transformers 5.0.0rc2__py3-none-any.whl → 5.1.0__py3-none-any.whl

transformers/models/qwen3_omni_moe/modeling_qwen3_omni_moe.py

@@ -4,7 +4,6 @@
 #             the file from the modular. If any change should be done, please apply the change to the
 #                          modular_qwen3_omni_moe.py file directly. One of our CI enforces this.
 #                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
-# coding=utf-8
 # Copyright 2025 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
 #
 #
@@ -23,7 +22,7 @@
 import math
 from collections.abc import Callable
 from dataclasses import dataclass
-from typing import Optional, Union
+from typing import Optional
 
 import numpy as np
 import torch
@@ -45,8 +44,8 @@ from ...masking_utils import create_causal_mask, create_sliding_window_causal_ma
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import GradientCheckpointingLayer
 from ...modeling_outputs import (
-    BaseModelOutput,
     BaseModelOutputWithPast,
+    BaseModelOutputWithPooling,
     CausalLMOutputWithPast,
     MoeCausalLMOutputWithPast,
     MoeModelOutputWithPast,
@@ -54,8 +53,14 @@ from ...modeling_outputs import (
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import auto_docstring, can_return_tuple, is_grouped_mm_available
-from ...utils.generic import OutputRecorder, TransformersKwargs, check_model_inputs, maybe_autocast
+from ...utils import auto_docstring, can_return_tuple, is_grouped_mm_available, torch_compilable_check
+from ...utils.generic import (
+    OutputRecorder,
+    TransformersKwargs,
+    check_model_inputs,
+    is_flash_attention_requested,
+    maybe_autocast,
+)
 from .configuration_qwen3_omni_moe import (
     Qwen3OmniMoeAudioEncoderConfig,
     Qwen3OmniMoeCode2WavConfig,
@@ -69,6 +74,17 @@ from .configuration_qwen3_omni_moe import (
 )
 
 
+@dataclass
+@auto_docstring
+class BaseModelOutputWithDeepstackFeatures(BaseModelOutputWithPooling):
+    r"""
+    deepstack_features (`List[torch.FloatTensor]`, *optional*):
+        List of hidden-states (feature maps) from deepstack layers.
+    """
+
+    deepstack_features: list[torch.FloatTensor] | None = None
+
+
 class SinusoidsPositionEmbedding(nn.Module):
     def __init__(self, length, channels, max_timescale=10000):
         super().__init__()
@@ -254,13 +270,13 @@ class Qwen3OmniMoePreTrainedModelForConditionalGeneration(Qwen3OmniMoePreTrained
 
     def get_rope_index(
         self,
-        input_ids: Optional[torch.LongTensor] = None,
-        image_grid_thw: Optional[torch.LongTensor] = None,
-        video_grid_thw: Optional[torch.LongTensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
+        input_ids: torch.LongTensor | None = None,
+        image_grid_thw: torch.LongTensor | None = None,
+        video_grid_thw: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
         use_audio_in_video: bool = False,
-        audio_seqlens: Optional[torch.LongTensor] = None,
-        second_per_grids: Optional[torch.Tensor] = None,
+        audio_seqlens: torch.LongTensor | None = None,
+        second_per_grids: torch.Tensor | None = None,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         """
         Calculate the 3D rope index based on image and video's temporal, height and width in LLM.
@@ -512,7 +528,7 @@ def eager_attention_forward(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attention_mask: Optional[torch.Tensor],
+    attention_mask: torch.Tensor | None,
     scaling: float,
     dropout: float = 0.0,
     **kwargs,
@@ -562,10 +578,10 @@ class Qwen3OmniMoeAudioAttention(nn.Module):
     def forward(
         self,
         hidden_states: torch.Tensor,
-        cu_seqlens: Optional[torch.Tensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
+        cu_seqlens: torch.Tensor | None = None,
+        attention_mask: torch.Tensor | None = None,
         **kwargs,
-    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
+    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
         """Input shape: Batch x Time x Channel"""
 
