Sync bundled HF code with upstream Laguna PR (v5 schema)

modeling_laguna.py

@@ -1,5 +1,4 @@
-# ruff: noqa
-# Copyright 2025 Poolside and the HuggingFace Inc. team. All rights reserved.
+# Copyright 2026 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.
@@ -13,37 +12,34 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Optional
 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.utils import auto_docstring, can_return_tuple, is_grouped_mm_available
-from transformers.generation import GenerationMixin
 from transformers.activations import ACT2FN
 from transformers.cache_utils import Cache, DynamicCache
-from transformers.integrations import (
-    use_kernelized_func,
-    use_kernel_func_from_hub,
-    use_kernel_forward_from_hub,
-)
-from transformers.masking_utils import create_causal_mask
-from transformers.utils.generic import OutputRecorder, TransformersKwargs, maybe_autocast, check_model_inputs
-from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.generation import GenerationMixin
+from transformers.integrations import use_experts_implementation, use_kernel_forward_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 MoeModelOutputWithPast, MoeCausalLMOutputWithPast
-from transformers.processing_utils import Unpack
+from transformers.modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast
 from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
-from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
-
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.utils import auto_docstring, can_return_tuple
+from transformers.utils.generic import TransformersKwargs, maybe_autocast
+from transformers.utils.output_capturing import OutputRecorder, capture_outputs
 from .configuration_laguna import LagunaConfig
 
 
 @use_kernel_forward_from_hub("RMSNorm")
 class LagunaRMSNorm(nn.Module):
-    def __init__(self, hidden_size, eps=1e-6):
+    def __init__(self, hidden_size, eps: float = 1e-6) -> None:
         """
         LagunaRMSNorm is equivalent to T5LayerNorm
         """
@@ -51,7 +47,7 @@ class LagunaRMSNorm(nn.Module):
         self.weight = nn.Parameter(torch.ones(hidden_size))
         self.variance_epsilon = eps
 
-    def forward(self, hidden_states):
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         input_dtype = hidden_states.dtype
         hidden_states = hidden_states.to(torch.float32)
         variance = hidden_states.pow(2).mean(-1, keepdim=True)
@@ -65,27 +61,35 @@ class LagunaRMSNorm(nn.Module):
 class LagunaRotaryEmbedding(nn.Module):
     inv_freq: torch.Tensor  # fix linting for `register_buffer`
 
-    def __init__(self, config: LagunaConfig, device=None):
+    def __init__(self, config: LagunaConfig, device=None, layer_type=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.layer_types = list(set(config.layer_types))
+        self.rope_type = {}
+        for layer_type in self.layer_types:
+            rope_params = self.config.rope_parameters[layer_type]
+            if rope_params is None:
+                continue
 
-        self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
+            self.rope_type[layer_type] = rope_params["rope_type"]
+            rope_init_fn: Callable = self.compute_default_rope_parameters
+            if self.rope_type[layer_type] != "default":
+                rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type[layer_type]]
+            curr_inv_freq, curr_attention_scaling = rope_init_fn(self.config, device, layer_type=layer_type)
+            self.register_buffer(f"{layer_type}_inv_freq", curr_inv_freq, persistent=False)
+            self.register_buffer(f"{layer_type}_original_inv_freq", curr_inv_freq.clone(), persistent=False)
+            setattr(self, f"{layer_type}_attention_scaling", curr_attention_scaling)
 
     @staticmethod
     def compute_default_rope_parameters(
         config: LagunaConfig | None = None,
         device: Optional["torch.device"] = None,
         seq_len: int | None = None,
+        layer_type: str | None = None,
     ) -> tuple["torch.Tensor", float]:
         """
         Computes the inverse frequencies according to the original RoPE implementation
@@ -96,14 +100,18 @@ class LagunaRotaryEmbedding(nn.Module):
                 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
-        -------
+            layer_type (`str`, *optional*):
+                The current layer type if the model has different RoPE parameters per type.
+                Should not be used unless `config.layer_types is not None`
+        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"]
-        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
+        base = config.rope_parameters[layer_type]["rope_theta"]
+        # key difference to gemma3: partial rope
+        partial_rotary_factor = config.rope_parameters[layer_type].get("partial_rotary_factor", 1.0)
+        head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
+        dim = int(head_dim * partial_rotary_factor)
 
         attention_factor = 1.0  # Unused in this type of RoPE
 
@@ -115,16 +123,19 @@ class LagunaRotaryEmbedding(nn.Module):
 
