
+
+        # === Export Joint Decision Head ===
+        typer.echo("Tracing and converting joint decision head…")
+        joint_decision = JointDecisionWrapper(joint, vocab_size=vocab_size)
+        traced_joint_decision = torch.jit.trace(
+            joint_decision,
+            (encoder_ref, decoder_ref),
+            strict=False,
+        )
+        traced_joint_decision.eval()
+        joint_decision_inputs = [
+            ct.TensorType(
+                name="encoder", shape=_tensor_shape(encoder_ref), dtype=np.float32
+            ),
+            ct.TensorType(
+                name="decoder", shape=_tensor_shape(decoder_ref), dtype=np.float32
+            ),
+        ]
+        joint_decision_outputs = [
+            ct.TensorType(name="token_id", dtype=np.int32),
+            ct.TensorType(name="token_prob", dtype=np.float32),
+        ]
+        joint_decision_model = _coreml_convert(
+            traced_joint_decision,
+            joint_decision_inputs,
+            joint_decision_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        joint_decision_path = output_dir / "parakeet_eou_joint_decision.mlpackage"
+        _save_mlpackage(
+            joint_decision_model,
+            joint_decision_path,
+            "Parakeet EOU joint + decision head (softmax, argmax)",
+        )
+
+        # === Export Single-Step Joint Decision ===
+        typer.echo("Tracing and converting single-step joint decision…")
+        jd_single = JointDecisionSingleStep(joint, vocab_size=vocab_size)
+        # Create single-step slices from refs
+        enc_step = encoder_ref[:, :, :1].contiguous()
+        dec_step = decoder_ref[:, :, :1].contiguous()
+        traced_jd_single = torch.jit.trace(
+            jd_single,
+            (enc_step, dec_step),
+            strict=False,
+        )
+        traced_jd_single.eval()
+        jd_single_inputs = [
+            ct.TensorType(
+                name="encoder_step",
+                shape=(1, enc_step.shape[1], 1),
+                dtype=np.float32,
+            ),
+            ct.TensorType(
+                name="decoder_step",
+                shape=(1, dec_step.shape[1], 1),
+                dtype=np.float32,
+            ),
+        ]
+        jd_single_outputs = [
+            ct.TensorType(name="token_id", dtype=np.int32),
+            ct.TensorType(name="token_prob", dtype=np.float32),
+            ct.TensorType(name="top_k_ids", dtype=np.int32),
+            ct.TensorType(name="top_k_logits", dtype=np.float32),
+        ]
+        jd_single_model = _coreml_convert(
+            traced_jd_single,
+            jd_single_inputs,
+            jd_single_outputs,
+            export_settings,
+            compute_units_override=ct.ComputeUnit.CPU_ONLY,
+        )
+        jd_single_path = output_dir / "parakeet_eou_joint_decision_single_step.mlpackage"
+        _save_mlpackage(
+            jd_single_model,
+            jd_single_path,
+            "Parakeet EOU single-step joint decision (current frame)",
+        )
+
+        # === Save Metadata ===
+        metadata: Dict[str, object] = {
+            "model_id": model_id,
+            "model_name": "parakeet_realtime_eou_120m-v1",
+            "model_class": type(asr_model).__name__,
+            "encoder_class": type(asr_model.encoder).__name__,
+            "sample_rate": sample_rate,
+            "max_audio_seconds": export_settings.max_audio_seconds,
+            "max_audio_samples": max_samples,
+            "max_symbol_steps": export_settings.max_symbol_steps,
+            "vocab_size": vocab_size,
+            "vocab_with_blank": vocab_size + 1,
+            "decoder_hidden": decoder_hidden,
+            "decoder_layers": decoder_layers,
+            "num_extra_outputs": num_extra,
+            "has_eou_token": True,
+            "checkpoint": checkpoint_meta,
+            "coreml": {
+                "compute_units": export_settings.compute_units.name,
+                "compute_precision": (
+                    export_settings.compute_precision.name
+                    if export_settings.compute_precision is not None
+                    else "FLOAT32"
+                ),
+            },
+            "components": {
+                "preprocessor": {
+                    "inputs": {
+                        "audio_signal": [1, max_samples],
+                        "audio_length": [1],
+                    },
+                    "outputs": {
+                        "mel": list(_tensor_shape(mel_ref)),
+                        "mel_length": [1],
+                    },
+                    "path": preprocessor_path.name,
+                },
+                "encoder": {
+                    "inputs": {
+                        "mel": list(_tensor_shape(mel_ref)),
+                        "mel_length": [1],
+                    },
+                    "outputs": {
+                        "encoder": list(_tensor_shape(encoder_ref)),
+                        "encoder_length": [1],
+                        "frame_times": [1, _tensor_shape(encoder_ref)[2]],
+                    },
+                    "path": encoder_path.name,
+                },
+                "mel_encoder": {
+                    "inputs": {
+                        "audio_signal": [1, max_samples],
+                        "audio_length": [1],
+                    },
+                    "outputs": {
+                        "encoder": list(_tensor_shape(encoder_ref)),
+                        "encoder_length": [1],
+                        "frame_times": [1, _tensor_shape(encoder_ref)[2]],
+                    },
+                    "path": mel_encoder_path.name,
+                },
+                "decoder": {
+                    "inputs": {
+                        "targets": list(_tensor_shape(targets)),
+                        "target_length": [1],
+                        "h_in": list(_tensor_shape(zero_state)),
+                        "c_in": list(_tensor_shape(zero_state)),
+                    },
+                    "outputs": {
+                        "decoder": list(_tensor_shape(decoder_ref)),
+                        "h_out": list(_tensor_shape(h_ref)),
+                        "c_out": list(_tensor_shape(c_ref)),
+                    },
+                    "path": decoder_path.name,
+                },
+                "joint": {
+                    "inputs": {
+                        "encoder": list(_tensor_shape(encoder_ref)),
+                        "decoder": list(_tensor_shape(decoder_ref)),
+                    },
+                    "outputs": {
+                        "logits": list(_tensor_shape(joint_ref)),
+                    },
+                    "path": joint_path.name,
+                },
+                "joint_decision": {
+                    "inputs": {
+                        "encoder": list(_tensor_shape(encoder_ref)),
+                        "decoder": list(_tensor_shape(decoder_ref)),
+                    },
+                    "outputs": {
+                        "token_id": [