         seq_length, _ = hidden_states.size()
@@ -579,9 +595,9 @@ class Qwen3OmniMoeAudioAttention(nn.Module):
         value_states = value_states.transpose(0, 1).unsqueeze(0)
         max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         attn_output, _ = attention_interface(
             self,
@@ -622,7 +638,7 @@ class Qwen3OmniMoeAudioEncoderLayer(GradientCheckpointingLayer):
         self,
         hidden_states: torch.Tensor,
         cu_seqlens: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
+        attention_mask: torch.Tensor | None = None,
         **kwargs,
     ) -> torch.Tensor:
         """
@@ -671,6 +687,10 @@ class Qwen3OmniMoeAudioEncoder(Qwen3OmniMoePreTrainedModel):
     input_modalities = "audio"
     _no_split_modules = ["Qwen3OmniMoeAudioEncoderLayer"]
     _supports_sdpa = True
+    _can_record_outputs = {
+        "hidden_states": Qwen3OmniMoeAudioEncoderLayer,
+        "attentions": Qwen3OmniMoeAudioAttention,
+    }
 
     def __init__(self, config: Qwen3OmniMoeAudioEncoderConfig):
         super().__init__(config)
@@ -717,7 +737,7 @@ class Qwen3OmniMoeAudioEncoder(Qwen3OmniMoePreTrainedModel):
         # NOTE: the created attention masl only approximates the ragged FA2 attention by
         # allowing bidirectional attention within `cu_seqlens` blocks, and not attending between
         # blocks. Though it will not be a 100% match for FA2's `varlen` path
-        if self.config._attn_implementation == "flash_attention_2":
+        if is_flash_attention_requested(self.config):
             return None
 
         seq_length = inputs_tensor.shape[0]
@@ -731,6 +751,7 @@ class Qwen3OmniMoeAudioEncoder(Qwen3OmniMoePreTrainedModel):
             attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = 0
         return attention_mask
 
+    @check_model_inputs(tie_last_hidden_states=False)
     @auto_docstring
     def forward(
         self,
@@ -748,11 +769,7 @@ class Qwen3OmniMoeAudioEncoder(Qwen3OmniMoePreTrainedModel):
         aftercnn_lens = _get_feat_extract_output_lengths(feature_lens)
         chunk_num = torch.ceil(feature_lens / (self.n_window * 2)).long()
 
-        chunk_lengths = torch.tensor(
-            [self.n_window * 2] * chunk_num.sum(),
-            dtype=torch.long,
-            device=feature_lens.device,
-        )
+        chunk_lengths = torch.full((chunk_num.sum(),), self.n_window * 2, dtype=torch.long, device=feature_lens.device)
         tail_chunk_index = F.pad(chunk_num, (1, 0), value=-1).cumsum(0)[1:]
         chunk_lengths[tail_chunk_index] = feature_lens % (self.n_window * 2)
         chunk_lengths[chunk_lengths == 0] = self.n_window * 2
@@ -804,7 +821,7 @@ class Qwen3OmniMoeAudioEncoder(Qwen3OmniMoePreTrainedModel):
         hidden_states = self.proj1(hidden_states)
         hidden_states = self.act(hidden_states)
         hidden_states = self.proj2(hidden_states)
-        return BaseModelOutput(last_hidden_state=hidden_states)
+        return BaseModelOutputWithPooling(last_hidden_state=hidden_states)
 
     def padded_and_mask_function(self, tensor_list, tensor_len, padding_value=0, padding_side="right"):
         """
@@ -893,8 +910,8 @@ class Qwen3OmniMoeVisionAttention(nn.Module):
         self,
         hidden_states: torch.Tensor,
         cu_seqlens: torch.Tensor,
-        rotary_pos_emb: Optional[torch.Tensor] = None,
-        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        rotary_pos_emb: torch.Tensor | None = None,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
         **kwargs,
     ) -> torch.Tensor:
         seq_length = hidden_states.shape[0]
@@ -908,11 +925,11 @@ class Qwen3OmniMoeVisionAttention(nn.Module):
         key_states = key_states.transpose(0, 1).unsqueeze(0)
         value_states = value_states.transpose(0, 1).unsqueeze(0)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
-        if "flash" in self.config._attn_implementation:
+        if is_flash_attention_requested(self.config):
             # Flash Attention: Use cu_seqlens for variable length attention
             max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
             attn_output, _ = attention_interface(
@@ -997,6 +1014,25 @@ class Qwen3OmniMoeVisionRotaryEmbedding(nn.Module):
         return freqs
 
