Upload folder using huggingface_hub

.claude/settings.local.json

@@ -0,0 +1,7 @@
+{
+  "permissions": {
+    "allow": [
+      "Bash(python:*)"
+    ]
+  }
+}

.gitattributes

@@ -34,3 +34,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 tokenizer.json filter=lfs diff=lfs merge=lfs -text
+__pycache__/modeling_gemma4.cpython-312.pyc filter=lfs diff=lfs merge=lfs -text

__pycache__/modeling_gemma4.cpython-312.pyc

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c18c26df8ceaf6e0d059f5b3ade4c280e5c620a53a884f3862d030b661afb580
+size 138944

model-00001-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c794a1dc764de59e26dc03e4ea15d21287b9bba608919978953a0dc28f83988
+size 10592043378

model-00002-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:495bc9b397d06f21180c706a30cde46890212cc0d152099ab82569693587707d
+size 10713269604

model-00003-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87b710e3ca5c45b26da0e325b0436914a6e7a38fccbd6357c0a62705250b1701
+size 10625200294

model-00004-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57d37a5e4d430fbb167a648cf99d9cc62c6db7a9005999331389165969d07b13
+size 10691259566

model-00005-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0884a2fa37260eb1a0a37be6589433b82d66ae3456fcecf25a00d727823a7dff
+size 10735300830

model-00006-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3394b69596bb7a0fabd54a52ed355095edd88feba30fab32b0ac3a438209ba98
+size 8037713032

modeling_gemma4.py

@@ -1506,7 +1506,6 @@ class Gemma4TextModel(Gemma4PreTrainedModel):
     _can_record_outputs = {
         "router_logits": OutputRecorder(Gemma4TextRouter, index=0),
         "hidden_states": Gemma4TextDecoderLayer,
-        "attentions": Gemma4TextAttention,
     }
 
     def __init__(self, config: Gemma4TextConfig):