+                            _tensor_shape(encoder_ref)[0],
+                            _tensor_shape(encoder_ref)[2],
+                            _tensor_shape(decoder_ref)[2],
+                        ],
+                        "token_prob": [
+                            _tensor_shape(encoder_ref)[0],
+                            _tensor_shape(encoder_ref)[2],
+                            _tensor_shape(decoder_ref)[2],
+                        ],
+                    },
+                    "path": joint_decision_path.name,
+                },
+                "joint_decision_single_step": {
+                    "inputs": {
+                        "encoder_step": [1, _tensor_shape(encoder_ref)[1], 1],
+                        "decoder_step": [1, _tensor_shape(decoder_ref)[1], 1],
+                    },
+                    "outputs": {
+                        "token_id": [1, 1, 1],
+                        "token_prob": [1, 1, 1],
+                        "top_k_ids": [1, 1, 1, 64],
+                        "top_k_logits": [1, 1, 1, 64],
+                    },
+                    "path": jd_single_path.name,
+                },
+            },
+        }
+
+        # Export tokenizer vocab if available
+        try:
+            tokenizer = asr_model.tokenizer
+            vocab = {
+                "blank_id": int(asr_model.decoder.blank_idx),
+                "vocab_size": vocab_size,
+            }
+            # Try to get special tokens
+            if hasattr(tokenizer, "tokenizer"):
+                inner_tokenizer = tokenizer.tokenizer
+                if hasattr(inner_tokenizer, "get_vocab"):
+                    full_vocab = inner_tokenizer.get_vocab()
+                    # Find EOU token
+                    eou_token = None
+                    for token, idx in full_vocab.items():
+                        if "<EOU>" in token.upper() or "eou" in token.lower():
+                            eou_token = {"token": token, "id": idx}
+                            break
+                    if eou_token:
+                        vocab["eou_token"] = eou_token
+            metadata["tokenizer"] = vocab
+        except Exception as e:
+            typer.echo(f"Warning: Could not export tokenizer info: {e}")
+
+        metadata_path = output_dir / "metadata.json"
+        metadata_path.write_text(json.dumps(metadata, indent=2))
+        typer.echo(f"\nExport complete. Metadata written to {metadata_path}")
+        typer.echo(f"Output directory: {output_dir}")
+
+    finally:
+        asr_model.decoder._rnnt_export = decoder_export_flag
+
+
+if __name__ == "__main__":
+    app()
diff --git a/convert_streaming_encoder.py b/convert_streaming_encoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..a1c199c8a6c540dab48b973ab4232db2901b8e72
--- /dev/null
+++ b/convert_streaming_encoder.py
@@ -0,0 +1,193 @@
+
+import torch
+import torch.nn as nn
+import coremltools as ct
+import numpy as np
+import typer
+from pathlib import Path
+from typing import Tuple, List, Optional
+import json
+import shutil
+
+# Iimport torch
+import coremltools as ct
+import numpy as np
+import argparse
+from nemo.collections.asr.models import EncDecRNNTBPEModel
+
+app = typer.Typer()
+
+class LoopbackEncoderWrapper(nn.Module):
+    """
+    Wraps the entire Parakeet Encoder (PreEncode + Conformer) for CoreML Loopback Streaming.
+    
+    Inputs:
+      - audio_signal: [B, D, T] (Mel spectrogram chunk)
+      - audio_length: [B]
+      - pre_cache: [B, D, pre_cache_size] (Previous audio context)
+      - cache_last_channel: [layers, B, cache_size, hidden]
+      - cache_last_time: [layers, B, hidden, time_cache]
+      - cache_last_channel_len: [B]
+      
+    Outputs:
+      - encoded_output: [B, D_out, T_out]
+      - encoded_length: [B]
+      - new_pre_cache: [B, D, pre_cache_size]
+      - new_cache_last_channel
+      - new_cache_last_time
+      - new_cache_last_channel_len
+    """
+    def __init__(self, encoder, pre_cache_size=16):
+        super().__init__()
+        self.encoder = encoder
+        self.pre_cache_size = pre_cache_size
+        
+    def forward(
+        self, 
+        audio_signal: torch.Tensor, 
+        audio_length: torch.Tensor,
+        pre_cache: torch.Tensor,
+        cache_last_channel: torch.Tensor,
+        cache_last_time: torch.Tensor,
+        cache_last_channel_len: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        
+        # 1. Prepend pre_cache to audio_signal
+        # audio_signal: [B, D, T]
+        # pre_cache: [B, D, T_cache]
+        full_input = torch.cat([pre_cache, audio_signal], dim=2)
+        full_length = audio_length + self.pre_cache_size
+        
+        # 2. Extract NEW pre_cache (last N frames of full_input)
+        # Note: We do this BEFORE processing because we want the raw audio context
+        new_pre_cache = full_input[:, :, -self.pre_cache_size:]
+        
+        # 3. Process with Encoder
+        # Reconstruct NeMo cache object
+        current_cache = [cache_last_channel, cache_last_time, cache_last_channel_len]
+        
+        encoded, encoded_len, new_cache_channel, new_cache_time, new_cache_len = self.encoder.cache_aware_stream_step(
+            processed_signal=full_input,
+            processed_signal_length=full_length,
+            cache_last_channel=cache_last_channel,
+            cache_last_time=cache_last_time,
+            cache_last_channel_len=cache_last_channel_len
+        )
+        
+        # 4. Drop the first few frames corresponding to pre_cache?
+        # NeMo's cache_aware_stream_step usually handles the "valid" output frames.
+        # But since we manually prepended, we might get extra output frames.
+        # However, for streaming, we usually want the model to see the context but only output the new tokens.
+        # Let's trust NeMo's streaming logic for now, or check if we need to slice.
+        # Given we are using 'cache_aware_stream_step', it expects the full context window?
+        # Actually, standard usage is: input IS the new chunk, but internal convolution looks at past.
+        # But since we are stateless, we MUST provide the past.
+        # So passing (pre_cache + chunk) is correct.
+        
+        # Cast lengths to Int32 for CoreML
+        encoded_len_32 = encoded_len.to(dtype=torch.int32)
+        new_channel_len_32 = new_cache_len.to(dtype=torch.int32)
+        
+        return encoded, encoded_len_32, new_pre_cache, new_cache_channel, new_cache_time, new_channel_len_32
+
+def _coreml_convert(
+    traced_model,
+    inputs,
+    outputs,
+    compute_units=ct.ComputeUnit.CPU_ONLY
+):
+    return ct.convert(
+        traced_model,
+        inputs=inputs,
+        outputs=outputs,
+        compute_units=compute_units,
+        minimum_deployment_target=ct.target.macOS14,
+    )
+
+def main():
+    model_id: str = "nvidia/parakeet_realtime_eou_120m-v1"
+    output_dir: str = "temp_swift_models/StreamingLoopback"
+    output_path = Path(output_dir)
+    output_path.mkdir(parents=True, exist_ok=True)
+    
+    print(f"Loading model: {model_id}...")