     @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)
+    def forward(self, x, position_ids, layer_type=None):
+        inv_freq = getattr(self, f"{layer_type}_inv_freq")
+        attention_scaling = getattr(self, f"{layer_type}_attention_scaling")
+
+        inv_freq_expanded = 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
+            cos = emb.cos() * attention_scaling
+            sin = emb.sin() * attention_scaling
 
         return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
 
@@ -146,71 +157,97 @@ class LagunaMLP(nn.Module):
 
 
 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))
+        self.e_score_correction_bias = nn.Parameter(torch.zeros(config.num_experts), requires_grad=False)
+        self.router_logit_softcapping = config.moe_router_logit_softcapping
 
-    def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+    ) -> 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())
-        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)
+        router_logits = F.linear(hidden_states, self.weight).float()
+        # Optional logits softcapping
+        if self.router_logit_softcapping > 0.0:
+            router_logits = torch.tanh(router_logits / self.router_logit_softcapping) * self.router_logit_softcapping
+        # Sigmoid instead of softmax normalization
+        routing_scores = torch.sigmoid(router_logits)
+
+        scores_for_selection = routing_scores + self.e_score_correction_bias.to(routing_scores.dtype)
+        _, selected_experts = torch.topk(scores_for_selection, self.top_k, dim=-1)
+        routing_weights = routing_scores.gather(-1, selected_experts)
+        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."""
+@use_experts_implementation
+class LagunaExperts(nn.Module):
+    """Collection of expert weights stored as 3D tensors."""
 
     def __init__(self, config):
         super().__init__()
         self.num_experts = config.num_experts
-        self.top_k = config.num_experts_per_tok
+        self.hidden_dim = config.hidden_size
+        self.intermediate_dim = config.moe_intermediate_size
+        self.gate_up_proj = nn.Parameter(torch.empty(self.num_experts, 2 * self.intermediate_dim, self.hidden_dim))
+        self.down_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_dim, self.intermediate_dim))
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        top_k_index: torch.Tensor,
+        top_k_weights: torch.Tensor,
+    ) -> torch.Tensor:
+        final_hidden_states = torch.zeros_like(hidden_states)
+        with torch.no_grad():
+            expert_mask = torch.nn.functional.one_hot(top_k_index, num_classes=self.num_experts)
+            expert_mask = expert_mask.permute(2, 1, 0)
+            expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()
+
+        for expert_idx in expert_hit:
+            expert_idx = expert_idx[0]
+            if expert_idx == self.num_experts:
+                continue
+            top_k_pos, token_idx = torch.where(expert_mask[expert_idx])
+            current_state = hidden_states[token_idx]
+            gate, up = nn.functional.linear(current_state, self.gate_up_proj[expert_idx]).chunk(2, dim=-1)
+            current_hidden_states = self.act_fn(gate) * up
+            current_hidden_states = nn.functional.linear(current_hidden_states, self.down_proj[expert_idx])
+            current_hidden_states = current_hidden_states * top_k_weights[token_idx, top_k_pos, None]
+            final_hidden_states.index_add_(0, token_idx, current_hidden_states.to(final_hidden_states.dtype))
+
+        return final_hidden_states
+
+
+class LagunaSparseMoeBlock(nn.Module):
+    def __init__(self, config: LagunaConfig):
+        super().__init__()
+        self.experts = LagunaExperts(config)
         self.gate = LagunaTopKRouter(config)
-        self.experts = nn.ModuleList(
-            [LagunaMLP(config, intermediate_size=config.moe_intermediate_size) for _ in range(self.num_experts)]
-        )
-        self.shared_expert = LagunaMLP(config, intermediate_size=config.shared_expert_intermediate_size)
-        self.shared_expert_gate = (
-            nn.Linear(config.hidden_size, 1, bias=False) if getattr(config, "moe_shared_gate", False) else None
-        )
+        self.shared_experts = LagunaMLP(config, intermediate_size=config.shared_expert_intermediate_size)
+        self.routed_scaling_factor = config.moe_routed_scaling_factor
 
     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_output = self.shared_experts(hidden_states)
 