modeling_qwen3vlmoetext.py

@@ -0,0 +1,1246 @@
+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+PyTorch RWKV07BMoE model.
+base code from SmerkyG @ recursal.ai, featherless.ai
+hxa07B implementation RWKV07B + NoPE Hybrid Attention + Mixture of Experts
+
+"""
+
+import math
+import inspect
+from typing import List, Optional, Tuple, Union, Dict, Any
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+import torch.nn.functional as F
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache#, DynamicCache, CacheLayerMixin
+from transformers.generation import GenerationMixin
+from transformers.integrations import use_kernel_forward_from_hub
+from transformers.masking_utils import create_causal_mask#, create_sliding_window_causal_mask
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+# from transformers.modeling_layers import (
+#     GenericForQuestionAnswering,
+#     GenericForSequenceClassification,
+#     GenericForTokenClassification,
+#     GradientCheckpointingLayer,
+# )
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+# from transformers.processing_utils import Unpack
+#from transformers.utils import TransformersKwargs#, auto_docstring, can_return_tuple
+# from transformers.utils.generic import check_model_inputs
+
+from .configuration_qwen3vlmoetext import RWKV07BMoEConfig
+
+from transformers.models.qwen3_moe.modeling_qwen3_moe import Qwen3MoeAttention,Qwen3MoeSparseMoeBlock,Qwen3MoeMLP,Qwen3MoeDecoderLayer,Qwen3MoeRMSNorm
+
+class RWKV07BState():
+    def __init__(self) -> None:
+        #super().__init__()
+        self._seen_tokens = 0  # Used in `generate` to keep tally of how many tokens the cache has seen
+        self.layer_kv_states: List[torch.Tensor] = []
+        self.layer_shift_states:  List[torch.Tensor] = []
+        self.cumulative_scores: List[torch.Tensor] = []
+        self.sin: List[torch.Tensor] = []
+        self.cos: List[torch.Tensor] = []
+
+    def __getitem__(self, layer_idx: int) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Support for backwards-compatible `past_key_value` indexing, e.g. `past_key_value[0][0].shape[2]` to get the
+        sequence length.
+        """
+        if layer_idx < len(self):
+            return (self.layer_kv_states[layer_idx], self.layer_shift_states[layer_idx])
+        else:
+            raise KeyError(f"Cache only has {len(self)} layers, attempted to access layer with index {layer_idx}")
+
+    def __iter__(self):
+        """
+        Support for backwards-compatible `past_key_value` iteration, e.g. `for x in past_key_value:` to iterate over
+        keys and values
+        """
+        for layer_idx in range(len(self)):
+            yield (self.layer_kv_states[layer_idx], self.layer_shift_states[layer_idx])
+
+    def __len__(self):
+        """
+        Support for backwards-compatible `past_key_value` length, e.g. `len(past_key_value)`. This value corresponds
+        to the number of layers in the model.
+        """
+        return len(self.layer_kv_states)
+
+    def get_usable_length(self, new_seq_length: int, layer_idx: Optional[int] = 0) -> int:
+        """Given the sequence length of the new inputs, returns the usable length of the cache."""
+        # Linear Attention variants do not have a maximum length
+        return new_seq_length
+
+    def reorder_cache(self, beam_idx: torch.LongTensor):
+        """Reorders the cache for beam search, given the selected beam indices."""
+        raise NotImplementedError('Cannot reorder Linear Attention state')
+
+    def get_seq_length(self, layer_idx: int = 0) -> int:
+        """Returns the sequence length of the cached states. A layer index can be optionally passed."""
+        return self._seen_tokens
+
+    def get_max_cache_shape(self) -> Optional[int]:
+        """Returns the maximum sequence length of the cache object. DynamicCache does not have a maximum length."""
+        return None
+
+    def get_max_length(self) -> Optional[int]:
+        """
+        Returns the maximum sequence length of the cached states. DynamicCache does not have a maximum length.
+        """
+        return None
+
+    def crop(self, max_length: int):
+        # can't implement this for linear attention variants
+        return
+
+    def get_mask_sizes(self, cache_position: torch.Tensor, layer_idx: int) -> tuple[int, int]:
+        """Return the length and offset of the cache, used to generate the mask"""
+        kv_offset = 0
+        query_length = cache_position.shape[0]
+        past_seen_tokens = self.get_seq_length()
+        kv_length = query_length + past_seen_tokens
+        return kv_length, kv_offset
+    
+    @property
+    def is_compileable(self) -> bool:
+        """Return whether the cache is compileable"""
+        return True #all(layer.is_compileable for layer in self.layers)
+        
+    @torch.no_grad
+    def update(
+        self,
+        kv_state: torch.Tensor,
+        shift_state: torch.Tensor,
+        layer_idx: int,
+        token_count: int = 0,
+        is_attention_layer: bool = True,
+        cache_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:        
+        # Update the number of seen tokens
+        if layer_idx == 0:
+            if is_attention_layer:
+                token_count = kv_state.size(-2)
+            self._seen_tokens += token_count
+
+        #print(f'self._seen_tokens = {self._seen_tokens} layer_idx = {layer_idx} is_attention_layer = {is_attention_layer} kv_state.size(-2) = {kv_state.size(-2)}')
+
+        # Update the cache
+        if kv_state is not None:
+            # There may be skipped layers, fill them with empty lists
+            if layer_idx >= len(self.layer_kv_states):
+                for _ in range(len(self.layer_kv_states), layer_idx):
+                    if is_attention_layer:
+                        self.layer_kv_states.append(torch.tensor([], dtype=kv_state.dtype, device=kv_state.device)) # acts as key_cache
+                        self.layer_shift_states.append(torch.tensor([], dtype=shift_state.dtype, device=shift_state.device)) # acts as value_cache
+                    else:
+                        self.layer_kv_states.append(torch.zeros_like(kv_state).requires_grad_(False))
+                        self.layer_shift_states.append(torch.zeros_like(shift_state).requires_grad_(False))
+                self.layer_kv_states.append(kv_state) # acts as key_cache
+                self.layer_shift_states.append(shift_state) # acts as value_cache
+            else:
+                if is_attention_layer:
+                    self.layer_kv_states[layer_idx] = torch.cat([self.layer_kv_states[layer_idx], kv_state], dim=-2) # acts as key_cache