+    asr_model = EncDecRNNTBPEModel.from_pretrained(model_name=model_id)
+    asr_model.eval()
+    
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--chunk-frames", type=int, default=17, help="Number of frames in the input chunk (e.g. 17 for 160ms, 129 for 1.28s)")
+    args = parser.parse_args()
+
+    encoder = asr_model.encoder
+    
+    # --- Configuration ---
+    # 160ms chunk = 16 frames (but preprocessor produces 17 with padding/centering)
+    # 1.28s chunk = 128 frames (preprocessor produces 129)
+    chunk_size_in = args.chunk_frames 
+    mel_dim = 128
+    hidden_dim = encoder.d_model # 512
+    num_layers = len(encoder.layers) # 17
+    
+    # Cache sizes
+    cache_channel_size = 70
+    cache_time_size = 8
+    pre_cache_size = 16
+    
+    print(f"Config: Chunk={chunk_size_in}, Mel={mel_dim}, Hidden={hidden_dim}, Layers={num_layers}")
+    print(f"Cache: Channel={cache_channel_size}, Time={cache_time_size}, Pre={pre_cache_size}")
+
+    # --- Wrapper ---
+    wrapper = LoopbackEncoderWrapper(encoder, pre_cache_size=pre_cache_size)
+    wrapper.eval()
+    
+    # --- Test Inputs (for Tracing) ---
+    batch_size = 1
+    test_mel = torch.randn(batch_size, mel_dim, chunk_size_in)
+    test_mel_len = torch.tensor([chunk_size_in], dtype=torch.int32)
+    test_pre_cache = torch.zeros(batch_size, mel_dim, pre_cache_size)
+    
+    # Initial Cache (Zeros)
+    test_cache_channel = torch.zeros(num_layers, batch_size, cache_channel_size, hidden_dim)
+    test_cache_time = torch.zeros(num_layers, batch_size, hidden_dim, cache_time_size)
+    test_cache_len = torch.zeros(batch_size, dtype=torch.int32) 
+    
+    print("Tracing model...")
+    traced_model = torch.jit.trace(
+        wrapper,
+        (test_mel, test_mel_len, test_pre_cache, test_cache_channel, test_cache_time, test_cache_len),
+        strict=False
+    )
+    
+    # --- CoreML Conversion ---
+    print("Converting to CoreML...")
+    
+    inputs = [
+        ct.TensorType(name="audio_signal", shape=(1, 128, chunk_size_in), dtype=np.float32),
+        ct.TensorType(name="audio_length", shape=(1,), dtype=np.int32),
+        ct.TensorType(name="pre_cache", shape=(1, 128, pre_cache_size), dtype=np.float32),
+        ct.TensorType(name="cache_last_channel", shape=(num_layers, 1, cache_channel_size, hidden_dim), dtype=np.float32),
+        ct.TensorType(name="cache_last_time", shape=(num_layers, 1, hidden_dim, cache_time_size), dtype=np.float32),
+        ct.TensorType(name="cache_last_channel_len", shape=(1,), dtype=np.int32),
+    ]
+    
+    outputs = [
+        ct.TensorType(name="encoded_output", dtype=np.float32),
+        ct.TensorType(name="encoded_length", dtype=np.int32),
+        ct.TensorType(name="new_pre_cache", dtype=np.float32),
+        ct.TensorType(name="new_cache_last_channel", dtype=np.float32),
+        ct.TensorType(name="new_cache_last_time", dtype=np.float32),
+        ct.TensorType(name="new_cache_last_channel_len", dtype=np.int32),
+    ]
+    
+    mlmodel = _coreml_convert(traced_model, inputs, outputs)
+    
+    save_path = output_path / "streaming_encoder.mlpackage"
+    mlmodel.save(str(save_path))
+    print(f"Saved: {save_path}")
+    
+    # Also export Preprocessor, Decoder, Joint for completeness? 
+    # For now, let's assume we reuse the existing ones or export them separately if needed.
+    # But the user asked specifically for the Encoder loopback.
+
+if __name__ == "__main__":
+    main()
diff --git a/individual_components.py b/individual_components.py
new file mode 100644
index 0000000000000000000000000000000000000000..47271397bc8d9d17cc0fabcf4bb63be7e7c2109c
--- /dev/null
+++ b/individual_components.py
@@ -0,0 +1,250 @@
+#!/usr/bin/env python3
+"""Export Parakeet Realtime EOU 120M RNNT components into CoreML."""
+from __future__ import annotations
+
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Optional, Tuple
+
+import coremltools as ct
+import torch
+
+
+@dataclass
+class ExportSettings:
+    output_dir: Path
+    compute_units: ct.ComputeUnit
+    deployment_target: Optional[ct.target]
+    compute_precision: Optional[ct.precision]
+    max_audio_seconds: float
+    max_symbol_steps: int
+
+
+class PreprocessorWrapper(torch.nn.Module):
+    """Wrapper for the audio preprocessor (mel spectrogram extraction)."""
+
+    def __init__(self, module: torch.nn.Module) -> None:
+        super().__init__()
+        self.module = module
+
+    def forward(
+        self, audio_signal: torch.Tensor, length: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        mel, mel_length = self.module(
+            input_signal=audio_signal, length=length.to(dtype=torch.long)
+        )
+        return mel, mel_length
+
+
+class EncoderWrapper(torch.nn.Module):
+    """Wrapper for the cache-aware FastConformer encoder.
+
+    Note: For the realtime EOU model, the encoder is cache-aware which means
+    it can operate in a streaming fashion. For CoreML export, we export
+    without cache state for simplicity (full-context mode).
+    """
+
+    def __init__(self, module: torch.nn.Module) -> None:
+        super().__init__()
+        self.module = module
+
+    def forward(
+        self, features: torch.Tensor, length: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        encoded, encoded_lengths = self.module(
+            audio_signal=features, length=length.to(dtype=torch.long)
+        )
+        # Synthesize per-frame timestamps (seconds) using the 80 ms encoder stride.
+        # Shape: [B, T_enc]
+        frame_times = (
+            torch.arange(encoded.shape[-1], device=encoded.device, dtype=torch.float32)
+            * 0.08
+        )
+        return encoded, encoded_lengths, frame_times
+
+
+class DecoderWrapper(torch.nn.Module):
+    """Wrapper for the RNNT prediction network (decoder)."""
+
+    def __init__(self, module: torch.nn.Module) -> None:
+        super().__init__()
+        self.module = module
+
+    def forward(
+        self,
+        targets: torch.Tensor,
+        target_lengths: torch.Tensor,
+        h_in: torch.Tensor,
+        c_in: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        state = [h_in, c_in]
+        decoder_output, _, new_state = self.module(
+            targets=targets.to(dtype=torch.long),
+            target_length=target_lengths.to(dtype=torch.long),
+            states=state,
+        )
+        return decoder_output, new_state[0], new_state[1]
+
+
+class JointWrapper(torch.nn.Module):
+    """Wrapper for the RNNT joint network.
+
+    Note: Unlike Parakeet TDT v3, the realtime EOU model does NOT have
+    duration outputs (num_extra_outputs). The joint network outputs only
+    token logits over the vocabulary + blank.