-        shared_expert_output = self.shared_expert(hidden_states)
-        if self.shared_expert_gate is not None:
-            shared_expert_output = shared_expert_output * torch.sigmoid(self.shared_expert_gate(hidden_states))
-
-        # Routed experts
         _, routing_weights, selected_experts = self.gate(hidden_states)
-        final_hidden_states = torch.zeros_like(hidden_states)
-
-        expert_mask = F.one_hot(selected_experts, num_classes=self.num_experts)
-        expert_mask = expert_mask.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
-            current_state = hidden_states[token_idx]
-            current_hidden_states = self.experts[expert_idx](current_state)
-            current_hidden_states = current_hidden_states * routing_weights[token_idx, top_k_pos, None]
-            final_hidden_states.index_add_(0, token_idx, current_hidden_states.to(final_hidden_states.dtype))
+        hidden_states = self.experts(hidden_states, selected_experts, routing_weights)
+        # Additional scaling
+        hidden_states = hidden_states * self.routed_scaling_factor
+        hidden_states = hidden_states + shared_output
 
-        final_hidden_states = final_hidden_states + shared_expert_output
-        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
-        return final_hidden_states
+        hidden_states = hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+        return hidden_states
 
 
 def rotate_half(x):
@@ -220,10 +257,12 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-@use_kernel_func_from_hub("rotary_pos_emb")
+# Adapted from transformers.models.glm.modular_glm.apply_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.
 
+    Removes the interleaving of cos and sin from GLM
+
     Args:
         q (`torch.Tensor`): The query tensor.
         k (`torch.Tensor`): The key tensor.
@@ -236,15 +275,24 @@ def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
             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
-    -------
+    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)
+
+    # Keep half or full tensor for later concatenation
+    rotary_dim = cos.shape[-1]
+    q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
+    k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]
+
+    # Apply rotary embeddings on the first half or full tensor
+    q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin)
+    k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin)
+
+    # Concatenate back to full shape
+    q_embed = torch.cat([q_embed, q_pass], dim=-1)
+    k_embed = torch.cat([k_embed, k_pass], dim=-1)
     return q_embed, k_embed
 
 
@@ -275,8 +323,7 @@ def eager_attention_forward(
 
     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 = attn_weights + attention_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)
@@ -286,33 +333,39 @@ def eager_attention_forward(
     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):
+    """Afmoe-style SWA/GQA attention with Laguna-specific gating and per-layer head count."""
+
+    def __init__(self, config: LagunaConfig, layer_idx: int, num_heads: int):
         super().__init__()
+        # Number of heads is controlled via `config.num_attention_heads_per_layer` which is passed from the parent for the specific layer
+        self.num_heads = num_heads
         self.config = config
         self.layer_idx = layer_idx
-        self.head_dim = config.head_dim
-        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.num_key_value_groups = self.num_heads // config.num_key_value_heads
         self.scaling = self.head_dim**-0.5
         self.attention_dropout = config.attention_dropout
         self.is_causal = True
 
-        # Laguna: no QKV bias, explicit head_dim
-        self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * config.head_dim, bias=False)
-        self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * config.head_dim, bias=False)
-        self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * config.head_dim, bias=False)
-        self.o_proj = nn.Linear(config.num_attention_heads * config.head_dim, config.hidden_size, bias=False)
-        # Laguna-specific: gating projection
-        self.g_proj = nn.Linear(config.hidden_size, config.num_attention_heads * config.head_dim, bias=False)
-        # QK normalization (RMSNorm applied per-head after reshape, before RoPE)
-        self.q_norm = LagunaRMSNorm(config.head_dim, eps=config.rms_norm_eps)
-        self.k_norm = LagunaRMSNorm(config.head_dim, eps=config.rms_norm_eps)
+        # Per-layer head count: rebuild q_proj and o_proj using self.num_heads (parent uses config.num_attention_heads).
+        self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
+        self.k_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.v_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, config.hidden_size, bias=config.attention_bias)
+        # Parent LlamaAttention already sets: layer_idx, num_heads, num_key_value_heads, num_key_value_groups, head_dim
+        # We only add Laguna-specific attributes
+        self.is_local_attention = config.layer_types[layer_idx] == "sliding_attention"
+        self.sliding_window = config.sliding_window if self.is_local_attention else None
+
+        self.q_norm = LagunaRMSNorm(self.head_dim, eps=config.rms_norm_eps)
+        self.k_norm = LagunaRMSNorm(self.head_dim, eps=config.rms_norm_eps)
+        self.g_proj = nn.Linear(config.hidden_size, self.num_heads, bias=False)
 