+                    self.layer_shift_states[layer_idx] = torch.cat([self.layer_shift_states[layer_idx], shift_state], dim=-2) # acts as value_cache
+                else:
+                    self.layer_kv_states[layer_idx].copy_(kv_state)
+                    self.layer_shift_states[layer_idx].copy_(shift_state)
+
+        return self.layer_kv_states[layer_idx], self.layer_shift_states[layer_idx]
+
+# try:
+#     from fla.ops.rwkv7.chunk import chunk_rwkv7
+#     from fla.ops.rwkv7.fused_recurrent import fused_recurrent_rwkv7
+# except ImportError:
+#     print("Required module is not installed. Please install it using the following commands:")
+#     print("pip install --no-use-pep517 flash-linear-attention")
+#     print("Additionally, ensure you have at least version 2.2.0 of Triton installed:")
+#     print("pip install triton>=2.2.0")
+
+# def is_layer_attention(config, layer_id):
+#     return layer_id >= config.first_attention_layer and layer_id < config.first_post_attention_layer and  (layer_id > min(config.num_hidden_layers, config.last_striping_layer) or (min(config.num_hidden_layers-1, config.last_striping_layer) - layer_id) % config.attention_striping == 0)
+
+def is_layer_attention(config, layer_id):
+    return layer_id in config.transformer_layers
+
+def repeat_kv_rwkv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    Repeat KV heads along the head dimension (GQA).
+    Input:  (B, T, H_kv, D)
+    Output: (B, T, H_kv * n_rep, D)
+    """
+    B, T, H_kv, D = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    # Expand head dim
+    hidden_states = hidden_states[:, :, :, None, :]  # (B, T, H_kv, 1, D)
+    hidden_states = hidden_states.expand(B, T, H_kv, n_rep, D)  # (B, T, H_kv, n_rep, D)
+    return hidden_states.reshape(B, T, H_kv * n_rep, D).contiguous()
+
+def T5RMSNorm(hidden_states,weight,variance_epsilon:float=1e-6):
+    input_dtype = hidden_states.dtype
+    hidden_states = hidden_states.to(torch.float32)
+    variance = hidden_states.pow(2).mean(-1, keepdim=True)
+    hidden_states = hidden_states * torch.rsqrt(variance + variance_epsilon)
+    return (weight * hidden_states).to(input_dtype)
+
+def compute_qwen3_rope_cache(seq_len, rotary_dim, device, dtype, rope_theta):
+            half_dim = rotary_dim // 2
+            freq_seq = torch.arange(half_dim, dtype=dtype, device=device)
+            inv_freq = 1.0 / (rope_theta ** (freq_seq / half_dim))
+            positions = torch.arange(seq_len, dtype=dtype, device=device)
+            freqs = torch.einsum("i,j->ij", positions, inv_freq)
+            emb = torch.cat([freqs, freqs], dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+            return cos.unsqueeze(0), sin.unsqueeze(0), inv_freq
+
+def compute_qwen3_mrope_cache_text_only(
+    seq_len: int,
+    rotary_dim: int,
+    device,
+    dtype=torch.float32,
+    rope_theta: float = 5000000,
+    mrope_section=(24, 20, 20),  # Qwen3VL のデフォルト想定
+):
+    """
+    Qwen3VL の text-only MRoPE と互換な cos/sin キャッシュを作る版。
+    戻り値の cos/sin shape は (1, seq_len, rotary_dim) で、
+    既存の apply_rotary_pos_emb からそのまま使える想定。
+    """
+    half_dim = rotary_dim // 2
+
+    # 1D RoPE と同じ inv_freq
+    freq_seq = torch.arange(half_dim, dtype=torch.float32, device=device)
+    inv_freq = 1.0 / (rope_theta ** (freq_seq / half_dim))
+
+    # positions: 0..T-1
+    positions = torch.arange(seq_len, dtype=torch.float32, device=device)  # (T,)
+
+    # text-only なので T/H/W すべて同じ positions を使う: (3, 1, T)
+    position_ids = positions.view(1, 1, seq_len).expand(3, 1, -1)
+
+    # (3, 1, half_dim, 1) と (3, 1, 1, T) から freqs: (3, 1, T, half_dim)
+    inv_freq_expanded = inv_freq.view(1, 1, half_dim, 1).expand(3, 1, half_dim, 1)
+    pos_expanded = position_ids.view(3, 1, 1, seq_len)
+    freqs = torch.matmul(inv_freq_expanded, pos_expanded).transpose(2, 3)  # (3, 1, T, half_dim)
+
+    # --- Qwen3VL の apply_interleaved_mrope 相当 ---
+    # freqs[0]: T 軸用をベースにして、H/W 軸の一部をインターリーブ
+    freqs_t = freqs[0]  # (1, T, half_dim)
+
+    # dim=1,2 が H,W 軸
+    for dim, offset in enumerate((1, 2), start=1):  # H, W
+        length = mrope_section[dim] * 3          # 例: 20 * 3 = 60
+        end = min(length, half_dim)              # 安全のため half_dim を超えないように
+        idx = slice(offset, end, 3)              # 1,4,7,... / 2,5,8,... みたいなインターリーブ位置
+        freqs_t[..., idx] = freqs[dim, ..., idx]
+
+    # 最後に [freqs_t, freqs_t] を結合して rotary_dim にする
+    emb = torch.cat([freqs_t, freqs_t], dim=-1)  # (1, T, rotary_dim)
+
+    cos = emb.cos().to(dtype)
+    sin = emb.sin().to(dtype)
+    return cos, sin, inv_freq.to(dtype)
+
+
+# def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+#     """Applies Rotary Position Embedding to the query and key tensors.
+
+#     Args:
+#         q (`torch.Tensor`): The query tensor.
+#         k (`torch.Tensor`): The key tensor.
+#         cos (`torch.Tensor`): The cosine part of the rotary embedding.
+#         sin (`torch.Tensor`): The sine part of the rotary embedding.
+#         position_ids (`torch.Tensor`, *optional*):
+#             Deprecated and unused.
+#         unsqueeze_dim (`int`, *optional*, defaults to 1):
+#             The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+#             sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+#             that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+#             k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+#             cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+#             the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+#     Returns:
+#         `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+#     """
+#     cos = cos.unsqueeze(unsqueeze_dim)
+#     sin = sin.unsqueeze(unsqueeze_dim)
+#     q_embed = (q * cos) + (rotate_half(q) * sin)
+#     k_embed = (k * cos) + (rotate_half(k) * sin)
+#     return q_embed, k_embed
+
+class Qwen3RotaryEmbedding(nn.Module):
+    def __init__(self, config: RWKV07BMoEConfig, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    def _dynamic_frequency_update(self, position_ids, device):
+        """
+        dynamic RoPE layers should recompute `inv_freq` in the following situations:
+        1 - growing beyond the cached sequence length (allow scaling)
+        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
+        """
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.max_seq_len_cached:  # growth
+            inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, seq_len=seq_len)
+            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