+    """
+
+    def __init__(self, module: torch.nn.Module) -> None:
+        super().__init__()
+        self.module = module
+
+    def forward(
+        self, encoder_outputs: torch.Tensor, decoder_outputs: torch.Tensor
+    ) -> torch.Tensor:
+        # Input: encoder_outputs [B, D, T], decoder_outputs [B, D, U]
+        # Transpose to match what projection layers expect
+        encoder_outputs = encoder_outputs.transpose(1, 2)  # [B, T, D]
+        decoder_outputs = decoder_outputs.transpose(1, 2)  # [B, U, D]
+
+        # Apply projections
+        enc_proj = self.module.enc(encoder_outputs)  # [B, T, joint_hidden]
+        dec_proj = self.module.pred(decoder_outputs)  # [B, U, joint_hidden]
+
+        # Explicit broadcasting along T and U to avoid converter ambiguity
+        x = enc_proj.unsqueeze(2) + dec_proj.unsqueeze(1)  # [B, T, U, joint_hidden]
+        x = self.module.joint_net[0](x)  # ReLU
+        x = self.module.joint_net[1](x)  # Dropout (no-op in eval)
+        out = self.module.joint_net[2](x)  # Linear -> logits [B, T, U, vocab+blank]
+        return out
+
+
+class MelEncoderWrapper(torch.nn.Module):
+    """Fused wrapper: waveform -> mel -> encoder.
+
+    Inputs:
+      - audio_signal: [B, S]
+      - audio_length: [B]
+
+    Outputs:
+      - encoder: [B, D, T_enc]
+      - encoder_length: [B]
+      - frame_times: [T_enc]
+    """
+
+    def __init__(
+        self, preprocessor: PreprocessorWrapper, encoder: EncoderWrapper
+    ) -> None:
+        super().__init__()
+        self.preprocessor = preprocessor
+        self.encoder = encoder
+
+    def forward(
+        self, audio_signal: torch.Tensor, audio_length: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        mel, mel_length = self.preprocessor(audio_signal, audio_length)
+        encoded, enc_len, frame_times = self.encoder(mel, mel_length.to(dtype=torch.int32))
+        return encoded, enc_len, frame_times
+
+
+class JointDecisionWrapper(torch.nn.Module):
+    """Joint + decision head: outputs label id and label prob.
+
+    Unlike Parakeet TDT v3, this model does NOT have duration outputs.
+
+    Inputs:
+      - encoder_outputs: [B, D, T]
+      - decoder_outputs: [B, D, U]
+
+    Returns:
+      - token_id: [B, T, U] int32
+      - token_prob: [B, T, U] float32
+    """
+
+    def __init__(self, joint: JointWrapper, vocab_size: int) -> None:
+        super().__init__()
+        self.joint = joint
+        self.vocab_with_blank = int(vocab_size) + 1
+
+    def forward(self, encoder_outputs: torch.Tensor, decoder_outputs: torch.Tensor):
+        logits = self.joint(encoder_outputs, decoder_outputs)
+
+        # Token selection
+        token_ids = torch.argmax(logits, dim=-1).to(dtype=torch.int32)
+        token_probs_all = torch.softmax(logits, dim=-1)
+        # gather expects int64 (long) indices; cast only for gather
+        token_prob = torch.gather(
+            token_probs_all, dim=-1, index=token_ids.long().unsqueeze(-1)
+        ).squeeze(-1)
+
+        return token_ids, token_prob
+
+
+class JointDecisionSingleStep(torch.nn.Module):
+    """Single-step variant for streaming: encoder_step -> token decision.
+
+    Inputs:
+      - encoder_step: [B=1, D, T=1]
+      - decoder_step: [B=1, D, U=1]
+
+    Returns:
+      - token_id: [1, 1, 1] int32
+      - token_prob: [1, 1, 1] float32
+      - top_k_ids: [1, 1, 1, K] int32
+      - top_k_logits: [1, 1, 1, K] float32
+    """
+
+    def __init__(self, joint: JointWrapper, vocab_size: int, top_k: int = 64) -> None:
+        super().__init__()
+        self.joint = joint
+        self.vocab_with_blank = int(vocab_size) + 1
+        self.top_k = int(top_k)
+
+    def forward(self, encoder_step: torch.Tensor, decoder_step: torch.Tensor):
+        # Reuse JointWrapper which expects [B, D, T] and [B, D, U]
+        logits = self.joint(encoder_step, decoder_step)  # [1, 1, 1, V+blank]
+
+        token_ids = torch.argmax(logits, dim=-1, keepdim=False).to(dtype=torch.int32)
+        token_probs_all = torch.softmax(logits, dim=-1)
+        token_prob = torch.gather(
+            token_probs_all, dim=-1, index=token_ids.long().unsqueeze(-1)
+        ).squeeze(-1)
+
+        # Also expose top-K candidates for host-side processing
+        topk_logits, topk_ids_long = torch.topk(
+            logits, k=min(self.top_k, logits.shape[-1]), dim=-1
+        )
+        topk_ids = topk_ids_long.to(dtype=torch.int32)
+        return token_ids, token_prob, topk_ids, topk_logits
+
+
+def _coreml_convert(
+    traced: torch.jit.ScriptModule,
+    inputs,
+    outputs,
+    settings: ExportSettings,
+    compute_units_override: Optional[ct.ComputeUnit] = None,
+    compute_precision: Optional[ct.precision] = None,
+) -> ct.models.MLModel:
+    cu = (
+        compute_units_override
+        if compute_units_override is not None
+        else settings.compute_units
+    )
+    kwargs = {
+        "convert_to": "mlprogram",