 
+class Qwen3OmniMoeTextTopKRouter(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.top_k = config.num_experts_per_tok
+        self.num_experts = config.num_experts
+        self.hidden_dim = config.hidden_size
+        self.weight = nn.Parameter(torch.zeros(self.num_experts, self.hidden_dim))
+
+    def forward(self, hidden_states):
+        hidden_states = hidden_states.reshape(-1, self.hidden_dim)
+        router_logits = F.linear(hidden_states, self.weight)  # (seq_len, num_experts)
+        router_logits = torch.nn.functional.softmax(router_logits, dtype=torch.float, dim=-1)
+        router_top_value, router_indices = torch.topk(router_logits, self.top_k, dim=-1)  # (seq_len, top_k)
+        router_top_value /= router_top_value.sum(dim=-1, keepdim=True)
+        router_top_value = router_top_value.to(router_logits.dtype)
+        router_scores = router_top_value
+        return router_logits, router_scores, router_indices
+
+
 class Qwen3OmniMoeVisionMLP(nn.Module):
     def __init__(self, config):
         super().__init__()
@@ -1010,26 +1046,6 @@ class Qwen3OmniMoeVisionMLP(nn.Module):
         return self.linear_fc2(self.act_fn(self.linear_fc1(hidden_state)))
 
 
-class Qwen3OmniMoeVisionPatchEmbed(nn.Module):
-    def __init__(self, config) -> None:
-        super().__init__()
-        self.patch_size = config.patch_size
-        self.temporal_patch_size = config.temporal_patch_size
-        self.in_channels = config.in_channels
-        self.embed_dim = config.hidden_size
-
-        kernel_size = [self.temporal_patch_size, self.patch_size, self.patch_size]
-        self.proj = nn.Conv3d(self.in_channels, self.embed_dim, kernel_size=kernel_size, stride=kernel_size, bias=True)
-
-    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
-        target_dtype = self.proj.weight.dtype
-        hidden_states = hidden_states.view(
-            -1, self.in_channels, self.temporal_patch_size, self.patch_size, self.patch_size
-        )
-        hidden_states = self.proj(hidden_states.to(dtype=target_dtype)).view(-1, self.embed_dim)
-        return hidden_states
-
-
 class Qwen3OmniMoeVisionBlock(GradientCheckpointingLayer):
     def __init__(self, config, attn_implementation: str = "sdpa") -> None:
         super().__init__()
@@ -1042,8 +1058,8 @@ class Qwen3OmniMoeVisionBlock(GradientCheckpointingLayer):
         self,
         hidden_states: torch.Tensor,
         cu_seqlens: torch.Tensor,
-        rotary_pos_emb: Optional[torch.Tensor] = None,
-        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        rotary_pos_emb: torch.Tensor | None = None,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
         **kwargs,
     ) -> torch.Tensor:
         hidden_states = hidden_states + self.attn(
@@ -1057,9 +1073,34 @@ class Qwen3OmniMoeVisionBlock(GradientCheckpointingLayer):
         return hidden_states
 
 
+class Qwen3OmniMoeVisionPatchEmbed(nn.Module):
+    def __init__(self, config) -> None:
+        super().__init__()
+        self.patch_size = config.patch_size
+        self.temporal_patch_size = config.temporal_patch_size
+        self.in_channels = config.in_channels
+        self.embed_dim = config.hidden_size
+
+        kernel_size = [self.temporal_patch_size, self.patch_size, self.patch_size]
+        self.proj = nn.Conv3d(self.in_channels, self.embed_dim, kernel_size=kernel_size, stride=kernel_size, bias=True)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        target_dtype = self.proj.weight.dtype
+        hidden_states = hidden_states.view(
+            -1, self.in_channels, self.temporal_patch_size, self.patch_size, self.patch_size
+        )
+        hidden_states = self.proj(hidden_states.to(dtype=target_dtype)).view(-1, self.embed_dim)
+        return hidden_states
+
+
 class Qwen3OmniMoeVisionEncoder(Qwen3OmniMoePreTrainedModel):
     config: Qwen3OmniMoeVisionEncoderConfig
     _no_split_modules = ["Qwen3OmniMoeVisionBlock"]
+    _can_record_outputs = {
+        "router_logits": OutputRecorder(Qwen3OmniMoeTextTopKRouter, layer_name="mlp.gate", index=0),
+        "hidden_states": Qwen3OmniMoeVisionBlock,
+        "attentions": Qwen3OmniMoeVisionAttention,
+    }
 
     def __init__(self, config, *inputs, **kwargs) -> None:
         super().__init__(config, *inputs, **kwargs)
@@ -1197,7 +1238,10 @@ class Qwen3OmniMoeVisionEncoder(Qwen3OmniMoePreTrainedModel):
         patch_pos_embeds = torch.cat(patch_pos_embeds_permute)
         return patch_pos_embeds
 
