transformers 5.0.0rc3__py3-none-any.whl → 5.1.0__py3-none-any.whl

transformers/models/qwen3_omni_moe/configuration_qwen3_omni_moe.py

@@ -209,8 +209,6 @@ class Qwen3OmniMoeTextConfig(PreTrainedConfig):
         use_cache (`bool`, *optional*, defaults to `True`):
             Whether or not the model should return the last key/values attentions (not used by all models). Only
             relevant if `config.is_decoder=True`.
-        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
-            Whether the model's input and output word embeddings should be tied.
         rope_parameters (`RopeParameters`, *optional*):
             Dictionary containing the configuration parameters for the RoPE embeddings. The dictionary should contain
             a value for `rope_theta` and optionally parameters used for scaling in case you want to use RoPE
@@ -242,6 +240,12 @@ class Qwen3OmniMoeTextConfig(PreTrainedConfig):
             Indicate which layers use Qwen3OmniMoeTextMLP rather than Qwen3OmniMoeTextSparseMoeBlock
             The list contains layer index, from 0 to num_layers-1 if we have num_layers layers
             If `mlp_only_layers` is empty, `decoder_sparse_step` is used to determine the sparsity.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        bos_token_id (`int`, *optional*):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*):
+            End of stream token id.
 
     ```python
     >>> from transformers import Qwen3OmniMoeTextModel, Qwen3OmniMoeTextConfig
@@ -266,9 +270,8 @@ class Qwen3OmniMoeTextConfig(PreTrainedConfig):
         "layers.*.self_attn.k_proj": "colwise",
         "layers.*.self_attn.v_proj": "colwise",
         "layers.*.self_attn.o_proj": "rowwise",
-        "layers.*.mlp.experts.gate_up_proj": "local_rowwise",
-        "layers.*.mlp.experts.down_proj": "local_rowwise",
-        "layers.*.mlp.experts": "gather",
+        "layers.*.mlp.experts.gate_up_proj": "packed_colwise",
+        "layers.*.mlp.experts.down_proj": "rowwise",
         "layers.*.mlp.gate_proj": "colwise",
         "layers.*.mlp.up_proj": "colwise",
         "layers.*.mlp.down_proj": "rowwise",
@@ -292,7 +295,6 @@ class Qwen3OmniMoeTextConfig(PreTrainedConfig):
         initializer_range: float | None = 0.02,
         rms_norm_eps: float | None = 1e-6,
         use_cache: bool | None = True,
-        tie_word_embeddings: bool | None = False,
         rope_parameters: RopeParameters | dict[str, RopeParameters] | None = None,
         attention_bias: bool | None = False,
         sliding_window: int | None = None,
@@ -305,6 +307,9 @@ class Qwen3OmniMoeTextConfig(PreTrainedConfig):
         output_router_logits: bool | None = False,
         router_aux_loss_coef: float | None = 0.001,
         mlp_only_layers: list[int] | None = None,
+        pad_token_id: int | None = None,
+        bos_token_id: int | None = None,
+        eos_token_id: int | None = None,
         **kwargs,
     ):
         self.vocab_size = vocab_size
@@ -333,9 +338,11 @@ class Qwen3OmniMoeTextConfig(PreTrainedConfig):
         self.output_router_logits = output_router_logits
         self.router_aux_loss_coef = router_aux_loss_coef
         self.mlp_only_layers = [] if mlp_only_layers is None else mlp_only_layers
+        self.pad_token_id = pad_token_id
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
 
         super().__init__(
-            tie_word_embeddings=tie_word_embeddings,
             ignore_keys_at_rope_validation={"mrope_section", "interleaved", "mrope_interleaved"},
             **kwargs,
         )
@@ -374,6 +381,8 @@ class Qwen3OmniMoeThinkerConfig(PreTrainedConfig):
             The user token id to encode the user token.
         initializer_range (`float`, *optional*, defaults to 0.02):
             The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
 
     Example:
 
@@ -411,12 +420,14 @@ class Qwen3OmniMoeThinkerConfig(PreTrainedConfig):
         audio_start_token_id=151647,
         user_token_id=872,
         initializer_range=0.02,
+        tie_word_embeddings=False,
         **kwargs,
     ):
         self.user_token_id = user_token_id
         self.position_id_per_seconds = position_id_per_seconds
         self.audio_start_token_id = audio_start_token_id
         self.initializer_range = initializer_range
+        self.tie_word_embeddings = tie_word_embeddings
 
         if isinstance(vision_config, dict):
             vision_config = Qwen3OmniMoeVisionEncoderConfig(**vision_config)
@@ -504,6 +515,12 @@ class Qwen3OmniMoeTalkerCodePredictorConfig(PreTrainedConfig):
             Attention pattern for each layer.
         attention_dropout (`float`, *optional*, defaults to 0.0):
             The dropout ratio for the attention probabilities.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        bos_token_id (`int`, *optional*):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*):
+            End of stream token id.
 