+        "inputs": inputs,
+        "outputs": outputs,
+        "compute_units": cu,
+    }
+    print("Converting:", traced.__class__.__name__)
+    print("Conversion kwargs:", kwargs)
+    if settings.deployment_target is not None:
+        kwargs["minimum_deployment_target"] = settings.deployment_target
+    
+    # Priority: explicit argument > settings
+    if compute_precision is not None:
+        kwargs["compute_precision"] = compute_precision
+    elif settings.compute_precision is not None:
+        kwargs["compute_precision"] = settings.compute_precision
+        
+    return ct.convert(traced, **kwargs)
diff --git a/metadata.json b/metadata.json
new file mode 100644
index 0000000000000000000000000000000000000000..25813362b8502d5f97b8b92d03aec535f93ed0a8
--- /dev/null
+++ b/metadata.json
@@ -0,0 +1,23 @@
+{
+  "model_id": "nvidia/parakeet_realtime_eou_120m-v1",
+  "model_name": "parakeet_realtime_eou_120m-v1-split",
+  "streaming_mode": "split_encoder",
+  "sample_rate": 16000,
+  "mel_dim": 128,
+  "hidden_dim": 512,
+  "num_layers": 17,
+  "mel_frames_per_chunk": 1005,
+  "vocab_size": 1026,
+  "blank_id": 1026,
+  "decoder_hidden": 640,
+  "decoder_layers": 1,
+  "cache_channel_size": 70,
+  "cache_time_size": 8,
+  "components": {
+    "preprocessor": "preprocessor.mlpackage",
+    "pre_encode": "pre_encode.mlpackage",
+    "conformer": "conformer_streaming.mlpackage",
+    "decoder": "decoder.mlpackage",
+    "joint_decision": "joint_decision.mlpackage"
+  }
+}
\ No newline at end of file
diff --git a/tokenizer.model b/tokenizer.model
new file mode 100644
index 0000000000000000000000000000000000000000..ac8cf05bb6c18f530e2bf89d965c62dbd4feda33
--- /dev/null
+++ b/tokenizer.model
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d556e51ba5b89db64a8cb2be6798fb29974edcadb58b0c7b80418eb5d8752303
+size 258183
diff --git a/vocab.json b/vocab.json
new file mode 100644
index 0000000000000000000000000000000000000000..462081190dbb331a0dde13e99d5cfa99b302f0a7
--- /dev/null
+++ b/vocab.json
@@ -0,0 +1,1028 @@
+{
+  "0": "<unk>",
+  "1": "▁t",
+  "2": "▁th",
+  "3": "▁a",
+  "4": "▁i",
+  "5": "▁the",
+  "6": "▁s",
+  "7": "re",
+  "8": "▁w",
+  "9": "▁o",
+  "10": "in",
+  "11": "at",
+  "12": "er",
+  "13": "nd",
+  "14": "ou",
+  "15": "▁c",
+  "16": "▁b",
+  "17": "▁h",
+  "18": "en",
+  "19": "on",
+  "20": "▁m",
+  "21": "▁f",
+  "22": "ing",
+  "23": "▁p",
+  "24": "▁to",
+  "25": "▁and",
+  "26": "▁d",
+  "27": "an",
+  "28": "or",
+  "29": "es",
+  "30": "▁y",
+  "31": "▁l",
+  "32": "▁of",
+  "33": "ll",
+  "34": "▁in",
+  "35": "ed",
+  "36": "it",
+  "37": "▁g",
+  "38": "is",
+  "39": "▁you",
+  "40": "▁n",
+  "41": "ar",
+  "42": "om",
+  "43": "as",
+  "44": "ve",
+  "45": "▁e",
+  "46": "ic",
+  "47": "▁it",
+  "48": "al",
+  "49": "us",
+  "50": "▁wh",
+  "51": "▁we",
+  "52": "▁be",
+  "53": "ion",
+  "54": "ow",
+  "55": "le",
+  "56": "▁is",
+  "57": "et",
+  "58": "ent",
+  "59": "ot",
+  "60": "ut",
+  "61": "▁re",
+  "62": "▁on",
+  "63": "ay",
+  "64": "▁ha",
+  "65": "ig",
+  "66": "▁so",
+  "67": "ct",
+  "68": "▁he",
+  "69": "▁for",
+  "70": "ver",
+  "71": "ke",
+  "72": "ro",
+  "73": "▁st",
+  "74": "id",
+  "75": "▁go",
+  "76": "all",
+  "77": "se",
+  "78": "ly",
+  "79": "▁u",
+  "80": "ch",
+  "81": "st",
+  "82": "ld",
+  "83": "▁k",
+  "84": "ce",
+  "85": "ur",
+  "86": "▁li",
+  "87": "am",
+  "88": "▁r",
+  "89": "ht",
+  "90": "▁j",
+  "91": "ith",
+  "92": "▁se",
+  "93": "ir",
+  "94": "▁as",
+  "95": "▁an",
+  "96": "im",
+  "97": "▁do",
+  "98": "ad",
+  "99": "▁was",
+  "100": "ight",
+  "101": "th",
+  "102": "▁are",
+  "103": "▁but",
+  "104": "▁sh",
+  "105": "ust",
+  "106": "ally",
+  "107": "▁not",
+  "108": "▁or",
+  "109": "▁com",
+  "110": "▁can",
+  "111": "▁me",
+  "112": "op",
+  "113": "▁mo",
+  "114": "▁at",
+  "115": "ill",
+  "116": "▁ch",
+  "117": "▁ne",
+  "118": "ant",
+  "119": "▁de",
+  "120": "▁kn",
+  "121": "▁one",
+  "122": "il",
+  "123": "ol",
+  "124": "▁con",
+  "125": "ter",
+  "126": "▁ab",
+  "127": "▁fr",
+  "128": "ere",
+  "129": "ck",
+  "130": "▁al",
+  "131": "▁all",
+  "132": "qu",
+  "133": "▁pro",
+  "134": "▁som",
+  "135": "ould",
+  "136": "▁tw",
+  "137": "ul",
+  "138": "ra",
+  "139": "od",
+  "140": "ers",
+  "141": "▁su",
+  "142": "ive",
+  "143": "▁v",
+  "144": "use",
+  "145": "ate",
+  "146": "ge",
+  "147": "if",
+  "148": "▁ex",
+  "149": "ess",
+  "150": "pp",
+  "151": "▁lo",
+  "152": "out",
+  "153": "▁if",
+  "154": "est",
+  "155": "ain",
+  "156": "ist",
+  "157": "and",
+  "158": "ea",
+  "159": "very",
+  "160": "art",
+  "161": "▁wor",
+  "162": "▁my",
+  "163": "ab",
+  "164": "ment",
+  "165": "▁bec",
+  "166": "un",
+  "167": "ity",
+  "168": "ri",
+  "169": "pe",
+  "170": "ions",
+  "171": "▁by",
+  "172": "ok",
+  "173": "our",
+  "174": "ort",
+  "175": "ind",
+  "176": "ink",
+  "177": "nt",
