Laguna-XS v1.4 base (step 1207000) bf16

config.json

@@ -0,0 +1,151 @@
+{
+  "architectures": [
+    "LagunaForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_laguna.LagunaConfig",
+    "AutoModelForCausalLM": "modeling_laguna.LagunaForCausalLM"
+  },
+  "model_type": "laguna",
+  "vocab_size": 100352,
+  "hidden_size": 2048,
+  "intermediate_size": 8192,
+  "num_hidden_layers": 40,
+  "num_attention_heads": 48,
+  "num_key_value_heads": 8,
+  "head_dim": 128,
+  "max_position_embeddings": 131072,
+  "qkv_bias": false,
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "rms_norm_eps": 1e-06,
+  "num_experts": 256,
+  "num_experts_per_tok": 8,
+  "moe_intermediate_size": 512,
+  "shared_expert_intermediate_size": 512,
+  "norm_topk_prob": true,
+  "router_aux_loss_coef": 0.001,
+  "decoder_sparse_step": 1,
+  "mlp_only_layers": [
+    0
+  ],
+  "bos_token_id": 2,
+  "eos_token_id": [
+    2,
+    24
+  ],
+  "pad_token_id": 9,
+  "tie_word_embeddings": false,
+  "use_cache": true,
+  "torch_dtype": "bfloat16",
+  "gating": "per-head",
+  "sliding_window": 512,
+  "rope_parameters": {
+    "rope_theta": 500000.0,
+    "rope_type": "yarn",
+    "factor": 32.0,
+    "original_max_position_embeddings": 4096,
+    "beta_slow": 1.0,
+    "beta_fast": 64.0,
+    "attention_factor": 1.0
+  },
+  "layer_types": [
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention"
+  ],
+  "num_attention_heads_per_layer": [
+    48,
+    64,
+    64,
+    64,
+    48,
+    64,
+    64,
+    64,
+    48,
+    64,
+    64,
+    64,
+    48,
+    64,
+    64,
+    64,
+    48,
+    64,
+    64,
+    64,
+    48,
+    64,
+    64,
+    64,
+    48,
+    64,
+    64,
+    64,
+    48,
+    64,
+    64,
+    64,
+    48,
+    64,
+    64,
+    64,
+    48,
+    64,
+    64,
+    64
+  ],
+  "swa_rope_parameters": {
+    "rope_theta": 10000.0,
+    "rope_type": "linear",
+    "factor": 1.0,
+    "partial_rotary_factor": 1.0
+  },
+  "moe_router_use_sigmoid": true,
+  "moe_apply_router_weight_on_input": false,
+  "moe_shared_gate": false,
+  "moe_routed_scaling_factor": 2.5,
+  "qk_norm_type": "rmsnorm",
+  "norm_type": "rmsnorm",
+  "rope_style": "rotate-half",
+  "partial_rotary_factor": 0.5,
+  "swa_attention_sink_enabled": false
+}