     ```python
     >>> from transformers import Qwen3OmniMoeTalkerCodePredictorModel, Qwen3OmniMoeTalkerCodePredictorConfig
@@ -555,11 +572,16 @@ class Qwen3OmniMoeTalkerCodePredictorConfig(PreTrainedConfig):
         rope_parameters: int | None = None,
         attention_bias: bool | None = False,
         sliding_window: int | None = None,
+        max_window_layers: int | None = 28,
         layer_types: list[str] | None = None,
         attention_dropout: int | None = 0,
         num_code_groups: int | None = 32,
+        pad_token_id: int | None = None,
+        bos_token_id: int | None = None,
+        eos_token_id: int | None = None,
         **kwargs,
     ):
+        self.num_code_groups = num_code_groups
         self.vocab_size = vocab_size
         self.max_position_embeddings = max_position_embeddings
         self.hidden_size = hidden_size
@@ -567,6 +589,7 @@ class Qwen3OmniMoeTalkerCodePredictorConfig(PreTrainedConfig):
         self.num_hidden_layers = num_hidden_layers
         self.num_attention_heads = num_attention_heads
         self.sliding_window = sliding_window
+        self.max_window_layers = max_window_layers
 
         # for backward compatibility
         if num_key_value_heads is None:
@@ -591,13 +614,13 @@ class Qwen3OmniMoeTalkerCodePredictorConfig(PreTrainedConfig):
             ]
         layer_type_validation(self.layer_types, self.num_hidden_layers)
 
+        self.pad_token_id = pad_token_id
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.tie_word_embeddings = tie_word_embeddings
         self.rope_parameters = rope_parameters
 
-        super().__init__(
-            tie_word_embeddings=tie_word_embeddings,
-            **kwargs,
-        )
-        self.num_code_groups = num_code_groups
+        super().__init__(**kwargs)
 
 
 class Qwen3OmniMoeTalkerTextConfig(PreTrainedConfig):
@@ -674,6 +697,12 @@ class Qwen3OmniMoeTalkerTextConfig(PreTrainedConfig):
             Indicate which layers use Qwen3OmniMoeTalkerTextMLP rather than Qwen3OmniMoeTalkerTextSparseMoeBlock
             The list contains layer index, from 0 to num_layers-1 if we have num_layers layers
             If `mlp_only_layers` is empty, `decoder_sparse_step` is used to determine the sparsity.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        bos_token_id (`int`, *optional*):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*):
+            End of stream token id.
 
     ```python
     >>> from transformers import Qwen3OmniMoeTalkerTextModel, Qwen3OmniMoeTalkerTextConfig
@@ -691,15 +720,18 @@ class Qwen3OmniMoeTalkerTextConfig(PreTrainedConfig):
     model_type = "qwen3_omni_moe_talker_text"
     keys_to_ignore_at_inference = ["past_key_values"]
 
+    attribute_map = {
+        "num_experts": "num_local_experts",
+    }
+
     # Default tensor parallel plan for base model `Qwen3OmniMoeTalkerText`
     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.o_proj": "rowwise",
-        "layers.*.mlp.experts.gate_up_proj": "local_rowwise",
-        "layers.*.mlp.experts.down_proj": "local_rowwise",
-        "layers.*.mlp.experts": "gather",
+        "layers.*.mlp.experts.gate_up_proj": "packed_colwise",
+        "layers.*.mlp.experts.down_proj": "rowwise",
         "layers.*.mlp.gate_proj": "colwise",
         "layers.*.mlp.up_proj": "colwise",
         "layers.*.mlp.down_proj": "rowwise",
@@ -736,6 +768,9 @@ class Qwen3OmniMoeTalkerTextConfig(PreTrainedConfig):
         output_router_logits: bool | None = False,
         router_aux_loss_coef: float | None = 0.001,
         mlp_only_layers: list[int] | None = None,
+        pad_token_id: int | None = None,
+        bos_token_id: int | None = None,
+        eos_token_id: int | None = None,
         **kwargs,
     ):
         self.vocab_size = vocab_size
@@ -765,10 +800,11 @@ class Qwen3OmniMoeTalkerTextConfig(PreTrainedConfig):
         self.router_aux_loss_coef = router_aux_loss_coef
         self.mlp_only_layers = [] if mlp_only_layers is None else mlp_only_layers
 
-        super().__init__(
-            tie_word_embeddings=tie_word_embeddings,
-            **kwargs,
-        )
+        self.pad_token_id = pad_token_id
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.tie_word_embeddings = tie_word_embeddings
+        super().__init__(**kwargs)
 
 
 class Qwen3OmniMoeTalkerConfig(PreTrainedConfig):

transformers/models/qwen3_omni_moe/modeling_qwen3_omni_moe.py

@@ -44,8 +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,
@@ -53,8 +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,
@@ -68,6 +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__()
@@ -578,9 +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,
@@ -670,6 +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)
@@ -716,7 +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]
@@ -730,6 +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,
@@ -747,11 +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
@@ -803,7 +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"):
         """
@@ -907,11 +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(
@@ -996,6 +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__()
@@ -1009,26 +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__()
@@ -1056,9 +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)
@@ -1196,7 +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)`):
@@ -1244,9 +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):
@@ -1524,9 +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,
@@ -1920,7 +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
@@ -1934,52 +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: torch.LongTensor | None = 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: torch.LongTensor | None = 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: torch.LongTensor | None = None,
         audio_feature_lengths: torch.LongTensor | None = 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.
+        **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)
@@ -1991,10 +2043,11 @@ 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,
@@ -2029,16 +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)
@@ -2141,13 +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
@@ -2155,8 +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
@@ -2405,9 +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,
@@ -2656,7 +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"
@@ -3049,7 +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"
@@ -3446,9 +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,