     def forward(
         self,
@@ -320,36 +373,28 @@ class LagunaAttention(nn.Module):
         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)
+        query_states = self.q_proj(hidden_states).view(hidden_shape)
+        key_states = self.k_proj(hidden_states).view(hidden_shape)
+        value_states = self.v_proj(hidden_states).view(hidden_shape)
 
-        # QK normalization (applied per-head before RoPE)
-        query_states = self.q_norm(query_states)
-        key_states = self.k_norm(key_states)
+        query_states = self.q_norm(query_states).transpose(1, 2)
+        key_states = self.k_norm(key_states).transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
 
         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]
+            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)
 
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
         attn_output, attn_weights = attention_interface(
             self,
             query_states,
@@ -358,37 +403,30 @@ class LagunaAttention(nn.Module):
             attention_mask,
             dropout=0.0 if not self.training else self.attention_dropout,
             scaling=self.scaling,
+            sliding_window=self.sliding_window,
             **kwargs,
         )
 
         attn_output = attn_output.reshape(*input_shape, -1).contiguous()
 
-        # Laguna-specific: apply gating BEFORE o_proj
-        # gate values are computed from original hidden_states, applied in attention dimension
         gate = F.softplus(self.g_proj(hidden_states).float()).to(attn_output.dtype)
-        attn_output = attn_output * gate
+        attn_output = (attn_output.view(*input_shape, -1, self.head_dim) * gate.unsqueeze(-1)).view(*input_shape, -1)
 
         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.self_attn = LagunaAttention(config, layer_idx)
-        # Use MoE or dense MLP based on layer configuration
-        if (layer_idx not in config.mlp_only_layers) and (
-            config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
-        ):
+        self.hidden_size = config.hidden_size
+        self.self_attn = LagunaAttention(config, layer_idx, config.num_attention_heads_per_layer[layer_idx])
+        if config.mlp_layer_types[layer_idx] == "sparse":
             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 forward(
         self,
@@ -397,7 +435,6 @@ class LagunaDecoderLayer(GradientCheckpointingLayer):
         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:
@@ -410,7 +447,6 @@ class LagunaDecoderLayer(GradientCheckpointingLayer):
             position_ids=position_ids,
             past_key_values=past_key_values,
             use_cache=use_cache,
-            cache_position=cache_position,
             position_embeddings=position_embeddings,
             **kwargs,
         )
@@ -434,9 +470,8 @@ class LagunaPreTrainedModel(PreTrainedModel):
     _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
+
+    _can_compile_fullgraph = True
     _supports_attention_backend = True
     _can_record_outputs = {
         "router_logits": OutputRecorder(LagunaTopKRouter, index=0),
@@ -448,10 +483,24 @@ class LagunaPreTrainedModel(PreTrainedModel):
     def _init_weights(self, module):
         super()._init_weights(module)
         std = self.config.initializer_range
-        if isinstance(module, LagunaTopKRouter):
+        if isinstance(module, LagunaExperts):
+            init.normal_(module.gate_up_proj, mean=0.0, std=std)
+            init.normal_(module.down_proj, mean=0.0, std=std)
+        elif isinstance(module, LagunaTopKRouter):
             init.normal_(module.weight, mean=0.0, std=std)
+        if isinstance(module, LagunaTopKRouter):
+            torch.nn.init.zeros_(module.e_score_correction_bias)
+        elif isinstance(module, LagunaRotaryEmbedding):
+            for layer_type in module.layer_types:
+                rope_init_fn = module.compute_default_rope_parameters
+                if module.rope_type[layer_type] != "default":
+                    rope_init_fn = ROPE_INIT_FUNCTIONS[module.rope_type[layer_type]]
+                curr_inv_freq, _ = rope_init_fn(module.config, layer_type=layer_type)
+                init.copy_(getattr(module, f"{layer_type}_inv_freq"), curr_inv_freq)
+                init.copy_(getattr(module, f"{layer_type}_original_inv_freq"), curr_inv_freq)
 