configuration_laguna.py

@@ -0,0 +1,194 @@
+# Copyright 2025 Poolside and the HuggingFace Inc. team. All rights reserved.
+#
+# 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.
+from transformers.configuration_utils import PreTrainedConfig
+from transformers.modeling_rope_utils import RopeParameters
+
+
+class LagunaConfig(PreTrainedConfig):
+    r"""
+    Configuration class for Laguna model.
+
+    Laguna is Poolside's MoE architecture with:
+    - Attention output gating (softplus gate)
+    - Sigmoid routing instead of softmax
+    - No QKV bias
+    - Explicit head_dim parameter
+
+    Args:
+        head_dim (`int`, *optional*, defaults to 128):
+            Dimension of attention heads. Laguna uses explicit head_dim rather than
+            computing it from hidden_size // num_attention_heads.
+        qkv_bias (`bool`, *optional*, defaults to `False`):
+            Whether to add bias to QKV projections. Laguna uses no QKV bias.
+        attention_bias (`bool`, *optional*, defaults to `False`):
+            Whether to add bias to attention output projection. Laguna uses no attention bias.
+        gating (`bool`, *optional*, defaults to `True`):
+            Whether to use softplus output gating on attention. When True, a g_proj linear
+            layer is added and attn_output = attn_output * softplus(g_proj(x)).
+        sliding_window (`int`, *optional*):
+            Sliding window attention size. Used by layers whose type in ``layer_types``
+            is ``"sliding_attention"``. When ``None``, all layers use full attention.
+        layer_types (`list[str]`, *optional*):
+            Per-layer attention type. Each element should be ``"sliding_attention"`` or
+            ``"full_attention"``. Length must equal ``num_hidden_layers``. When ``None``,
+            all layers default to global attention.
+        swa_attention_sink_enabled (`bool`, *optional*, defaults to `False`):
+            Whether to enable learnable attention sinks on sliding-window attention layers.
+            When enabled, a per-head bias parameter is added that allows the model to attend
+            to position 0 even when it falls outside the sliding window.
+        swa_rope_parameters (`RopeParameters`, *optional*):
+            Separate RoPE configuration for sliding-window attention layers. When ``None``,
+            SWA layers use the same RoPE as global attention layers.
+        vocab_size (`int`, *optional*, defaults to 100352):
+            Vocabulary size of the Laguna model.
+        hidden_size (`int`, *optional*, defaults to 2048):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 8192):
+            Dimension of the MLP representations for dense layers.
+        num_hidden_layers (`int`, *optional*, defaults to 48):
+            Number of hidden layers in the Transformer.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads.
+        num_key_value_heads (`int`, *optional*, defaults to 8):
+            Number of key-value heads for GQA.
+        max_position_embeddings (`int`, *optional*, defaults to 4096):
+            Maximum sequence length.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-6):
+            Epsilon for RMSNorm layers.
+        num_experts (`int`, *optional*, defaults to 256):
+            Number of routed experts.
+        num_experts_per_tok (`int`, *optional*, defaults to 16):
+            Number of experts selected per token (top-k).
+        moe_intermediate_size (`int`, *optional*, defaults to 1024):
+            Intermediate size of routed experts.
+        shared_expert_intermediate_size (`int`, *optional*, defaults to 1024):
+            Intermediate size of the shared expert.
+        norm_topk_prob (`bool`, *optional*, defaults to `True`):
+            Whether to normalize top-k routing probabilities.
+        decoder_sparse_step (`int`, *optional*, defaults to 1):
+            Frequency of MoE layers (1 = every layer is MoE after mlp_only_layers).
+        mlp_only_layers (`list[int]`, *optional*, defaults to `[0]`):
+            Layer indices that use dense MLP instead of MoE.
+        router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
+            Auxiliary loss coefficient for load balancing.
+        rope_parameters (`RopeParameters`, *optional*):
+            RoPE configuration. Defaults to rope_theta=500000.0.
+    """
+
+    model_type = "laguna"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.g_proj": "colwise",  # Laguna-specific gating projection
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        vocab_size: int = 100352,
+        hidden_size: int = 2048,
+        intermediate_size: int = 8192,
+        num_hidden_layers: int = 48,
+        num_attention_heads: int = 32,
+        num_key_value_heads: int = 8,
+        head_dim: int = 128,
+        qkv_bias: bool = False,
+        attention_bias: bool = False,
+        gating: bool | str = True,
+        hidden_act: str = "silu",
+        max_position_embeddings: int = 4096,
+        initializer_range: float = 0.02,
+        rms_norm_eps: float = 1e-6,
+        use_cache: bool = True,
+        tie_word_embeddings: bool = False,
+        rope_parameters: RopeParameters | dict[str, RopeParameters] | None = None,
+        partial_rotary_factor: float = 1.0,
+        attention_dropout: float = 0.0,
+        sliding_window: int | None = None,
+        layer_types: list[str] | None = None,
+        swa_attention_sink_enabled: bool = False,
+        swa_rope_parameters: RopeParameters | None = None,
+        num_attention_heads_per_layer: list[int] | None = None,
+        num_experts: int = 256,
+        num_experts_per_tok: int = 16,
+        moe_intermediate_size: int = 1024,
+        shared_expert_intermediate_size: int = 1024,
+        norm_topk_prob: bool = True,
+        decoder_sparse_step: int = 1,
+        mlp_only_layers: list[int] | None = None,
+        router_aux_loss_coef: float = 0.001,
+        output_router_logits: bool = False,
+        moe_routed_scaling_factor: float = 1.0,
+        moe_apply_router_weight_on_input: bool = False,
+        **kwargs,
+    ):
+        # Default mlp_only_layers: first layer is dense (moe_first_k_dense_replace=1)
+        if mlp_only_layers is None:
+            mlp_only_layers = [0]
+
+        # Default rope_parameters with Laguna's theta
+        if rope_parameters is None:
+            rope_parameters = {"rope_type": "default", "rope_theta": 500000.0}
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.head_dim = head_dim
+        self.qkv_bias = qkv_bias
+        self.attention_bias = attention_bias
+        self.gating = gating
+        self.hidden_act = hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_parameters = rope_parameters
+        self.partial_rotary_factor = partial_rotary_factor
+        self.attention_dropout = attention_dropout
+        # Sliding window attention arguments
+        self.sliding_window = sliding_window
+        self.layer_types = layer_types
+        self.swa_attention_sink_enabled = swa_attention_sink_enabled
+        self.swa_rope_parameters = swa_rope_parameters
+        self.num_attention_heads_per_layer = num_attention_heads_per_layer
+        # MoE arguments
+        self.num_experts = num_experts
+        self.num_experts_per_tok = num_experts_per_tok
+        self.moe_intermediate_size = moe_intermediate_size
+        self.shared_expert_intermediate_size = shared_expert_intermediate_size
+        self.norm_topk_prob = norm_topk_prob
+        self.decoder_sparse_step = decoder_sparse_step
+        self.mlp_only_layers = mlp_only_layers
+        self.router_aux_loss_coef = router_aux_loss_coef
+        self.output_router_logits = output_router_logits
+        self.moe_routed_scaling_factor = moe_routed_scaling_factor
+        self.moe_apply_router_weight_on_input = moe_apply_router_weight_on_input
+
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
+
+
+__all__ = ["LagunaConfig"]