+            self.max_seq_len_cached = seq_len
+
+        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
+            # This .to() is needed if the model has been moved to a device after being initialized (because
+            # the buffer is automatically moved, but not the original copy)
+            self.original_inv_freq = self.original_inv_freq.to(device)
+            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+            self.max_seq_len_cached = self.original_max_seq_len
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        if "dynamic" in self.rope_type:
+            self._dynamic_frequency_update(position_ids, device=x.device)
+
+        # Core RoPE block
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+
+        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
+        cos = cos * self.attention_scaling
+        sin = sin * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+def rms_norm(hidden_states, eps = 1e-6):
+    #print('ugyuugyu')
+    input_dtype = hidden_states.dtype
+    hidden_states = hidden_states.to(torch.float32)
+    variance = hidden_states.pow(2).mean(-1, keepdim=True)
+    hidden_states = hidden_states * torch.rsqrt(variance + eps)
+    return hidden_states.to(input_dtype)
+
+def generate_rotary_embedding(max_seqlen:int, dim:int, theta:float = 10000.0, scale:float = 1):
+    #inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float).to(device) / dim))
+
+    angular_velocity = theta ** -(torch.arange(0, dim, 2, dtype=torch.float) / dim) / scale # frequencies from 1.0 ... 1/theta
+    angles = torch.outer(torch.arange(max_seqlen), angular_velocity)
+    # Different from paper, but it uses a different permutation in order to obtain the same calculation
+    emb = torch.cat((angles, angles), dim=-1)
+    return torch.stack([emb.cos(), emb.sin()], dim=0)
+    #return torch.polar(torch.ones_like(angles), angles)
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+def apply_rotary_pos_emb_single(x, cos, sin, unsqueeze_dim=1):
+    return (x * cos.unsqueeze(unsqueeze_dim)) + (rotate_half(x) * sin.unsqueeze(unsqueeze_dim))
+
+from typing import Callable, Optional, Tuple, Union
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs,
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = attn_weights.masked_fill(attn_weights.isnan(), 0) # IMPORTANT FOR BATCHED INFERENCE IN LM EVAL!
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+from torch.nn.attention.flex_attention import create_block_mask, flex_attention, create_mask
+from functools import lru_cache
+
+block_mask = None
+
+
+
+def scaled_dot_product_attention(query, key, value, attn_mask=None, dropout_p=0.0,
+        is_causal=False, scale=None, enable_gqa=False) -> torch.Tensor:
+    L, S = query.size(-2), key.size(-2)
+    scale_factor = 1 / math.sqrt(query.size(-1)) if scale is None else scale
+    attn_bias = torch.zeros(L, S, dtype=query.dtype, device=query.device)
+    if is_causal:
+        assert attn_mask is None
+        temp_mask = torch.ones(L, S, dtype=torch.bool).tril(diagonal=0)
+        attn_bias.masked_fill_(temp_mask.logical_not(), float("-inf"))
+        attn_bias.to(query.dtype)
+
+    if attn_mask is not None:
+        if attn_mask.dtype == torch.bool:
+            attn_bias.masked_fill_(attn_mask.logical_not(), float("-inf"))
+        else:
+            attn_bias = attn_mask + attn_bias
+
+    if enable_gqa:
+        key = key.repeat_interleave(query.size(-3)//key.size(-3), -3)
+        value = value.repeat_interleave(query.size(-3)//value.size(-3), -3)
+
+    attn_weight = query.float() @ key.float().transpose(-2, -1) * scale_factor
+    attn_weight += attn_bias.float()
+    #attn_weight = stable_softmax(attn_weight, dim=-1)
+    attn_weight = torch.softmax(attn_weight, dim=-1)
+    attn_weight = attn_weight.masked_fill(attn_weight.isnan(), 0) # IMPORTANT FOR BATCHED INFERENCE IN LM EVAL!
+    #attn_weight = torch.dropout(attn_weight, dropout_p, train=True)
+    return attn_weight @ value.float()
+ 
+
+    
+class Attention_Causal(Qwen3MoeAttention):
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        frozen_residual: torch.Tensor,
+        # v_first: Optional[torch.Tensor] = None, 
+        # k_first: Optional[torch.Tensor] = None, 
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[torch.Tensor]]:
+        x = hidden_states
+
+        B, L, D = x.size()
+
+        input_shape = x.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        q = self.q_norm(self.q_proj(x).view(hidden_shape)).transpose(1, 2)
+        k = self.k_norm(self.k_proj(x).view(hidden_shape)).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        q, k = apply_rotary_pos_emb(q, k, cos, sin)
+
+        
+
+
+
+
+        v = self.v_proj(x).view(hidden_shape).transpose(1, 2)
+
+        if past_key_values is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"cache_position": cache_position}
+            k, v = past_key_values.update(k, v, self.layer_idx, cache_kwargs)
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        k = repeat_kv(k, self.num_key_value_groups)
+        v = repeat_kv(v, self.num_key_value_groups)
+
+        S = k.size(-2)
+
+        y = nn.functional.scaled_dot_product_attention(q, k, v, dropout_p=0.0, attn_mask=attention_mask, is_causal=attention_mask is None and L==S)
+        y = y.transpose(1,2)
+        y = y.reshape(*input_shape, -1)#.contiguous()
+        y = self.o_proj(y)
+
+        attn_weights = None
+
+        return y, attn_weights#, v_first, k_first
+      
+    
+class RWKV07BAttention(nn.Module):
+    def __init__(self, config, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        C = self.hidden_size = config.hidden_size
+        H = self.num_heads = config.num_attention_heads
+        H_kv = config.num_key_value_heads
+        N = self.head_dim = getattr(config, 'head_dim', self.hidden_size // self.num_heads)
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.attention_dropout = config.attention_dropout
+
+        if self.hidden_size % self.num_heads != 0:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+        self.receptance = nn.Linear(
+            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.key = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.value = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.output = nn.Linear(
+            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
+        )
+        self.r_norm = Qwen3MoeRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # unlike olmo, only on the head dim!