-    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs) -> torch.Tensor:
+    @check_model_inputs
+    def forward(
+        self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs: Unpack[TransformersKwargs]
+    ) -> tuple | BaseModelOutputWithDeepstackFeatures:
         """
         Args:
             hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`):
@@ -1245,9 +1289,13 @@ class Qwen3OmniMoeVisionEncoder(Qwen3OmniMoePreTrainedModel):
                 )
                 deepstack_feature_lists.append(deepstack_feature)
 
-        hidden_states = self.merger(hidden_states)
+        merged_hidden_states = self.merger(hidden_states)
 
-        return hidden_states, deepstack_feature_lists
+        return BaseModelOutputWithDeepstackFeatures(
+            last_hidden_state=hidden_states,
+            pooler_output=merged_hidden_states,
+            deepstack_features=deepstack_feature_lists,
+        )
 
     @property
     def deepstack_merger_list(self):
@@ -1277,9 +1325,9 @@ class Qwen3OmniMoeThinkerTextRotaryEmbedding(nn.Module):
 
     @staticmethod
     def compute_default_rope_parameters(
-        config: Optional[Qwen3OmniMoeTextConfig] = None,
+        config: Qwen3OmniMoeTextConfig | None = None,
         device: Optional["torch.device"] = None,
-        seq_len: Optional[int] = None,
+        seq_len: int | None = None,
     ) -> tuple["torch.Tensor", float]:
         """
         Computes the inverse frequencies according to the original RoPE implementation
@@ -1442,7 +1490,7 @@ class Qwen3OmniMoeThinkerTextRMSNorm(nn.Module):
 
 
 @use_kernel_func_from_hub("rotary_pos_emb")
-def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
     """Applies Rotary Position Embedding to the query and key tensors.
 
     Args:
@@ -1450,8 +1498,6 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
         k (`torch.Tensor`): The key tensor.
         cos (`torch.Tensor`): The cosine part of the rotary embedding.
         sin (`torch.Tensor`): The sine part of the rotary embedding.
-        position_ids (`torch.Tensor`, *optional*):
-            Deprecated and unused.
         unsqueeze_dim (`int`, *optional*, defaults to 1):
             The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
             sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
@@ -1507,11 +1553,11 @@ class Qwen3OmniMoeThinkerTextAttention(nn.Module):
         self,
         hidden_states: torch.Tensor,
         position_embeddings: tuple[torch.Tensor, torch.Tensor],
-        attention_mask: Optional[torch.Tensor],
-        past_key_values: Optional[Cache] = None,
-        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: torch.Tensor | None,
+        past_key_values: Cache | None = None,
+        cache_position: torch.LongTensor | None = None,
         **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
         input_shape = hidden_states.shape[:-1]
         hidden_shape = (*input_shape, -1, self.head_dim)
 
@@ -1527,9 +1573,9 @@ class Qwen3OmniMoeThinkerTextAttention(nn.Module):
             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]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         attn_output, attn_weights = attention_interface(
             self,
@@ -1581,12 +1627,12 @@ class Qwen3OmniMoeThinkerTextDecoderLayer(GradientCheckpointingLayer):
     def forward(
         self,
         hidden_states: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[Cache] = None,
-        use_cache: Optional[bool] = False,
-        cache_position: Optional[torch.LongTensor] = None,
-        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        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
@@ -1701,18 +1747,18 @@ class Qwen3OmniMoeThinkerTextModel(Qwen3OmniMoePreTrainedModel):
     @auto_docstring
     def forward(
         self,
-        input_ids: Optional[torch.LongTensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[Cache] = None,
-        inputs_embeds: Optional[torch.FloatTensor] = None,
-        use_cache: Optional[bool] = None,
-        cache_position: Optional[torch.LongTensor] = None,
+        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,
         # args for deepstack
-        visual_pos_masks: Optional[torch.Tensor] = None,
-        deepstack_visual_embeds: Optional[list[torch.Tensor]] = None,
+        visual_pos_masks: torch.Tensor | None = None,
+        deepstack_visual_embeds: list[torch.Tensor] | None = None,
         **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> Union[tuple, BaseModelOutputWithPast]:
+    ) -> tuple | BaseModelOutputWithPast:
         r"""
         visual_pos_masks (`torch.Tensor` of shape `(batch_size, seqlen)`, *optional*):
             The mask of the visual positions.
@@ -1810,15 +1856,15 @@ class Qwen3OmniMoeThinkerCausalLMOutputWithPast(MoeCausalLMOutputWithPast):
             The rope index difference between sequence length and multimodal rope.
     """
 