generation_config.json

@@ -0,0 +1,12 @@
+{
+  "bos_token_id": 2,
+  "do_sample": true,
+  "eos_token_id": [
+    2,
+    24
+  ],
+  "max_new_tokens": 2048,
+  "pad_token_id": 9,
+  "temperature": 0.7,
+  "top_p": 0.9
+}

model-00001-of-00014.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:096bec47fccb4e593cda439e96c441b4df24da603f6996ad4cc2f42b07b62979
+size 5120041576

model-00002-of-00014.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b033cde77d0dfc467217228ac1fe56955da6f6f0539d217c0e87bc9c6141a02
+size 5119449520

model-00003-of-00014.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4322f9a3659ac1b3f1aa6445d23e00294b876d76c2dcb940b103a94afb68290
+size 5119449504

model-00004-of-00014.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc9a1c934aa3e438031f7272ab103fc42d8dbbaad5b35a6a9041fe8b2615c03b
+size 5119450272

model-00005-of-00014.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:52aac8a7fb885688771a7c74a9d06e62b57cdbbecb5282347e7d9c9ad0ebf59c
+size 5119451824

model-00006-of-00014.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:05f9030c4d16a4b858e31cd470511784d3917a2a6f023ed8a5362bb239b7997c
+size 5119451944

model-00007-of-00014.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c5fb7baabed09175615fe9d9fd93544bfb8c70b24d81a139719eeaae0b105ab
+size 5119451960

model-00008-of-00014.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:452ccc8d15c66187c90845a504b8eb66105ed185da996d180ff2a93aea19889b
+size 5119451960

model-00009-of-00014.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:91343e6489e08e6b8c1f94ad333dade5c1dd34ff10b9bcd7600aff346337c7e5
+size 5119451872

model-00010-of-00014.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0fd9ba3702aff6e57e11362b8347382279ceef6a9ef0896571771ea5c3d3da08
+size 5119451824

model-00011-of-00014.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b131998e1f04900a4c809675ccbbb33ee2d3fd8237ab364d809331d59c0f09bb
+size 5119451856

model-00012-of-00014.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07749587fc5f27ce84ca6889afd840b68dd5878019d37e22881ec727cdbf59aa
+size 5119451960

model-00013-of-00014.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:933f10a0e0b31fb9f9904f21a2bbd0beaf5ec211ad1ebb7ff91b6086a304d243
+size 5119451960

model-00014-of-00014.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:52bf6dae97176c8476d198fb820912f9f6a6b51b682b10560befd88f2969c384
+size 335563984