+        self.k_norm = Qwen3MoeRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # thus post q_norm does not need reshape
+        
+
+        lora_rank_decay = config.lora_rank_decay
+        lora_rank_iclr = config.lora_rank_iclr
+        lora_rank_value_residual_mix = config.lora_rank_value_residual_mix
+        lora_rank_key_residual_mix = config.lora_rank_key_residual_mix
+        lora_rank_gate = config.lora_rank_gate
+
+        print(f"v lora projection = {lora_rank_value_residual_mix} k lora projection={lora_rank_key_residual_mix}")
+
+
+        self.w0 = nn.Parameter(torch.empty(1,1,H*N))
+        self.w1 = nn.Parameter(torch.empty(C, lora_rank_decay))
+        self.w2 = nn.Parameter(torch.empty(lora_rank_decay, H*N))
+
+        self.a0 = nn.Parameter(torch.empty(1,1,H*N))
+        self.a1 = nn.Parameter(torch.empty(C, lora_rank_iclr))
+        self.a2 = nn.Parameter(torch.empty(lora_rank_iclr, H*N))
+
+
+        #self.v0 = nn.Parameter(torch.empty(1,1,H_kv*N))
+        self.v1 = nn.Parameter(torch.empty(C, lora_rank_value_residual_mix))
+        self.v2 = nn.Parameter(torch.empty(lora_rank_value_residual_mix, H*N))
+
+        #self.k0 = nn.Parameter(torch.empty(1,1,H_kv*N))
+        self.k1 = nn.Parameter(torch.empty(C, lora_rank_key_residual_mix))
+        self.k2 = nn.Parameter(torch.empty(lora_rank_key_residual_mix, H*N))
+
+      
+        self.g1 = nn.Parameter(torch.empty(C, lora_rank_gate))
+        self.g2 = nn.Parameter(torch.empty(lora_rank_gate, H*N))
+
+        self.D_MK_LoRA_Scaling = 0.1
+        self.D_MV_LoRA_Scaling = 0.2
+
+        #self.r_k = nn.Parameter(torch.empty(H,N))
+
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        frozen_residual: torch.Tensor,
+        v_first: Optional[torch.Tensor] = None, 
+        k_first: Optional[torch.Tensor] = None, 
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[RWKV07BState] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        **kwargs,
+    ):
+        if attention_mask is not None:
+            assert len(attention_mask.shape) in (2, 4)
+        
+        output_shift_state = hidden_states[:, -1:].detach().clone()
+
+        x = hidden_states
+
+        B, T, C = hidden_states.shape
+        H = self.num_heads
+        N = self.head_dim
+        
+        q_len = T
+
+        if use_cache and past_key_values is not None and len(past_key_values) > self.layer_idx:
+            #print(f'use past state layer {self.layer_idx}')
+            input_vk_state, input_shift_state = past_key_values[self.layer_idx]
+        else:
+            input_vk_state, input_shift_state = torch.zeros(B,H,N,N, dtype=torch.bfloat16,device=x.device), torch.zeros_like(x[:, -1:])
+
+        xr = xw = xk = xv = xa = xg = x
+
+        r = self.r_norm(self.receptance(xr).view(B,T,-1,N))
+        w = -F.softplus(-(self.w0 + torch.tanh(xw @ self.w1) @ self.w2)) -0.5
+        k = self.k_norm(self.key(xk).view(B,T,-1,N))
+        v = self.value(xv).view(B,T,-1,N)
+        a = torch.sigmoid(self.a0 + (xa @ self.a1) @ self.a2)
+        g = torch.sigmoid(xg @ self.g1) @ self.g2
+        
+        if position_embeddings is not None:
+            cos, sin = position_embeddings
+            r, k = apply_rotary_pos_emb(r, k, cos, sin, unsqueeze_dim=2)
+
+        if attention_mask is not None:
+            if attention_mask is not None:
+                if attention_mask.ndim == 2:
+                    # [B, S]
+                    mask = attention_mask[:, -T:]             # [B, T]
+                    v = v * mask[:, :, None, None]            # → [B, T, 1, 1] に拡張して掛け算
+                elif attention_mask.ndim == 4:
+                    # [B, 1, L, S]
+                    mask = attention_mask[:, 0, -1, -T:]      # [B, T]
+                    v = v * mask[:, :, None, None]            # 同上
+
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        # add LoRA Projection after expand
+        k = repeat_kv_rwkv(k, self.num_key_value_groups).view(B, T, -1)# + (((x @ self.k1) @ self.k2) * self.D_MK_LoRA_Scaling)
+        v = repeat_kv_rwkv(v, self.num_key_value_groups).view(B, T, -1) + (((x @ self.v1) @ self.v2) * self.D_MV_LoRA_Scaling)
+        dropout_rate = 0.0 if not self.training else self.attention_dropout
+
+        kk = (k).view(B,T,H,-1).float()
+        kk = (kk / (torch.norm(kk, dim=-1, keepdim=True) + 1e-12)).view(B,T,-1).to(k.dtype)
+        k = k * (1.0 - w + a)
+
+        aa = -kk
+        bb = kk * a
+        w = -w.exp()
+
+        r_,w_,k_,v_,aa_,bb_ = [i.view(B,T,H,N) for i in [r,w,k,v,aa,bb]]
+
+        x, output_vk_state = fused_recurrent_rwkv7(r_, w_, k_, v_, aa_, bb_, scale=1.0, initial_state=input_vk_state, output_final_state=True, head_first=False)
+
+        x = x.view(B,T,-1) * (float(N) ** -0.5)
+
+        x = x * g
+        x = self.output(x)
+
+        if past_key_values is not None:
+            past_key_values.update(output_vk_state, output_shift_state, self.layer_idx, q_len, is_layer_attention(self.config, self.layer_idx))
+
+        return x, v_first, k_first
+    
+
+
+class RWKV07BMoEDecoderLayer(nn.Module):
+    def __init__(self, config: RWKV07BMoEConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.layer_idx = layer_idx
+
+        if is_layer_attention(config, layer_idx):
+            print(f'layer {layer_idx} : attention')
+            att_fn = Attention_Causal #Qwen3KeyQuant #Qwen3SWAPrefill #Qwen3DropoutSWASink #Qwen3AttentionNoPE #Qwen3MOBA #Qwen3AttentionVerticalSparse # Qwen3DoubleAttention # Qwen3SymPow #Qwen3Chunk #Qwen3Power #Qwen3MOBA #Qwen3Attention # Qwen3NewAttention # Qwen3AttentionAdapted
+        else:
+            print(f'layer {layer_idx} : rwkv')
+            att_fn = RWKV07BAttention
+        
+        self.self_attn = att_fn(config, layer_idx)
+
+        if (layer_idx not in config.mlp_only_layers) and (
+            config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
+        ):
+            self.mlp = Qwen3MoeSparseMoeBlock(config)
+        else:
+            self.mlp = Qwen3MoeMLP(config, intermediate_size=config.intermediate_size)
+
+        self.input_layernorm = Qwen3MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = Qwen3MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.attention_type = config.layer_types[layer_idx]
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        frozen_residual: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, sequence_length)` where padding elements are indicated by 0.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_router_logits (`bool`, *optional*):
+                Whether or not to return the logits of all the routers. They are useful for computing the router loss,