 
+@auto_docstring
 class LagunaModel(LagunaPreTrainedModel):
     def __init__(self, config: LagunaConfig):
         super().__init__(config)
@@ -469,7 +518,8 @@ class LagunaModel(LagunaPreTrainedModel):
         # Initialize weights and apply final processing
         self.post_init()
 
-    @check_model_inputs
+    @capture_outputs
+    @auto_docstring
     def forward(
         self,
         input_ids: torch.LongTensor | None = None,
@@ -478,49 +528,50 @@ class LagunaModel(LagunaPreTrainedModel):
         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],
-    ):
+    ) -> MoeModelOutputWithPast:
         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 use_cache and past_key_values is None:
+            past_key_values = DynamicCache(config=self.config)
 
         if position_ids is None:
-            position_ids = cache_position.unsqueeze(0)
-
-        # Laguna uses full attention only (no sliding window)
-        causal_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,
-        )
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens
+            position_ids = position_ids.unsqueeze(0)
+
+        if not isinstance(causal_mask_mapping := attention_mask, dict):
+            mask_kwargs = {
+                "config": self.config,
+                "inputs_embeds": inputs_embeds,
+                "attention_mask": attention_mask,
+                "past_key_values": past_key_values,
+                "position_ids": position_ids,
+            }
+            mask_creation_functions = {
+                "full_attention": lambda: create_causal_mask(**mask_kwargs),
+                "sliding_attention": lambda: create_sliding_window_causal_mask(**mask_kwargs),
+            }
+            causal_mask_mapping = {}
+            for layer_type in set(self.config.layer_types):
+                causal_mask_mapping[layer_type] = mask_creation_functions[layer_type]()
 
         hidden_states = inputs_embeds
-        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+        position_embeddings = {}
+        for layer_type in set(self.config.layer_types):
+            position_embeddings[layer_type] = self.rotary_emb(hidden_states, position_ids, layer_type)
 
-        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+        for i, decoder_layer in enumerate(self.layers[: self.config.num_hidden_layers]):
             hidden_states = decoder_layer(
                 hidden_states,
-                attention_mask=causal_mask,
+                attention_mask=causal_mask_mapping[self.config.layer_types[i]],
+                position_embeddings=position_embeddings[self.config.layer_types[i]],
                 position_ids=position_ids,
                 past_key_values=past_key_values,
-                use_cache=use_cache,
-                cache_position=cache_position,
-                position_embeddings=position_embeddings,
                 **kwargs,
             )
 
@@ -528,7 +579,7 @@ class LagunaModel(LagunaPreTrainedModel):
 
         return MoeModelOutputWithPast(
             last_hidden_state=hidden_states,
-            past_key_values=past_key_values,
+            past_key_values=past_key_values if use_cache else None,
         )
 
 
@@ -558,8 +609,7 @@ def load_balancing_loss_func(
             The attention_mask used in forward function
             shape [batch_size X sequence_length] if not None.
 
-    Returns
-    -------
+    Returns:
         The auxiliary loss.
     """
     if gate_logits is None or not isinstance(gate_logits, tuple):
@@ -618,7 +668,7 @@ def load_balancing_loss_func(
 @auto_docstring
 class LagunaForCausalLM(LagunaPreTrainedModel, GenerationMixin):
     _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
-    _tp_plan = {"lm_head": "colwise_rep"}
+    _tp_plan = {"lm_head": "colwise_gather_output"}
     _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
 
     def __init__(self, config):
@@ -645,17 +695,15 @@ class LagunaForCausalLM(LagunaPreTrainedModel, GenerationMixin):
         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 (`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
@@ -670,7 +718,6 @@ class LagunaForCausalLM(LagunaPreTrainedModel, GenerationMixin):
             inputs_embeds=inputs_embeds,
             use_cache=use_cache,
             output_router_logits=output_router_logits,
-            cache_position=cache_position,
             **kwargs,
         )
 
@@ -691,8 +738,8 @@ class LagunaForCausalLM(LagunaPreTrainedModel, GenerationMixin):
                 self.num_experts_per_tok,
                 attention_mask,
             )
-            if labels is not None and isinstance(aux_loss, torch.Tensor):
-                loss += self.router_aux_loss_coef * aux_loss.to(loss.device)
+            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,