modeling_laguna.py

@@ -0,0 +1,785 @@
+# Copyright 2025 Poolside and the HuggingFace Inc. team. All rights reserved.
+#
+# 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.
+
+import copy
+from collections.abc import Callable
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from transformers import initialization as init
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.generation import GenerationMixin
+from transformers.integrations import (
+    use_kernel_forward_from_hub,
+    use_kernel_func_from_hub,
+    use_kernelized_func,
+)
+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 GradientCheckpointingLayer
+from transformers.modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast
+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 auto_docstring, can_return_tuple, is_grouped_mm_available
+from transformers.utils.generic import TransformersKwargs, check_model_inputs, maybe_autocast
+
+try:
+    # transformers >= 5.5 relocated OutputRecorder to a dedicated module.
+    from transformers.utils.output_capturing import OutputRecorder
+except ImportError:
+    from transformers.utils.generic import OutputRecorder  # type: ignore[no-redef]
+from .configuration_laguna import LagunaConfig
+
+
+def _build_rope_config(base_config, rope_params, partial_rotary_factor):
+    """Shallow-copy the config with rope_parameters / partial_rotary_factor overridden."""
+    cfg = copy.copy(base_config)
+    if rope_params is not None:
+        cfg.rope_parameters = dict(rope_params)
+    if partial_rotary_factor is not None:
+        cfg.partial_rotary_factor = float(partial_rotary_factor)
+    return cfg
+
+
+@use_kernel_forward_from_hub("RMSNorm")
+class LagunaRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        LagunaRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        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 + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+class LagunaRotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: LagunaConfig, 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.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
+
+    @staticmethod
+    def compute_default_rope_parameters(
+        config: LagunaConfig | None = None,
+        device: Optional["torch.device"] = None,
+        seq_len: int | None = None,
+    ) -> tuple["torch.Tensor", float]:
+        """
+        Computes the inverse frequencies according to the original RoPE implementation
+        Args:
+            config ([`~transformers.PreTrainedConfig`]):
+                The model configuration.
+            device (`torch.device`):
+                The device to use for initialization of the inverse frequencies.
+            seq_len (`int`, *optional*):
+                The current sequence length. Unused for this type of RoPE.
+        Returns:
+            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
+            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
+        """
+        base = config.rope_parameters["rope_theta"]
+        head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
+        partial = getattr(config, "partial_rotary_factor", 1.0)
+        dim = int(head_dim * partial)
+
+        attention_factor = 1.0  # Unused in this type of RoPE
+
+        # Compute the inverse frequencies
+        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
+
+    @torch.no_grad()
+    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+class LagunaMLP(nn.Module):
+    def __init__(self, config, intermediate_size=None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size if intermediate_size is None else intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+class LagunaTopKRouter(nn.Module):
+    """Laguna MoE router using sigmoid scoring (not softmax)."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.top_k = config.num_experts_per_tok
+        self.num_experts = config.num_experts
+        self.norm_topk_prob = config.norm_topk_prob
+        self.hidden_dim = config.hidden_size
+        self.weight = nn.Parameter(torch.zeros(self.num_experts, self.hidden_dim))
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        e_score_correction_bias: torch.Tensor | None = None,
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        hidden_states = hidden_states.reshape(-1, self.hidden_dim)
+        router_logits = F.linear(hidden_states, self.weight)
+        # Laguna-specific: sigmoid routing in float32 for precision
+        routing_weights = torch.sigmoid(router_logits.float())
+        if e_score_correction_bias is not None:
+            routing_weights = routing_weights + e_score_correction_bias.float()
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        if self.norm_topk_prob:
+            routing_weights = routing_weights / routing_weights.sum(dim=-1, keepdim=True)
+        routing_weights = routing_weights.to(hidden_states.dtype)
+        return router_logits, routing_weights, selected_experts
+
+
+class LagunaSparseMoeBlock(nn.Module):
+    """Laguna MoE block using sigmoid router, per-expert MLPs, and a shared expert."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.num_experts = config.num_experts