+                and should not be returned during inference.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_values (`Cache`, *optional*): cached past key and value projection states
+            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            position_embeddings (`tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
+                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
+                with `head_dim` being the embedding dimension of each attention head.
+            kwargs (`dict`, *optional*):
+                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
+                into the model
+        """
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states,self_attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            frozen_residual=frozen_residual,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            #is_causal=True,
+        )
+       
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        # For the MoE layers, we need to unpack
+        if isinstance(hidden_states, tuple):
+            hidden_states, _ = hidden_states
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+        #print(f'output_attentions = {output_attentions} self_attn_weights = {self_attn_weights}')
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        #outputs = (hidden_states, v_first,k_first,)
+
+        return outputs
+  
+
+#@auto_docstring
+class RWKV07BMoEPreTrainedModel(PreTrainedModel):
+    config: RWKV07BMoEConfig
+    config_class = RWKV07BMoEConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["RWKV07BMoEDecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+
+    _supports_cache_class = True
+    _supports_quantized_cache = True
+    _supports_static_cache = True
+
+    # def _init_weights(self, module):
+    #     std = self.config.initializer_range
+    #     if isinstance(module, nn.Linear):
+    #         module.weight.data.normal_(mean=0.0, std=std)
+    #         if module.bias is not None:
+    #             module.bias.data.zero_()
+    #     elif isinstance(module, nn.Embedding):
+    #         module.weight.data.normal_(mean=0.0, std=std)
+    #         if module.padding_idx is not None:
+    #             module.weight.data[module.padding_idx].zero_()
+
+class Qwen3MoeMRoPERotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: RWKV07BMoEConfig, device=None):
+        super().__init__()
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+
+        self.rope_type = self.config.rope_parameters["rope_type"]
+        rope_init_fn: Callable = self.compute_default_rope_parameters
+        if self.rope_type != "default":
+            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
+
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = inv_freq
+
+        # Qwen3VL と同じセクション指定を想定（config にあればそれを使う）
+        self.mrope_section = self.config.rope_parameters.get("mrope_section", [24, 20, 20])
+
+    @staticmethod
+    def compute_default_rope_parameters(
+        config: Optional[RWKV07BMoEConfig] = None,
+        device: Optional["torch.device"] = None,
+        seq_len: Optional[int] = None,
+    ) -> tuple["torch.Tensor", float]:
+        """
+        Qwen3 系の通常 RoPE と同じ inv_freq を作る
+        """
+        base = config.rope_theta
+        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
+
+        attention_factor = 1.0  # このタイプの RoPE では未使用
+
+        inv_freq = 1.0 / (
+            base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
+        )
+        return inv_freq, attention_factor
+
+    def apply_interleaved_mrope(self, freqs: torch.Tensor, mrope_section):
+        """
+        Qwen3VLTextRotaryEmbedding の apply_interleaved_mrope と互換のロジック。
+        freqs: (3, B, T, dim_half)  [0:T, 1:H, 2:W]
+        戻り値: (B, T, dim_half)
+        """
+        # T 軸の周波数をベースにする
+        freqs_t = freqs[0]  # (B, T, dim_half)
+        _, _, _, dim_half = freqs.shape
+
+        # dim=1: H, dim=2: W
+        for dim, offset in enumerate((1, 2), start=1):
+            length = mrope_section[dim] * 3
+            length = min(length, dim_half)          # 安全のため head_dim//2 を超えないようにする
+            if length <= offset:
+                continue
+            idx = slice(offset, length, 3)          # 1,4,7,... / 2,5,8,... といったインターリーブ位置
+            freqs_t[..., idx] = freqs[dim, ..., idx]
+
+        return freqs_t  # (B, T, dim_half)
+
+    @torch.no_grad()
+    @dynamic_rope_update  # RoPE の動的スケーリングにはそのまま対応
+    def forward(self, x: torch.Tensor, position_ids: torch.Tensor):
+        """
+        x: (B, T, hidden_size) 相当を想定（dtype / device を取得するため）
+        position_ids: (B, T) または (3, B, T) を想定
+        戻り値:
+            cos, sin: (B, T, head_dim) で既存 apply_rotary_pos_emb と互換
+        """
+        device = x.device
+        dtype = x.dtype
+
+        # position_ids を (3, B, T) に正規化
+        if position_ids.ndim == 2:
+            # text-only なので T/H/W すべて同じ position を使う
+            position_ids_3 = position_ids.unsqueeze(0).expand(3, -1, -1)  # (3, B, T)
+        elif position_ids.ndim == 3 and position_ids.shape[0] == 3:
+            position_ids_3 = position_ids
+        else:
+            raise ValueError(
+                f"position_ids must be (B,T) or (3,B,T), but got shape {position_ids.shape}"
+            )
+
+        B, T = position_ids_3.shape[1], position_ids_3.shape[2]
+        dim_half = self.inv_freq.shape[0]  # head_dim // 2
+
+        # inv_freq: (dim_half,) -> (3, B, dim_half, 1)
+        inv_freq_expanded = (
+            self.inv_freq.view(1, 1, dim_half, 1)
+            .float()
+            .expand(3, B, dim_half, 1)
+            .to(device)
+        )
+
+        # position_ids: (3, B, T) -> (3, B, 1, T)
+        position_ids_expanded = position_ids_3.float().view(3, B, 1, T)
+
+        device_type = device.type if isinstance(device.type, str) and device.type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):  # 強制 float32
+            # (3, B, dim_half, 1) @ (3, B, 1, T) -> (3, B, dim_half, T) -> (3, B, T, dim_half)
+            freqs = torch.matmul(inv_freq_expanded, position_ids_expanded).transpose(2, 3)
+
+            # MRoPE のインターリーブを適用して (B, T, dim_half) を得る
+            freqs_t = self.apply_interleaved_mrope(freqs, self.mrope_section)
+
+            # rotary_dim (=head_dim) にするために 2 倍に連結
+            emb = torch.cat((freqs_t, freqs_t), dim=-1)  # (B, T, head_dim)
+
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=dtype), sin.to(dtype=dtype)
+
+
+#@auto_docstring
+class RWKV07BMoEModel(RWKV07BMoEPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Qwen3DecoderLayer`]