-    rope_deltas: Optional[torch.LongTensor] = None
+    rope_deltas: torch.LongTensor | None = None
 
 
 def load_balancing_loss_func(
-    gate_logits: Union[torch.Tensor, tuple[torch.Tensor], None],
-    num_experts: Optional[int] = None,
+    gate_logits: torch.Tensor | tuple[torch.Tensor] | None,
+    num_experts: int | None = None,
     top_k=2,
-    attention_mask: Optional[torch.Tensor] = None,
-) -> Union[torch.Tensor, int]:
+    attention_mask: torch.Tensor | None = None,
+) -> torch.Tensor | int:
     r"""
     Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
 
@@ -1923,7 +1969,6 @@ class Qwen3OmniMoeThinkerForConditionalGeneration(
         self.vocab_size = config.text_config.vocab_size
         self.model = Qwen3OmniMoeThinkerTextModel._from_config(config.text_config)
         self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)
-        self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
         self.spatial_merge_size = config.vision_config.spatial_merge_size
         self.rope_deltas = None
         self.num_experts = config.text_config.num_experts
@@ -1937,52 +1982,56 @@ class Qwen3OmniMoeThinkerForConditionalGeneration(
     def set_input_embeddings(self, value):
         self.model.set_input_embeddings(value)
 
+    @can_return_tuple
+    @auto_docstring
     def get_video_features(
-        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: Optional[torch.LongTensor] = None
-    ):
-        """
-        Encodes videos into continuous embeddings that can be forwarded to the language model.
-
-        Args:
-            pixel_values_videos (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
-                The tensors corresponding to the input videos.
-            video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
-                The temporal, height and width of feature shape of each video in LLM.
+        self,
+        pixel_values_videos: torch.FloatTensor,
+        video_grid_thw: torch.LongTensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithDeepstackFeatures:
+        r"""
+        pixel_values_videos (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
+            The tensors corresponding to the input videos.
+        video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
+            The temporal, height and width of feature shape of each video in LLM.
         """
         pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
-        video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
-        return video_embeds
-
-    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: Optional[torch.LongTensor] = None):
-        """
-        Encodes images into continuous embeddings that can be forwarded to the language model.
+        return self.visual(pixel_values_videos, grid_thw=video_grid_thw, **kwargs)
 
-        Args:
-            pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
-                The tensors corresponding to the input images.
-            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
-                The temporal, height and width of feature shape of each image in LLM.
+    @can_return_tuple
+    @auto_docstring
+    def get_image_features(
+        self,
+        pixel_values: torch.FloatTensor,
+        image_grid_thw: torch.LongTensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithDeepstackFeatures:
+        r"""
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
+            The tensors corresponding to the input images.
+        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+            The temporal, height and width of feature shape of each image in LLM.
         """
         pixel_values = pixel_values.type(self.visual.dtype)
-        image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
-        return image_embeds
+        return self.visual(pixel_values, grid_thw=image_grid_thw, **kwargs)
 
+    @can_return_tuple
+    @auto_docstring
     def get_audio_features(
         self,
         input_features: torch.FloatTensor,
-        feature_attention_mask: Optional[torch.LongTensor] = None,
-        audio_feature_lengths: Optional[torch.LongTensor] = None,
-    ):
-        """
-        Encodes audios into continuous embeddings that can be forwarded to the language model.
-
-        Args:
-            input_features (`torch.FloatTensor`):
-                The tensors corresponding to the input audios.
-            feature_attention_mask (`torch.LongTensor`, *optional*):
-                Mask to avoid performing attention on padding feature indices. Mask values selected in `[0, 1]`:
-            audio_feature_lengths (`torch.LongTensor` of shape `(num_audios)`, *optional*):
-                The length of feature shape of each audio in LLM.
+        feature_attention_mask: torch.LongTensor | None = None,
+        audio_feature_lengths: torch.LongTensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPooling:
+        r"""
+        input_features (`torch.FloatTensor`):
+            The tensors corresponding to the input audios.
+        feature_attention_mask (`torch.LongTensor`, *optional*):
+            Mask to avoid performing attention on padding feature indices. Mask values selected in `[0, 1]`:
+        audio_feature_lengths (`torch.LongTensor` of shape `(num_audios)`, *optional*):
+            The length of feature shape of each audio in LLM.
         """
         if feature_attention_mask is not None:
             audio_feature_lengths = torch.sum(feature_attention_mask, dim=1)
@@ -1994,17 +2043,18 @@ class Qwen3OmniMoeThinkerForConditionalGeneration(
         audio_outputs = self.audio_tower(
             input_features,
             feature_lens=feature_lens,
+            return_dict=True,
+            **kwargs,
         )
-        audio_features = audio_outputs.last_hidden_state
 