+        self.top_k = config.num_experts_per_tok
+        self.routed_scaling_factor = float(getattr(config, "moe_routed_scaling_factor", 1.0))
+        self.apply_router_weight_on_input = bool(getattr(config, "moe_apply_router_weight_on_input", False))
+        self.gate = LagunaTopKRouter(config)
+        self.experts = nn.ModuleList(
+            [LagunaMLP(config, intermediate_size=config.moe_intermediate_size) for _ in range(self.num_experts)]
+        )
+        self.experts.e_score_correction_bias = nn.Parameter(torch.zeros(self.num_experts))
+        self.shared_expert = LagunaMLP(config, intermediate_size=config.shared_expert_intermediate_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+
+        shared_expert_output = self.shared_expert(hidden_states)
+
+        _, routing_weights, selected_experts = self.gate(
+            hidden_states, e_score_correction_bias=self.experts.e_score_correction_bias
+        )
+        routed_output = torch.zeros_like(hidden_states)
+
+        expert_mask = F.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
+
+        for expert_idx in range(self.num_experts):
+            top_k_pos, token_idx = torch.where(expert_mask[expert_idx])
+            if token_idx.shape[0] == 0:
+                continue
+            w = routing_weights[token_idx, top_k_pos, None]
+            if self.apply_router_weight_on_input:
+                current = self.experts[expert_idx](hidden_states[token_idx] * w)
+            else:
+                current = self.experts[expert_idx](hidden_states[token_idx]) * w
+            routed_output.index_add_(0, token_idx, current.to(routed_output.dtype))
+
+        routed_output = routed_output * self.routed_scaling_factor
+        final_hidden_states = routed_output + shared_expert_output
+        return final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+
+
+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)
+
+
+@use_kernel_func_from_hub("rotary_pos_emb")
+def apply_rotary_pos_emb(q, k, cos, sin, 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.
+        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)
+    rot_dim = cos.shape[-1]
+    if rot_dim == q.shape[-1]:
+        q_embed = (q * cos) + (rotate_half(q) * sin)
+        k_embed = (k * cos) + (rotate_half(k) * sin)
+        return q_embed, k_embed
+    q_rot, q_pass = q[..., :rot_dim], q[..., rot_dim:]
+    k_rot, k_pass = k[..., :rot_dim], k[..., rot_dim:]
+    q_rot = (q_rot * cos) + (rotate_half(q_rot) * sin)
+    k_rot = (k_rot * cos) + (rotate_half(k_rot) * sin)
+    return torch.cat([q_rot, q_pass], dim=-1), torch.cat([k_rot, k_pass], dim=-1)
+
+
+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: torch.Tensor | None,
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs: Unpack[TransformersKwargs],
+):
+    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 = 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
+
+
+# Laguna attention is identical to Qwen2MoE attention except:
+# - No QKV bias
+# - Explicit head_dim from config
+# - Output gating: attn_output = attn_output * softplus(g_proj(hidden_states))
+# - No sliding window (full attention only)
+@use_kernelized_func(apply_rotary_pos_emb)
+class LagunaAttention(nn.Module):
+    def __init__(self, config: LagunaConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = config.head_dim
+
+        per_layer_heads = getattr(config, "num_attention_heads_per_layer", None)
+        num_heads = per_layer_heads[layer_idx] if per_layer_heads is not None else config.num_attention_heads
+        self.num_heads = num_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = num_heads // config.num_key_value_heads
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_causal = True
+
+        self.q_proj = nn.Linear(config.hidden_size, num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(config.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(config.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(num_heads * self.head_dim, config.hidden_size, bias=False)
+
+        gating = getattr(config, "gating", True)
+        self.gating = bool(gating)
+        self.gate_per_head = gating == "per-head"
+        if self.gating:
+            g_out = num_heads if self.gate_per_head else num_heads * self.head_dim
+            self.g_proj = nn.Linear(config.hidden_size, g_out, bias=False)
+
+        self.q_norm = LagunaRMSNorm(config.head_dim, eps=config.rms_norm_eps)
+        self.k_norm = LagunaRMSNorm(config.head_dim, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: torch.Tensor | None,
+        past_key_values: Cache | None = None,
+        cache_position: torch.LongTensor | None = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(hidden_shape).transpose(1, 2)
+        key_states = key_states.view(hidden_shape).transpose(1, 2)
+        value_states = value_states.view(hidden_shape).transpose(1, 2)
+
+        # QK normalization (applied per-head before RoPE)
+        query_states = self.q_norm(query_states)
+        key_states = self.k_norm(key_states)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_values is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+
+        if self.gating:
+            gate = F.softplus(self.g_proj(hidden_states).float()).to(attn_output.dtype)
+            if self.gate_per_head:
+                shape = attn_output.shape
+                attn_output = (
+                    attn_output.view(*shape[:-1], self.num_heads, self.head_dim) * gate.unsqueeze(-1)
+                ).view(shape)
+            else:
+                attn_output = attn_output * gate
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, attn_weights
+
+
+class LagunaDecoderLayer(GradientCheckpointingLayer):
+    """Laguna decoder layer with gated attention and sigmoid-routed MoE."""
+
+    def __init__(self, config: LagunaConfig, layer_idx: int):
+        super().__init__()
+        self.layer_idx = layer_idx
+        layer_types = getattr(config, "layer_types", None)
+        self.attention_type = (
+            layer_types[layer_idx] if layer_types is not None else "full_attention"
+        )
+        self.self_attn = LagunaAttention(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 = LagunaSparseMoeBlock(config)
+        else:
+            self.mlp = LagunaMLP(config, intermediate_size=config.intermediate_size)
+        self.input_layernorm = LagunaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = LagunaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.hidden_size = config.hidden_size
+
+    def _pick(self, obj):
+        if isinstance(obj, dict):
+            return obj.get(self.attention_type, obj.get("full_attention"))
+        return obj
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        use_cache: bool | None = False,
+        cache_position: torch.LongTensor | None = None,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        # Self Attention
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=self._pick(attention_mask),
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=self._pick(position_embeddings),
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+@auto_docstring
+class LagunaPreTrainedModel(PreTrainedModel):
+    config: LagunaConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["LagunaDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _can_compile_fullgraph = (
+        is_grouped_mm_available()
+    )  # https://huggingface.co/docs/transformers/experts_interface#torchcompile
+    _supports_attention_backend = True
+    _can_record_outputs = {
+        "router_logits": OutputRecorder(LagunaTopKRouter, index=0),
+        "hidden_states": LagunaDecoderLayer,
+        "attentions": LagunaAttention,
+    }
+
+    @torch.no_grad()
+    def _init_weights(self, module):
+        super()._init_weights(module)
+        std = self.config.initializer_range
+        if isinstance(module, LagunaTopKRouter):
+            init.normal_(module.weight, mean=0.0, std=std)
+
+
+class LagunaModel(LagunaPreTrainedModel):
+    def __init__(self, config: LagunaConfig):
+        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(
+            [LagunaDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = LagunaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = LagunaRotaryEmbedding(config=config)
+
+        self._has_swa = (
+            config.layer_types is not None and "sliding_attention" in config.layer_types
+        )
+        swa_rp = getattr(config, "swa_rope_parameters", None)
+        if self._has_swa and swa_rp is not None:
+            swa_partial = swa_rp.get("partial_rotary_factor", None)
+            swa_cfg = _build_rope_config(config, swa_rp, swa_partial)
+            self.swa_rotary_emb = LagunaRotaryEmbedding(config=swa_cfg)
+        else:
+            self.swa_rotary_emb = None
+
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @check_model_inputs
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        use_cache: bool | None = None,
+        cache_position: torch.LongTensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ):
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache(config=self.config)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        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)
+
+        global_mask = create_causal_mask(
+            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,
+        )
+
+        hidden_states = inputs_embeds
+        global_pe = self.rotary_emb(hidden_states, position_ids)
+
+        if self._has_swa:
+            swa_mask = create_sliding_window_causal_mask(
+                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,
+            )
+            causal_mask = {"full_attention": global_mask, "sliding_attention": swa_mask}
+            swa_pe = self.swa_rotary_emb(hidden_states, position_ids) if self.swa_rotary_emb is not None else global_pe
+            position_embeddings = {"full_attention": global_pe, "sliding_attention": swa_pe}
+        else:
+            causal_mask = global_mask
+            position_embeddings = global_pe
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            hidden_states = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+                **kwargs,
+            )
+
+        hidden_states = self.norm(hidden_states)
+
+        return MoeModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+        )