+  "178": "▁up",
+  "179": "um",
+  "180": "▁don",
+  "181": "▁get",
+  "182": "red",
+  "183": "▁out",
+  "184": "el",
+  "185": "ause",
+  "186": "res",
+  "187": "▁ma",
+  "188": "ich",
+  "189": "▁us",
+  "190": "rou",
+  "191": "▁int",
+  "192": "em",
+  "193": "os",
+  "194": "ies",
+  "195": "ie",
+  "196": "▁pl",
+  "197": "▁tr",
+  "198": "ven",
+  "199": "ous",
+  "200": "▁le",
+  "201": "▁two",
+  "202": "ard",
+  "203": "ine",
+  "204": "▁co",
+  "205": "een",
+  "206": "▁now",
+  "207": "ty",
+  "208": "her",
+  "209": "ack",
+  "210": "▁pe",
+  "211": "ame",
+  "212": "▁how",
+  "213": "▁who",
+  "214": "▁see",
+  "215": "▁tim",
+  "216": "ect",
+  "217": "ast",
+  "218": "▁our",
+  "219": "ci",
+  "220": "ree",
+  "221": "ople",
+  "222": "gh",
+  "223": "▁no",
+  "224": "▁had",
+  "225": "▁man",
+  "226": "▁qu",
+  "227": "▁en",
+  "228": "ide",
+  "229": "ure",
+  "230": "ud",
+  "231": "so",
+  "232": "▁his",
+  "233": "▁sa",
+  "234": "▁sp",
+  "235": "▁say",
+  "236": "ose",
+  "237": "ther",
+  "238": "▁act",
+  "239": "▁ta",
+  "240": "▁cl",
+  "241": "ings",
+  "242": "pt",
+  "243": "king",
+  "244": "▁any",
+  "245": "▁has",
+  "246": "▁un",
+  "247": "iv",
+  "248": "▁im",
+  "249": "▁ag",
+  "250": "▁te",
+  "251": "▁fe",
+  "252": "one",
+  "253": "per",
+  "254": "ong",
+  "255": "▁po",
+  "256": "▁ad",
+  "257": "ff",
+  "258": "ore",
+  "259": "itt",
+  "260": "ans",
+  "261": "iz",
+  "262": "eah",
+  "263": "reat",
+  "264": "act",
+  "265": "own",
+  "266": "hing",
+  "267": "enty",
+  "268": "age",
+  "269": "ber",
+  "270": "ice",
+  "271": "▁am",
+  "272": "ple",
+  "273": "are",
+  "274": "▁per",
+  "275": "und",
+  "276": "ite",
+  "277": "ix",
+  "278": "pl",
+  "279": "▁way",
+  "280": "▁did",
+  "281": "▁pr",
+  "282": "▁got",
+  "283": "ars",
+  "284": "▁she",
+  "285": "▁let",
+  "286": "ag",
+  "287": "▁ac",
+  "288": "int",
+  "289": "▁ar",
+  "290": "ry",
+  "291": "ign",
+  "292": "ish",
+  "293": "▁fir",
+  "294": "ace",
+  "295": "ble",
+  "296": "og",
+  "297": "ue",
+  "298": "▁ye",
+  "299": "ap",
+  "300": "iff",
+  "301": "▁ro",
+  "302": "▁her",
+  "303": "nder",
+  "304": "▁ok",
+  "305": "▁res",
+  "306": "▁gu",
+  "307": "ence",
+  "308": "▁may",
+  "309": "ated",
+  "310": "ip",
+  "311": "▁bo",
+  "312": "▁him",
+  "313": "way",
+  "314": "ac",
+  "315": "ical",
+  "316": "ass",
+  "317": "ase",
+  "318": "▁dis",
+  "319": "able",
+  "320": "ick",
+  "321": "▁app",
+  "322": "ance",
+  "323": "▁pre",
+  "324": "▁six",
+  "325": "▁off",
+  "326": "▁new",
+  "327": "ia",
+  "328": "orm",
+  "329": "ank",
+  "330": "▁lot",
+  "331": "ach",
+  "332": "▁fo",
+  "333": "inet",
+  "334": "ire",
+  "335": "ary",
+  "336": "ult",
+  "337": "▁tal",
+  "338": "▁mu",
+  "339": "▁bl",
+  "340": "ount",
+  "341": "sel",
+  "342": "vel",
+  "343": "▁br",
+  "344": "▁imp",
+  "345": "ep",
+  "346": "cess",
+  "347": "ord",
+  "348": "▁sc",
+  "349": "▁inc",
+  "350": "ound",
+  "351": "ang",
+  "352": "be",
+  "353": "ress",
+  "354": "uct",
+  "355": "▁ind",
+  "356": "▁af",
+  "357": "ving",
+  "358": "▁oh",
+  "359": "▁bet",
+  "360": "▁use",
+  "361": "ome",
+  "362": "ens",
+  "363": "ys",
+  "364": "▁bu",
+  "365": "co",
+  "366": "ory",
+  "367": "ater",
+  "368": "ild",
+  "369": "ght",
+  "370": "ial",
+  "371": "▁day",
+  "372": "ning",
+  "373": "na",
+  "374": "ile",
+  "375": "▁spe",
+  "376": "▁mar",
+  "377": "ody",
+  "378": "ough",
+  "379": "ade",
+  "380": "vers",
+  "381": "xt",
+  "382": "▁fl",
+  "383": "▁ke",
+  "384": "ian",
+  "385": "▁sy",
+  "386": "▁put",
+  "387": "fore",
+  "388": "ub",
+  "389": "▁ph",
+  "390": "fe",
+  "391": "▁em",
+  "392": "▁ser",
+  "393": "form",
+  "394": "ting",
+  "395": "te",
+  "396": "av",
+  "397": "ious",
+  "398": "▁rec",
+  "399": "ks",
+  "400": "▁gr",
+  "401": "ces",
+  "402": "wn",
+  "403": "ors",
+  "404": "▁jo",
+  "405": "ents",
+  "406": "▁des",
+  "407": "▁try",
+  "408": "▁equ",
+  "409": "▁z",
+  "410": "▁rem",
+  "411": "▁str",
+  "412": "self",
+  "413": "▁bit",
+  "414": "ph",
+  "415": "ved",
+  "416": "▁why",
+  "417": "▁bas",
+  "418": "▁hel",
+  "419": "▁rel",
+  "420": "ath",
+  "421": "ject",
+  "422": "ail",
+  "423": "▁la",
+  "424": "ual",
+  "425": "▁god",
+  "426": "▁nat",
+  "427": "erm",
+  "428": "day",
+  "429": "▁id",
+  "430": "ft",
+  "431": "▁wr",
+  "432": "▁min",
+  "433": "ates",
+  "434": "▁gen",
+  "435": "tain",
+  "436": "▁ob",
+  "437": "ull",
+  "438": "ict",
+  "439": "▁tra",
+  "440": "▁end",
+  "441": "▁hig",
+  "442": "▁fif",
+  "443": "oth",
+  "444": "tern",
+  "445": "▁its",
+  "446": "vent",
+  "447": "▁sm",
+  "448": "ons",
+  "449": "▁add",
+  "450": "iss",
+  "451": "▁bel",
+  "452": "ful",
+  "453": "get",
+  "454": "▁ele",
+  "455": "▁rep",
+  "456": "ak",
+  "457": "▁ho",
+  "458": "▁pos",
+  "459": "▁num",
+  "460": "ange",
+  "461": "ves",
+  "462": "ific",
+  "463": "urn",