-        return audio_features
+        return audio_outputs
 
     def get_placeholder_mask(
         self,
         input_ids: torch.LongTensor,
         inputs_embeds: torch.FloatTensor,
-        image_features: Optional[torch.FloatTensor] = None,
-        video_features: Optional[torch.FloatTensor] = None,
+        image_features: torch.FloatTensor | None = None,
+        video_features: torch.FloatTensor | None = None,
     ):
         """
         Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is
@@ -2032,16 +2082,18 @@ class Qwen3OmniMoeThinkerForConditionalGeneration(
 
         n_image_tokens = special_image_mask.sum()
         special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
-        if image_features is not None and inputs_embeds[special_image_mask].numel() != image_features.numel():
-            raise ValueError(
-                f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {image_features.shape[0]}"
+        if image_features is not None:
+            torch_compilable_check(
+                inputs_embeds[special_image_mask].numel() == image_features.numel(),
+                f"Image features and image tokens do not match, tokens: {n_image_tokens}, features: {image_features.shape[0]}",
             )
 
         n_video_tokens = special_video_mask.sum()
         special_video_mask = special_video_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
-        if video_features is not None and inputs_embeds[special_video_mask].numel() != video_features.numel():
-            raise ValueError(
-                f"Videos features and video tokens do not match: tokens: {n_video_tokens}, features {video_features.shape[0]}"
+        if video_features is not None:
+            torch_compilable_check(
+                inputs_embeds[special_video_mask].numel() == video_features.numel(),
+                f"Video features and video tokens do not match, tokens: {n_video_tokens}, features: {video_features.shape[0]}",
             )
 
         special_audio_mask = special_audio_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
@@ -2066,12 +2118,12 @@ class Qwen3OmniMoeThinkerForConditionalGeneration(
         rope_deltas=None,
         labels=None,
         use_cache=None,
-        output_router_logits: Optional[bool] = None,
+        output_router_logits: bool | None = None,
         use_audio_in_video=None,
         cache_position=None,
         video_second_per_grid=None,
         **kwargs,
-    ) -> Union[tuple, Qwen3OmniMoeThinkerCausalLMOutputWithPast]:
+    ) -> tuple | Qwen3OmniMoeThinkerCausalLMOutputWithPast:
         r"""
         image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
             The temporal, height and width of feature shape of each image in LLM.
@@ -2144,13 +2196,18 @@ class Qwen3OmniMoeThinkerForConditionalGeneration(
                 input_features,
                 feature_attention_mask=feature_attention_mask,
                 audio_feature_lengths=audio_feature_lengths,
-            )
+                return_dict=True,
+            ).last_hidden_state
             audio_features = audio_features.to(inputs_embeds.device, inputs_embeds.dtype)
             _, _, audio_mask = self.get_placeholder_mask(input_ids, inputs_embeds=inputs_embeds)
             inputs_embeds = inputs_embeds.masked_scatter(audio_mask, audio_features)
 
         if pixel_values is not None:
-            image_embeds, image_embeds_multiscale = self.get_image_features(pixel_values, image_grid_thw)
+            image_outputs: BaseModelOutputWithDeepstackFeatures = self.get_image_features(
+                pixel_values, image_grid_thw, return_dict=True
+            )
+            image_embeds = image_outputs.pooler_output
+            image_embeds_multiscale = image_outputs.deepstack_features
             image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
             image_mask, _, _ = self.get_placeholder_mask(
                 input_ids, inputs_embeds=inputs_embeds, image_features=image_embeds
@@ -2158,8 +2215,11 @@ class Qwen3OmniMoeThinkerForConditionalGeneration(
             inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
 
         if pixel_values_videos is not None:
-            video_embeds, video_embeds_multiscale = self.get_video_features(pixel_values_videos, video_grid_thw)
-
+            video_outputs: BaseModelOutputWithDeepstackFeatures = self.get_video_features(
+                pixel_values_videos, video_grid_thw, return_dict=True
+            )
+            video_embeds = video_outputs.pooler_output
+            video_embeds_multiscale = video_outputs.deepstack_features
             video_embeds = video_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
             _, video_mask, _ = self.get_placeholder_mask(
                 input_ids, inputs_embeds=inputs_embeds, video_features=video_embeds
@@ -2327,7 +2387,7 @@ class Qwen3OmniMoeTalkerCodePredictorOutputWithPast(CausalLMOutputWithPast):
         Current generation step of code predictor model.
     """
 