+
+
+def load_balancing_loss_func(
+    gate_logits: torch.Tensor | tuple[torch.Tensor] | None,
+    num_experts: int | None = None,
+    top_k=2,
+    attention_mask: torch.Tensor | None = None,
+) -> torch.Tensor | int:
+    r"""
+    Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
+
+    See Switch Transformer (https://huggingface.co/papers/2101.03961) for more details. This function implements the loss
+    function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between
+    experts is too unbalanced.
+
+    Args:
+        gate_logits:
+            Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
+            shape [batch_size X sequence_length, num_experts].
+        num_experts:
+            Number of experts
+        top_k:
+            The number of experts to route per-token, can be also interpreted as the `top-k` routing
+            parameter.
+        attention_mask (`torch.Tensor`, *optional*):
+            The attention_mask used in forward function
+            shape [batch_size X sequence_length] if not None.
+
+    Returns:
+        The auxiliary loss.
+    """
+    if gate_logits is None or not isinstance(gate_logits, tuple):
+        return 0
+
+    if isinstance(gate_logits, tuple):
+        compute_device = gate_logits[0].device
+        concatenated_gate_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0)
+
+    routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1)
+
+    _, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
+
+    expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)
+
+    if attention_mask is None:
+        # Compute the percentage of tokens routed to each experts
+        tokens_per_expert = torch.mean(expert_mask.float(), dim=0)
+
+        # Compute the average probability of routing to these experts
+        router_prob_per_expert = torch.mean(routing_weights, dim=0)
+    else:
+        batch_size, sequence_length = attention_mask.shape
+        num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length)
+
+        # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
+        expert_attention_mask = (
+            attention_mask[None, :, :, None, None]
+            .expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts))
+            .reshape(-1, top_k, num_experts)
+            .to(compute_device)
+        )
+
+        # Compute the percentage of tokens routed to each experts
+        tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(
+            expert_attention_mask, dim=0
+        )
+
+        # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert
+        router_per_expert_attention_mask = (
+            attention_mask[None, :, :, None]
+            .expand((num_hidden_layers, batch_size, sequence_length, num_experts))
+            .reshape(-1, num_experts)
+            .to(compute_device)
+        )
+
+        # Compute the average probability of routing to these experts
+        router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum(
+            router_per_expert_attention_mask, dim=0
+        )
+
+    overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0))
+    return overall_loss * num_experts
+
+
+@auto_docstring
+class LagunaForCausalLM(LagunaPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = LagunaModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.router_aux_loss_coef = config.router_aux_loss_coef
+        self.num_experts = config.num_experts
+        self.num_experts_per_tok = config.num_experts_per_tok
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        labels: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        output_router_logits: bool | None = None,
+        cache_position: torch.LongTensor | None = None,
+        logits_to_keep: int | torch.Tensor = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> MoeCausalLMOutputWithPast:
+        r"""
+        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]`.
+        """
+        # TODO (Joe) add example here after we got rid of the stale mistral example
+
+        output_router_logits = (
+            output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        )
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs: MoeModelOutputWithPast = 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_router_logits=output_router_logits,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        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, labels, self.vocab_size, **kwargs)
+
+        aux_loss = None
+        if output_router_logits:
+            aux_loss = load_balancing_loss_func(
+                outputs.router_logits,
+                self.num_experts,
+                self.num_experts_per_tok,
+                attention_mask,
+            )
+            if labels is not None:
+                loss += self.router_aux_loss_coef * aux_loss.to(loss.device)  # make sure to reside in the same device
+
+        return MoeCausalLMOutputWithPast(
+            loss=loss,
+            aux_loss=aux_loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            router_logits=outputs.router_logits,
+        )
+
+
+__all__ = ["LagunaForCausalLM", "LagunaModel", "LagunaPreTrainedModel"]