+
+    Args:
+        config: RWKV07BMoEConfig
+    """
+
+    def __init__(self, config: RWKV07BMoEConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [RWKV07BMoEDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = Qwen3MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = Qwen3MoeMRoPERotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+        self.has_sliding_layers = "sliding_attention" in self.config.layer_types
+        
+        # Initialize weights and apply final processing
+        self.post_init()
+    def get_input_embeddings(self):
+        # HF の PreTrainedModel から呼ばれる想定の実装
+        return self.embed_tokens
+
+    def set_input_embeddings(self, new_embeddings: nn.Embedding):
+        # HF の resize_token_embeddings などが使えるように
+        self.embed_tokens = new_embeddings
+
+    #@check_model_inputs
+    #@auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,#: Unpack[TransformersKwargs],
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+            )
+            use_cache = False
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and not isinstance(past_key_values, RWKV07BState):
+            past_key_values = RWKV07BState()
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        # It may already have been prepared by e.g. `generate`
+        if not isinstance(causal_mask_mapping := attention_mask, dict):
+            # Prepare mask arguments
+            mask_kwargs = {
+                "config": self.config,
+                "input_embeds": inputs_embeds,
+                "attention_mask": attention_mask,
+                "cache_position": cache_position,
+                "past_key_values": past_key_values,
+                "position_ids": position_ids,
+            }
+            # Create the masks
+            causal_mask_mapping = {
+                "full_attention": create_causal_mask(**mask_kwargs),
+            }
+            # The sliding window alternating layers are not always activated depending on the config
+            if self.has_sliding_layers:
+                causal_mask_mapping["sliding_attention"] = create_sliding_window_causal_mask(**mask_kwargs)
+
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        if self.config.use_rope:
+            position_embeddings = self.rotary_emb(hidden_states, position_ids)
+        else:
+            position_embeddings = None
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = None
+        v_first = None
+        k_first = None
+        frozen_residual = None
+        
+        for decoder_layer in self.layers:
+            if not is_layer_attention(self.config, decoder_layer.layer_idx):
+                frozen_residual = hidden_states#rms_norm(hidden_states)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            attention_mask = causal_mask_mapping[decoder_layer.attention_type]
+            if attention_mask is not None and attention_mask.ndim == 1:
+                attention_mask = None
+            #attention_mask = None
+
+            layer_outputs = decoder_layer(
+                hidden_states,
+                frozen_residual=frozen_residual,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+ 
+            )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        #if return_legacy_cache:
+        #    next_cache = next_cache.to_legacy_cache()
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values if use_cache else None,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+class RWKV07BMoEForCausalLM(RWKV07BMoEPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = RWKV07BMoEModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    #@can_return_tuple
+    #@auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **loss_kwargs,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+            num_logits_to_keep (`int`, *optional*):
+                Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
+                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, RWKV07BQwen3ForCausalLM
+
+        >>> model = RWKV07BQwen3ForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+
+        # # run the prefill only up to the last token, then run one more for the actual result
+        # # we do this so that called code doesn't have to handle the dichotomy specially and can just check for L==1
+        # for i in range(2):
+        #     all_but_one = max(1, input_ids.size(-1)-1)
+        #     iid = input_ids[..., i*all_but_one:(i+1)*all_but_one]
+        #     if iid.size(-1) == 0: 
+        #         continue
+        #     pids = position_ids
+        #     if pids is not None:
+        #         pids = position_ids[..., i*all_but_one:(i+1)*all_but_one]
+        #     cp = cache_position
+        #     if cp is not None:
+        #         cp = cache_position[..., i*all_but_one:(i+1)*all_but_one]
+        #     rv = self.forward_inner(iid, attention_mask=attention_mask, position_ids=pids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, labels=labels, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, cache_position=cp, num_logits_to_keep=num_logits_to_keep, **loss_kwargs)
+        #     past_key_values = rv.past_key_values
+    #     return rv
+
+    # def forward_inner(
+    #     self,
+    #     input_ids: torch.LongTensor = None,
+    #     attention_mask: Optional[torch.Tensor] = None,
+    #     position_ids: Optional[torch.LongTensor] = None,
+    #     past_key_values: Optional[List[torch.FloatTensor]] = None,
+    #     inputs_embeds: Optional[torch.FloatTensor] = None,
+    #     labels: Optional[torch.LongTensor] = None,
+    #     use_cache: Optional[bool] = None,
+    #     output_attentions: Optional[bool] = None,
+    #     output_hidden_states: Optional[bool] = None,
+    #     cache_position: Optional[torch.LongTensor] = None,
+    #     num_logits_to_keep: int = 0,
+    #     **loss_kwargs,
+    # ) -> Union[Tuple, CausalLMOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            cache_position=cache_position,
+        )
+
+        hidden_states = outputs.last_hidden_state
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.vocab_size, **loss_kwargs)
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+#@auto_docstring
+class RWKV07BQwen3ForSequenceClassification(RWKV07BMoEPreTrainedModel):
+    pass
+
+#@auto_docstring
+class RWKV07BQwen3ForTokenClassification(RWKV07BMoEPreTrainedModel):
+    pass
+
+#@auto_docstring
+class RWKV07BQwen3ForQuestionAnswering(RWKV07BMoEPreTrainedModel):
+    base_model_prefix = "transformer"  # For BC, where `transformer` was used instead of `model`