-    generation_steps: Optional[int] = None
+    generation_steps: int | None = None
 
 
 @use_kernel_forward_from_hub("RMSNorm")
@@ -2388,11 +2448,11 @@ class Qwen3OmniMoeTalkerCodePredictorAttention(nn.Module):
         self,
         hidden_states: torch.Tensor,
         position_embeddings: tuple[torch.Tensor, torch.Tensor],
-        attention_mask: Optional[torch.Tensor],
-        past_key_values: Optional[Cache] = None,
-        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: torch.Tensor | None,
+        past_key_values: Cache | None = None,
+        cache_position: torch.LongTensor | None = None,
         **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
         input_shape = hidden_states.shape[:-1]
         hidden_shape = (*input_shape, -1, self.head_dim)
 
@@ -2408,9 +2468,9 @@ class Qwen3OmniMoeTalkerCodePredictorAttention(nn.Module):
             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]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         attn_output, attn_weights = attention_interface(
             self,
@@ -2459,12 +2519,12 @@ class Qwen3OmniMoeTalkerCodePredictorDecoderLayer(GradientCheckpointingLayer):
     def forward(
         self,
         hidden_states: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[Cache] = None,
-        use_cache: Optional[bool] = False,
-        cache_position: Optional[torch.LongTensor] = None,
-        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        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
@@ -2511,9 +2571,9 @@ class Qwen3OmniMoeRotaryEmbedding(nn.Module):
 
     @staticmethod
     def compute_default_rope_parameters(
-        config: Optional[Qwen3OmniMoeConfig] = None,
+        config: Qwen3OmniMoeConfig | None = None,
         device: Optional["torch.device"] = None,
-        seq_len: Optional[int] = None,
+        seq_len: int | None = None,
     ) -> tuple["torch.Tensor", float]:
         """
         Computes the inverse frequencies according to the original RoPE implementation
@@ -2589,13 +2649,13 @@ class Qwen3OmniMoeTalkerCodePredictorModel(Qwen3OmniMoePreTrainedModel):
     @auto_docstring
     def forward(
         self,
-        input_ids: Optional[torch.LongTensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[Cache] = None,
-        inputs_embeds: Optional[torch.FloatTensor] = None,
-        use_cache: Optional[bool] = None,
-        cache_position: Optional[torch.LongTensor] = None,
+        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],
     ) -> BaseModelOutputWithPast:
         if input_ids is not None:
@@ -2659,7 +2719,7 @@ class Qwen3OmniMoeTalkerCodePredictorModel(Qwen3OmniMoePreTrainedModel):
 @auto_docstring
 class Qwen3OmniMoeTalkerCodePredictorModelForConditionalGeneration(Qwen3OmniMoePreTrainedModel, 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"])}
     config_class = Qwen3OmniMoeTalkerCodePredictorConfig
     base_model_prefix = "talker.code_predictor"
@@ -2752,7 +2812,7 @@ class Qwen3OmniMoeTalkerOutputWithPast(MoeCausalLMOutputWithPast):
         Current generation step, used to track which `trailing_text_hidden` should be used.
     """
 
-    generation_step: Optional[int] = None
+    generation_step: int | None = None
 