+  "464": "ise",
+  "465": "▁cr",
+  "466": "▁um",
+  "467": "ward",
+  "468": "▁reg",
+  "469": "ady",
+  "470": "ower",
+  "471": "uc",
+  "472": "▁dec",
+  "473": "lic",
+  "474": "▁set",
+  "475": "▁gon",
+  "476": "▁op",
+  "477": "▁ear",
+  "478": "▁sub",
+  "479": "▁sl",
+  "480": "les",
+  "481": "stem",
+  "482": "cial",
+  "483": "olog",
+  "484": "atch",
+  "485": "ily",
+  "486": "body",
+  "487": "nds",
+  "488": "ular",
+  "489": "ren",
+  "490": "▁own",
+  "491": "▁too",
+  "492": "cent",
+  "493": "ible",
+  "494": "pect",
+  "495": "ered",
+  "496": "ways",
+  "497": "teen",
+  "498": "▁uh",
+  "499": "▁big",
+  "500": "▁mod",
+  "501": "▁att",
+  "502": "▁car",
+  "503": "gr",
+  "504": "▁acc",
+  "505": "ied",
+  "506": "mun",
+  "507": "ib",
+  "508": "▁mon",
+  "509": "▁sch",
+  "510": "▁pol",
+  "511": "▁dat",
+  "512": "▁fin",
+  "513": "▁sim",
+  "514": "▁inv",
+  "515": "▁def",
+  "516": "ked",
+  "517": "▁ent",
+  "518": "▁yes",
+  "519": "ows",
+  "520": "ics",
+  "521": "ited",
+  "522": "ute",
+  "523": "ism",
+  "524": "ps",
+  "525": "▁ed",
+  "526": "▁el",
+  "527": "ably",
+  "528": "ppen",
+  "529": "als",
+  "530": "▁ten",
+  "531": "ract",
+  "532": "ss",
+  "533": "▁ass",
+  "534": "▁met",
+  "535": "gan",
+  "536": "▁eng",
+  "537": "▁stu",
+  "538": "ween",
+  "539": "arch",
+  "540": "▁gl",
+  "541": "▁cor",
+  "542": "▁dr",
+  "543": "vern",
+  "544": "▁ty",
+  "545": "▁run",
+  "546": "hip",
+  "547": "cus",
+  "548": "cond",
+  "549": "▁ins",
+  "550": "irty",
+  "551": "▁pub",
+  "552": "lud",
+  "553": "llow",
+  "554": "▁cou",
+  "555": "ew",
+  "556": "iew",
+  "557": "▁sur",
+  "558": "ero",
+  "559": "ood",
+  "560": "ness",
+  "561": "▁fun",
+  "562": "▁eff",
+  "563": "cept",
+  "564": "▁ca",
+  "565": "▁exp",
+  "566": "duct",
+  "567": "▁sw",
+  "568": "ize",
+  "569": "ope",
+  "570": "▁par",
+  "571": "kes",
+  "572": "cy",
+  "573": "▁ev",
+  "574": "▁ref",
+  "575": "ell",
+  "576": "▁bus",
+  "577": "ug",
+  "578": "rib",
+  "579": "▁cur",
+  "580": "mo",
+  "581": "ock",
+  "582": "ures",
+  "583": "air",
+  "584": "▁war",
+  "585": "str",
+  "586": "▁med",
+  "587": "▁wa",
+  "588": "▁val",
+  "589": "▁sin",
+  "590": "blem",
+  "591": "▁fam",
+  "592": "li",
+  "593": "▁far",
+  "594": "▁cle",
+  "595": "▁col",
+  "596": "mon",
+  "597": "▁gra",
+  "598": "led",
+  "599": "ense",
+  "600": "tin",
+  "601": "ues",
+  "602": "its",
+  "603": "▁mem",
+  "604": "▁inf",
+  "605": "▁eas",
+  "606": "ideo",
+  "607": "▁top",
+  "608": "io",
+  "609": "pan",
+  "610": "▁hum",
+  "611": "▁old",
+  "612": "ead",
+  "613": "▁ord",
+  "614": "ric",
+  "615": "ants",
+  "616": "oy",
+  "617": "esn",
+  "618": "uck",
+  "619": "ason",
+  "620": "ced",
+  "621": "ool",
+  "622": "rat",
+  "623": "ouse",
+  "624": "▁lar",
+  "625": "▁art",
+  "626": "▁wee",
+  "627": "▁cer",
+  "628": "ized",
+  "629": "▁mat",
+  "630": "con",
+  "631": "erg",
+  "632": "land",
+  "633": "ines",
+  "634": "▁chr",
+  "635": "▁aut",
+  "636": "▁lea",
+  "637": "▁sou",
+  "638": "oney",
+  "639": "tty",
+  "640": "▁ple",
+  "641": "ulat",
+  "642": "oks",
+  "643": "▁few",
+  "644": "▁sol",
+  "645": "▁che",
+  "646": "chn",
+  "647": "ird",
+  "648": "▁bre",
+  "649": "▁dur",
+  "650": "▁wom",
+  "651": "me",
+  "652": "izat",
+  "653": "eric",
+  "654": "ote",
+  "655": "▁uni",
+  "656": "eren",
+  "657": "arn",
+  "658": "ross",
+  "659": "ices",
+  "660": "ten",
+  "661": "eral",
+  "662": "ever",
+  "663": "ieve",
+  "664": "lish",
+  "665": "ash",
+  "666": "▁opp",
+  "667": "alth",
+  "668": "ger",
+  "669": "▁sk",
+  "670": "▁red",
+  "671": "peri",
+  "672": "▁det",
+  "673": "▁ext",
+  "674": "ner",
+  "675": "ah",
+  "676": "▁var",
+  "677": "▁loc",
+  "678": "gram",
+  "679": "ists",
+  "680": "ives",
+  "681": "▁es",
+  "682": "▁nor",
+  "683": "tro",
+  "684": "ale",
+  "685": "▁iss",
+  "686": "▁pri",
+  "687": "gin",
+  "688": "az",
+  "689": "oc",
+  "690": "▁pop",
+  "691": "ern",
+  "692": "▁sit",
+  "693": "ket",
+  "694": "▁pa",
+  "695": "▁law",
+  "696": "ages",
+  "697": "br",
+  "698": "▁cam",
+  "699": "▁mom",
+  "700": "osed",
+  "701": "▁bro",
+  "702": "ne",
+  "703": "bs",
+  "704": "▁cre",
+  "705": "erat",
+  "706": "▁sec",
+  "707": "▁cap",
+  "708": "▁vis",
+  "709": "▁pat",
+  "710": "ield",
+  "711": "iet",
+  "712": "▁tri",
+  "713": "up",
+  "714": "▁bra",
+  "715": "ts",
+  "716": "▁mot",
+  "717": "▁unt",
+  "718": "put",
+  "719": "bo",
+  "720": "ork",
+  "721": "mer",
+  "722": "ital",
+  "723": "▁air",
+  "724": "ined",
+  "725": "▁beh",
+  "726": "▁adv",
+  "727": "▁ret",
+  "728": "imes",
+  "729": "▁tea",
+  "730": "ural",
+  "731": "sid",
+  "732": "ters",
+  "733": "▁pur",
+  "734": "▁sci",
+  "735": "bers",
+  "736": "ient",
+  "737": "ier",
+  "738": "cc",
+  "739": "sw",
+  "740": "▁av",
+  "741": "reen",
+  "742": "ode",
+  "743": "ont",
+  "744": "▁dra",