special_tokens_map.json

@@ -0,0 +1,9 @@
+{
+  "bos_token": "〈|EOS|〉",
+  "cls_token": "〈|CLS|〉",
+  "eos_token": "〈|EOS|〉",
+  "mask_token": "〈|MASK|〉",
+  "pad_token": "〈|PAD|〉",
+  "sep_token": "〈|SEP|〉",
+  "unk_token": "〈|UNK|〉"
+}

tokenizer_config.json

@@ -0,0 +1,575 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "〈|UNK|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "〈|CODE_START|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "〈|EOS|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "〈|CODE_END|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "4": {
+      "content": "〈|META_START|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "5": {
+      "content": "〈|META_END|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "6": {
+      "content": "〈|FIM_MIDDLE|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "7": {
+      "content": "〈|FIM_SUFFIX|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "8": {
+      "content": "〈|SEP|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "9": {
+      "content": "〈|PAD|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "10": {
+      "content": "〈|CLS|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "11": {
+      "content": "〈|FIM_START|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "12": {
+      "content": "〈|MASK|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "13": {
+      "content": "|◊|",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "14": {
+      "content": "〈|",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "15": {
+      "content": "|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "16": {
+      "content": "〈|/",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "17": {
+      "content": "/|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "18": {
+      "content": "〈|THINK_START|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "19": {
+      "content": "〈|THINK_END|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "20": {
+      "content": "〈|SPECIAL_1|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "21": {
+      "content": "〈|SPECIAL_2|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "22": {
+      "content": "〈|SPECIAL_3|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "23": {
+      "content": "〈|SPECIAL_4|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "24": {
+      "content": "〈|SPECIAL_5|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "25": {
+      "content": "〈|SPECIAL_6|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "26": {
+      "content": "〈|SPECIAL_7|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "27": {
+      "content": "〈|SPECIAL_8|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "28": {
+      "content": "〈|SPECIAL_9|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "29": {
+      "content": "〈|SPECIAL_10|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "30": {
+      "content": "〈|SPECIAL_11|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "31": {
+      "content": "〈|SPECIAL_12|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32": {
+      "content": "〈|SPECIAL_13|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "33": {
+      "content": "〈|SPECIAL_14|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "34": {
+      "content": "〈|SPECIAL_15|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "35": {
+      "content": "〈|SPECIAL_16|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "36": {
+      "content": "〈|SPECIAL_17|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "37": {
+      "content": "〈|SPECIAL_18|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "38": {
+      "content": "〈|SPECIAL_19|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "39": {
+      "content": "〈|SPECIAL_20|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "40": {
+      "content": "〈|SPECIAL_21|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "41": {
+      "content": "〈|SPECIAL_22|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "42": {
+      "content": "〈|SPECIAL_23|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "43": {
+      "content": "〈|SPECIAL_24|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "44": {
+      "content": "〈|SPECIAL_25|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "45": {
+      "content": "〈|SPECIAL_26|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "46": {
+      "content": "〈|SPECIAL_27|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "47": {
+      "content": "〈|SPECIAL_28|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "48": {
+      "content": "〈|SPECIAL_29|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49": {
+      "content": "〈|SPECIAL_30|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50": {
+      "content": "〈|SPECIAL_31|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "51": {
+      "content": "〈|SPECIAL_32|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "52": {
+      "content": "〈|SPECIAL_33|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "53": {
+      "content": "〈|SPECIAL_34|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "54": {
+      "content": "〈|SPECIAL_35|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "55": {
+      "content": "〈|SPECIAL_36|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "56": {
+      "content": "〈|SPECIAL_37|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "57": {
+      "content": "〈|SPECIAL_38|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "58": {
+      "content": "〈|SPECIAL_39|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "59": {
+      "content": "〈|SPECIAL_40|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "60": {
+      "content": "〈|SPECIAL_41|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "61": {
+      "content": "〈|SPECIAL_42|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "62": {
+      "content": "〈|SPECIAL_43|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "63": {
+      "content": "〈|SPECIAL_44|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "64": {
+      "content": "〈|SPECIAL_45|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65": {
+      "content": "〈|SPECIAL_46|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "66": {
+      "content": "〈|SPECIAL_47|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "67": {
+      "content": "〈|SPECIAL_48|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "68": {
+      "content": "〈|SPECIAL_49|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "69": {
+      "content": "〈|SPECIAL_50|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "〈|EOS|〉",
+  "clean_up_tokenization_spaces": false,
+  "cls_token": "〈|CLS|〉",
+  "eos_token": "〈|EOS|〉",
+  "extra_special_tokens": {},
+  "mask_token": "〈|MASK|〉",
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "〈|PAD|〉",
+  "sep_token": "〈|SEP|〉",
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "unk_token": "〈|UNK|〉"
+}