 
 class Qwen3OmniMoeTalkerRotaryEmbedding(Qwen3OmniMoeThinkerTextRotaryEmbedding):
@@ -2878,12 +2938,12 @@ class Qwen3OmniMoeTalkerDecoderLayer(GradientCheckpointingLayer):
     def forward(
         self,
         hidden_states: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[Cache] = None,
-        use_cache: Optional[bool] = False,
-        cache_position: Optional[torch.LongTensor] = None,
-        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        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
@@ -2945,18 +3005,18 @@ class Qwen3OmniMoeTalkerModel(Qwen3OmniMoePreTrainedModel):
     @auto_docstring
     def forward(
         self,
-        input_ids: Optional[torch.LongTensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[Cache] = None,
-        inputs_embeds: Optional[torch.FloatTensor] = None,
-        use_cache: Optional[bool] = None,
-        cache_position: Optional[torch.LongTensor] = None,
+        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,
         # args for deepstack
-        visual_pos_masks: Optional[torch.Tensor] = None,
-        deepstack_visual_embeds: Optional[list[torch.Tensor]] = None,
+        visual_pos_masks: torch.Tensor | None = None,
+        deepstack_visual_embeds: list[torch.Tensor] | None = None,
         **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> Union[tuple, BaseModelOutputWithPast]:
+    ) -> tuple | BaseModelOutputWithPast:
         r"""
         visual_pos_masks (`torch.Tensor` of shape `(batch_size, seqlen)`, *optional*):
             The mask of the visual positions.
@@ -3052,7 +3112,7 @@ class Qwen3OmniMoeTalkerModel(Qwen3OmniMoePreTrainedModel):
 @auto_docstring
 class Qwen3OmniMoeTalkerForConditionalGeneration(Qwen3OmniMoeThinkerTextPreTrainedModel, GenerationMixin):
     _tied_weights_keys = {"codec_head": "model.codec_embedding.weight"}
-    _tp_plan = {"codec_head": "colwise_rep"}
+    _tp_plan = {"codec_head": "colwise_gather_output"}
     _pp_plan = {"codec_head": (["hidden_states"], ["logits"])}
     config_class = Qwen3OmniMoeTalkerConfig
     base_model_prefix = "talker"
@@ -3200,13 +3260,13 @@ class Qwen3OmniMoeTalkerForConditionalGeneration(Qwen3OmniMoeThinkerTextPreTrain
     # Should inherit from PretrainedModel, but cannot inherit multiple classes in modular
     def get_rope_index(
         self,
-        input_ids: Optional[torch.LongTensor] = None,
-        image_grid_thw: Optional[torch.LongTensor] = None,
-        video_grid_thw: Optional[torch.LongTensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
+        input_ids: torch.LongTensor | None = None,
+        image_grid_thw: torch.LongTensor | None = None,
+        video_grid_thw: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
         use_audio_in_video: bool = False,
-        audio_seqlens: Optional[torch.LongTensor] = None,
-        second_per_grids: Optional[torch.Tensor] = None,
+        audio_seqlens: torch.LongTensor | None = None,
+        second_per_grids: torch.Tensor | None = None,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         return Qwen3OmniMoePreTrainedModelForConditionalGeneration.get_rope_index(
             self,
@@ -3429,11 +3489,11 @@ class Qwen3OmniMoeCode2WavAttention(nn.Module):
         self,
         hidden_states: torch.Tensor,
         position_embeddings: tuple[torch.Tensor, torch.Tensor],
-        attention_mask: Optional[torch.Tensor],
-        past_key_values: Optional[Cache] = None,
-        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: torch.Tensor | None,
+        past_key_values: Cache | None = None,
+        cache_position: torch.LongTensor | None = None,
         **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
         input_shape = hidden_states.shape[:-1]
         hidden_shape = (*input_shape, -1, self.head_dim)
 
@@ -3449,9 +3509,9 @@ class Qwen3OmniMoeCode2WavAttention(nn.Module):
             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]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         attn_output, attn_weights = attention_interface(
             self,
@@ -3537,13 +3597,13 @@ class Qwen3OmniMoeCode2WavTransformerLayer(GradientCheckpointingLayer):
     def forward(
         self,
         hidden_states: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[Cache] = None,
-        use_cache: Optional[bool] = False,
-        cache_position: Optional[torch.LongTensor] = None,
+        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,
         **kwargs,
-    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
+    ) -> tuple[torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]:
         """
         Args:
             hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
@@ -3936,10 +3996,10 @@ class Qwen3OmniMoeForConditionalGeneration(Qwen3OmniMoePreTrainedModel, Generati
     @torch.no_grad()
     def generate(
         self,
-        input_ids: Optional[torch.Tensor] = None,
+        input_ids: torch.Tensor | None = None,
         speaker: str = "Ethan",
         use_audio_in_video: bool = False,
-        return_audio: Optional[bool] = None,
+        return_audio: bool | None = None,
         thinker_max_new_tokens: int = 1024,
         thinker_eos_token_id: int = 151645,
         talker_max_new_tokens: int = 4096,