+  "745": "ann",
+  "746": "nect",
+  "747": "▁x",
+  "748": "▁eu",
+  "749": "ton",
+  "750": "inat",
+  "751": "ene",
+  "752": "ared",
+  "753": "els",
+  "754": "▁mor",
+  "755": "▁rat",
+  "756": "cri",
+  "757": "▁men",
+  "758": "▁ah",
+  "759": "ames",
+  "760": "▁arm",
+  "761": "eak",
+  "762": "▁pay",
+  "763": "▁hal",
+  "764": "ins",
+  "765": "ilit",
+  "766": "stit",
+  "767": "▁ra",
+  "768": "▁leg",
+  "769": "cl",
+  "770": "pr",
+  "771": "▁wal",
+  "772": "▁bad",
+  "773": "▁ge",
+  "774": "roup",
+  "775": "▁mus",
+  "776": "man",
+  "777": "▁gi",
+  "778": "eds",
+  "779": "▁aw",
+  "780": "po",
+  "781": "ark",
+  "782": "row",
+  "783": "▁dep",
+  "784": "ully",
+  "785": "ral",
+  "786": "lect",
+  "787": "pend",
+  "788": "▁sev",
+  "789": "ime",
+  "790": "gest",
+  "791": "here",
+  "792": "▁yet",
+  "793": "ted",
+  "794": "▁rev",
+  "795": "ds",
+  "796": "▁ask",
+  "797": "less",
+  "798": "▁di",
+  "799": "ets",
+  "800": "line",
+  "801": "▁aff",
+  "802": "ired",
+  "803": "▁est",
+  "804": "ken",
+  "805": "vid",
+  "806": "most",
+  "807": "ivid",
+  "808": "unch",
+  "809": "par",
+  "810": "med",
+  "811": "rop",
+  "812": "ased",
+  "813": "eone",
+  "814": "▁ve",
+  "815": "▁abs",
+  "816": "ergy",
+  "817": "ret",
+  "818": "▁saw",
+  "819": "▁ey",
+  "820": "▁cal",
+  "821": "uat",
+  "822": "▁mid",
+  "823": "vat",
+  "824": "ream",
+  "825": "vice",
+  "826": "ians",
+  "827": "rent",
+  "828": "ctor",
+  "829": "err",
+  "830": "ush",
+  "831": "ases",
+  "832": "▁suc",
+  "833": "erms",
+  "834": "ave",
+  "835": "angu",
+  "836": "ries",
+  "837": "▁wo",
+  "838": "arts",
+  "839": "▁fil",
+  "840": "▁fat",
+  "841": "▁cho",
+  "842": "orts",
+  "843": "▁fre",
+  "844": "ee",
+  "845": "ught",
+  "846": "eng",
+  "847": "ump",
+  "848": "▁bar",
+  "849": "ying",
+  "850": "ane",
+  "851": "▁tem",
+  "852": "anks",
+  "853": "ury",
+  "854": "iat",
+  "855": "mit",
+  "856": "trol",
+  "857": "▁net",
+  "858": "▁maj",
+  "859": "▁cra",
+  "860": "ling",
+  "861": "▁fig",
+  "862": "orn",
+  "863": "icat",
+  "864": "pany",
+  "865": "▁occ",
+  "866": "ott",
+  "867": "ands",
+  "868": "▁exc",
+  "869": "▁mr",
+  "870": "ency",
+  "871": "rope",
+  "872": "itch",
+  "873": "▁lit",
+  "874": "abil",
+  "875": "not",
+  "876": "ma",
+  "877": "▁typ",
+  "878": "▁opt",
+  "879": "ob",
+  "880": "ser",
+  "881": "ety",
+  "882": "ms",
+  "883": "peci",
+  "884": "aces",
+  "885": "aut",
+  "886": "▁hon",
+  "887": "cuss",
+  "888": "▁sal",
+  "889": "▁sor",
+  "890": "att",
+  "891": "▁lab",
+  "892": "▁har",
+  "893": "urch",
+  "894": "nded",
+  "895": "uce",
+  "896": "ids",
+  "897": "▁hy",
+  "898": "▁fut",
+  "899": "▁ste",
+  "900": "ours",
+  "901": "ems",
+  "902": "utes",
+  "903": "ng",
+  "904": "ta",
+  "905": "▁won",
+  "906": "▁fa",
+  "907": "▁env",
+  "908": "ards",
+  "909": "▁job",
+  "910": "ium",
+  "911": "▁dot",
+  "912": "▁obv",
+  "913": "ina",
+  "914": "side",
+  "915": "elve",
+  "916": "cu",
+  "917": "▁jes",
+  "918": "▁pot",
+  "919": "▁pie",
+  "920": "▁tre",
+  "921": "▁hey",
+  "922": "▁mag",
+  "923": "ron",
+  "924": "▁key",
+  "925": "swer",
+  "926": "▁win",
+  "927": "ucat",
+  "928": "work",
+  "929": "ides",
+  "930": "▁low",
+  "931": "▁vol",
+  "932": "▁oth",
+  "933": "atic",
+  "934": "lf",
+  "935": "ads",
+  "936": "inds",
+  "937": "com",
+  "938": "ths",
+  "939": "▁ver",
+  "940": "ised",
+  "941": "lo",
+  "942": "▁squ",
+  "943": "▁cut",
+  "944": "oked",
+  "945": "irit",
+  "946": "ateg",
+  "947": "ppy",
+  "948": "mitt",
+  "949": "come",
+  "950": "hn",
+  "951": "igin",
+  "952": "mand",
+  "953": "▁dam",
+  "954": "ho",
+  "955": "▁da",
+  "956": "▁fur",
+  "957": "iron",
+  "958": "ilar",
+  "959": "▁fac",
+  "960": "▁neg",
+  "961": "▁ago",
+  "962": "ged",
+  "963": "miss",
+  "964": "enth",
+  "965": "▁dou",
+  "966": "▁hit",
+  "967": "▁guy",
+  "968": "▁bi",
+  "969": "ove",
+  "970": "fess",
+  "971": "ples",
+  "972": "owed",
+  "973": "ured",
+  "974": "▁ris",
+  "975": "ints",
+  "976": "rew",
+  "977": "▁sum",
+  "978": "▁hu",
+  "979": "ploy",
+  "980": "ude",
+  "981": "ried",
+  "982": "▁cir",
+  "983": "▁dev",
+  "984": "ear",
+  "985": "▁tot",
+  "986": "▁ann",
+  "987": "duc",
+  "988": "ik",
+  "989": "pon",
+  "990": "sted",
+  "991": "▁ide",
+  "992": "▁'",
+  "993": "ipp",
+  "994": "▁eat",
+  "995": "▁dom",
+  "996": "▁",
+  "997": "e",
+  "998": "t",
+  "999": "o",
+  "1000": "a",
+  "1001": "i",
+  "1002": "n",
+  "1003": "s",
+  "1004": "r",
+  "1005": "h",
+  "1006": "l",
+  "1007": "d",
+  "1008": "u",
+  "1009": "c",
+  "1010": "m",
+  "1011": "y",
+  "1012": "g",
+  "1013": "w",
+  "1014": "f",
+  "1015": "p",
+  "1016": "b",
+  "1017": "v",
+  "1018": "k",
+  "1019": "'",
+  "1020": "j",
+  "1021": "x",
+  "1022": "q",
+  "1023": "z",
+  "1024": "<EOU>",
+  "1025": "<EOB>"
+}
\ No newline at end of file