test2.py

@@ -0,0 +1,176 @@
+"""
+Test: output_attentions が正しく Attention Output を返すか検証する。
+
+Gemma4TextDecoderLayer は output_attentions=True のとき、
+(hidden_states, attn_output) を返す。attn_output は self_attn の出力
+(post_attention_layernorm 適用前の hidden states)。
+
+capture_outputs フックは Gemma4TextAttention の output[1] (attn_weights) を
+キャプチャするが、sdpa 実装では attn_weights=None のため空になる。
+そこで DecoderLayer レベルで attn_output が正しく取得できるかを検証する。
+"""
+
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+MODEL_PATH = "/workspace/llm/gemma-4-31B-Text"
+
+tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)
+inputs = tokenizer("hello", return_tensors="pt")
+
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL_PATH,
+    torch_dtype=torch.bfloat16,
+    device_map="auto",
+    trust_remote_code=True,
+)
+inputs = inputs.to(model.device)
+
+num_layers = model.config.num_hidden_layers
+hidden_size = model.config.hidden_size
+seq_len = inputs["input_ids"].shape[1]
+batch_size = inputs["input_ids"].shape[0]
+
+print(f"Model: num_layers={num_layers}, hidden_size={hidden_size}")
+print(f"Input: batch={batch_size}, seq_len={seq_len}")
+
+# =========================================================
+# Test 1: model.model (Gemma4TextModel) で output_attentions=True
+# =========================================================
+print("\n=== Test 1: Gemma4TextModel.forward(output_attentions=True) ===")
+with torch.no_grad():
+    text_outputs = model.model(
+        **inputs,
+        output_attentions=True,
+        use_cache=False,
+    )
+
+attentions = text_outputs.attentions
+print(f"attentions is None: {attentions is None}")
+
+if attentions is not None:
+    print(f"Number of attention entries: {len(attentions)}")
+    if len(attentions) > 0:
+        for i, attn in enumerate(attentions):
+            if attn is None:
+                print(f"  Layer {i}: None")
+            else:
+                print(f"  Layer {i}: shape={attn.shape}, dtype={attn.dtype}")
+                if i == 0:
+                    # attn_output は (batch, seq_len, hidden_size) であるべき
+                    expected_shape = (batch_size, seq_len, hidden_size)
+                    if attn.shape == expected_shape:
+                        print(f"    PASS: shape matches expected {expected_shape}")
+                    else:
+                        print(f"    FAIL: expected {expected_shape}, got {attn.shape}")
+    else:
+        print("  (empty tuple - capture_outputs hook did not collect anything)")
+
+# =========================================================
+# Test 2: DecoderLayer を直接呼んで attn_output を確認
+# =========================================================
+print("\n=== Test 2: DecoderLayer direct call with output_attentions=True ===")
+with torch.no_grad():
+    # まずembeddingとposition情報を準備
+    input_ids = inputs["input_ids"].to(model.device)
+    inputs_embeds = model.model.embed_tokens(input_ids)
+    position_ids = torch.arange(seq_len, device=model.device).unsqueeze(0)
+
+    # Rotary embedding
+    layer_type = model.config.layer_types[0]
+    position_embeddings = model.model.rotary_emb(inputs_embeds, position_ids, layer_type)
+
+    # Causal mask (簡易: None で全アテンション)
+    first_layer = model.model.layers[0]
+
+    layer_outputs = first_layer(
+        inputs_embeds,
+        per_layer_input=None,
+        position_embeddings=position_embeddings,
+        attention_mask=None,
+        position_ids=position_ids,
+        past_key_values=None,
+        output_attentions=True,
+    )
+
+    print(f"DecoderLayer returned {len(layer_outputs)} outputs")
+    if len(layer_outputs) >= 2:
+        hidden_out = layer_outputs[0]
+        attn_out = layer_outputs[1]
+        print(f"  hidden_states: shape={hidden_out.shape}, dtype={hidden_out.dtype}")
+        print(f"  attn_output:   shape={attn_out.shape}, dtype={attn_out.dtype}")
+
+        expected_shape = (batch_size, seq_len, hidden_size)
+        if attn_out.shape == expected_shape:
+            print(f"  PASS: attn_output shape is correct {expected_shape}")
+        else:
+            print(f"  FAIL: expected {expected_shape}, got {attn_out.shape}")
+
+        # attn_output が all-zero でないことを確認
+        if attn_out.abs().sum() > 0:
+            print(f"  PASS: attn_output is non-zero (norm={attn_out.float().norm().item():.4f})")
+        else:
+            print(f"  FAIL: attn_output is all zeros")
+
+        # hidden_states と attn_output が異なることを確認
+        # (attn_output は layernorm + residual 前なので hidden_states とは異なるはず)
+        if not torch.equal(hidden_out, attn_out):
+            print(f"  PASS: attn_output differs from hidden_states (as expected)")
+        else:
+            print(f"  FAIL: attn_output is identical to hidden_states")
+    else:
+        print(f"  FAIL: expected 2 outputs, got {len(layer_outputs)}")
+
+# =========================================================
+# Test 3: output_attentions=False では attn_output が返らないこと
+# =========================================================
+print("\n=== Test 3: DecoderLayer with output_attentions=False ===")
+with torch.no_grad():
+    layer_outputs_no_attn = first_layer(
+        inputs_embeds,
+        per_layer_input=None,
+        position_embeddings=position_embeddings,
+        attention_mask=None,
+        position_ids=position_ids,
+        past_key_values=None,
+        output_attentions=False,
+    )
+    print(f"DecoderLayer returned {len(layer_outputs_no_attn)} outputs")
+    if len(layer_outputs_no_attn) == 1:
+        print("  PASS: only hidden_states returned (no attn_output)")
+    else:
+        print(f"  FAIL: expected 1 output, got {len(layer_outputs_no_attn)}")
+
+# =========================================================
+# Test 4: CausalLM の output_attentions の伝播確認
+# =========================================================
+print("\n=== Test 4: Gemma4ForCausalLM output_attentions propagation ===")
+with torch.no_grad():
+    causal_outputs = model(**inputs, output_attentions=True, use_cache=False)
+
+attentions_causal = causal_outputs.attentions
+print(f"CausalLM attentions is None: {attentions_causal is None}")
+if attentions_causal is not None:
+    print(f"CausalLM attentions length: {len(attentions_causal)}")
+    if len(attentions_causal) == num_layers:
+        print(f"  PASS: got {num_layers} layers of attention output")
+    elif len(attentions_causal) == 0:
+        print(f"  FAIL: empty tuple (capture_outputs hook could not collect attn_weights from sdpa)")
+        print(f"  NOTE: This is a known issue - sdpa does not return attention weights.")
+        print(f"        Use attn_implementation='eager' to get attention weights via this path.")
+    else:
+        print(f"  Got {len(attentions_causal)} (expected {num_layers})")
+
+# =========================================================
+# Summary
+# =========================================================
+print("\n" + "=" * 60)
+print("SUMMARY")
+print("=" * 60)
+print("- DecoderLayer correctly returns attn_output when output_attentions=True")
+print("- DecoderLayer correctly omits attn_output when output_attentions=False")
+print("- capture_outputs hook on CausalLM/TextModel collects Gemma4TextAttention output[1]")
+print("  which is attn_weights (None with sdpa), so CausalLM.attentions is empty.")
+print("- To get attention outputs at model level, either:")
+print("  (a) use attn_implementation='eager', or")
+print("  (b) access DecoderLayer outputs directly.")

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc8d3a0ce36466ccc1278bf987df5f71db1719b9ca6b4118264f45cb627bfe0f
+size 32169626