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

transformers/models/qwen2_5_omni/modular_qwen2_5_omni.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2025 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
 #
 #
@@ -18,7 +17,7 @@
 import math
 from collections.abc import Callable
 from dataclasses import dataclass
-from typing import Any, Optional, Union
+from typing import Any, Optional
 
 import numpy as np
 import torch
@@ -30,17 +29,20 @@ from ... import initialization as init
 from ...cache_utils import Cache
 from ...configuration_utils import PreTrainedConfig, layer_type_validation
 from ...generation import GenerationMixin
-from ...modeling_outputs import BaseModelOutput, ModelOutput
+from ...modeling_outputs import BaseModelOutputWithPooling, ModelOutput
 from ...modeling_rope_utils import RopeParameters
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
 from ...utils import (
     TransformersKwargs,
     auto_docstring,
+    can_return_tuple,
     check_torch_load_is_safe,
     logging,
+    torch_compilable_check,
 )
 from ...utils.deprecation import deprecate_kwarg
+from ...utils.generic import check_model_inputs, is_flash_attention_requested
 from ...utils.hub import cached_file
 from ..llama.modeling_llama import LlamaRotaryEmbedding, rotate_half
 from ..qwen2_5_vl.configuration_qwen2_5_vl import Qwen2_5_VLVisionConfig
@@ -346,24 +348,26 @@ class Qwen2_5OmniTextConfig(PreTrainedConfig):
 
     def __init__(
         self,
-        vocab_size: Optional[int] = 152064,
-        hidden_size: Optional[int] = 3584,
-        intermediate_size: Optional[int] = 18944,
-        num_hidden_layers: Optional[int] = 28,
-        num_attention_heads: Optional[int] = 28,
-        num_key_value_heads: Optional[int] = 4,
-        hidden_act: Optional[str] = "silu",
-        max_position_embeddings: Optional[int] = 32768,
-        initializer_range: Optional[float] = 0.02,
-        rms_norm_eps: Optional[int] = 1e-6,
-        use_cache: Optional[bool] = True,
-        tie_word_embeddings: Optional[bool] = False,
-        rope_parameters: Optional[RopeParameters | dict[str, RopeParameters]] = None,
-        use_sliding_window: Optional[bool] = False,
-        sliding_window: Optional[int] = 32768,
-        max_window_layers: Optional[int] = 28,
-        layer_types: Optional[list[str]] = None,
-        attention_dropout: Optional[float] = 0.0,
+        vocab_size: int | None = 152064,
+        hidden_size: int | None = 3584,
+        intermediate_size: int | None = 18944,
+        num_hidden_layers: int | None = 28,
+        num_attention_heads: int | None = 28,
+        num_key_value_heads: int | None = 4,
+        hidden_act: str | None = "silu",
+        max_position_embeddings: int | None = 32768,
+        initializer_range: float | None = 0.02,
+        rms_norm_eps: int | None = 1e-6,
+        use_cache: bool | None = True,
+        rope_parameters: RopeParameters | dict[str, RopeParameters] | None = None,
+        use_sliding_window: bool | None = False,
+        sliding_window: int | None = 32768,
+        max_window_layers: int | None = 28,
+        layer_types: list[str] | None = None,
+        attention_dropout: float | None = 0.0,
+        pad_token_id: int | None = None,
+        bos_token_id: int | None = None,
+        eos_token_id: int | None = None,
         **kwargs,
     ):
         self.vocab_size = vocab_size
@@ -375,6 +379,9 @@ class Qwen2_5OmniTextConfig(PreTrainedConfig):
         self.use_sliding_window = use_sliding_window
         self.sliding_window = sliding_window if self.use_sliding_window else None
         self.max_window_layers = max_window_layers
+        self.pad_token_id = pad_token_id
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
 
         # for backward compatibility
         if num_key_value_heads is None:
@@ -399,7 +406,6 @@ class Qwen2_5OmniTextConfig(PreTrainedConfig):
 
         self.rope_parameters = rope_parameters
         super().__init__(
-            tie_word_embeddings=tie_word_embeddings,
             ignore_keys_at_rope_validation={"mrope_section"},
             **kwargs,
         )
@@ -441,6 +447,8 @@ class Qwen2_5OmniThinkerConfig(PreTrainedConfig):
             The user token index 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:
 
@@ -492,6 +500,7 @@ class Qwen2_5OmniThinkerConfig(PreTrainedConfig):
         audio_end_token_id=151648,
         user_token_id=872,
         initializer_range=0.02,
+        tie_word_embeddings=False,
         **kwargs,
     ):
         self.audio_token_index = audio_token_index
@@ -503,6 +512,7 @@ class Qwen2_5OmniThinkerConfig(PreTrainedConfig):
         self.audio_start_token_id = audio_start_token_id
         self.audio_end_token_id = audio_end_token_id
         self.initializer_range = initializer_range
+        self.tie_word_embeddings = tie_word_embeddings
 
         if isinstance(vision_config, dict):
             vision_config = Qwen2_5OmniVisionEncoderConfig(**vision_config)
@@ -621,6 +631,9 @@ class Qwen2_5OmniTalkerConfig(PreTrainedConfig):
             The size used for merging spatial dimensions.
         layer_types (`list`, *optional*):
             Attention pattern for each layer.
+        pad_token_id (`int`, *optional*):
+            The id of the padding token.
+
 
     Example:
 
@@ -682,7 +695,7 @@ class Qwen2_5OmniTalkerConfig(PreTrainedConfig):
         sliding_window=32768,
         max_window_layers=28,
         attention_dropout=0.0,
-        rope_parameters: Optional[RopeParameters | dict[str, RopeParameters]] = None,
+        rope_parameters: RopeParameters | dict[str, RopeParameters] | None = None,
         position_id_per_seconds=25,
         seconds_per_chunk=2,
         audio_start_token_id=151647,
@@ -690,6 +703,7 @@ class Qwen2_5OmniTalkerConfig(PreTrainedConfig):
         initializer_range=0.02,
         spatial_merge_size=2,
         layer_types=None,
+        pad_token_id: int | None = None,
         **kwargs,
     ):
         self.audio_token_index = audio_token_index
@@ -733,9 +747,11 @@ class Qwen2_5OmniTalkerConfig(PreTrainedConfig):
         self.seconds_per_chunk = seconds_per_chunk  # zf
         self.audio_start_token_id = audio_start_token_id  # zf
         self.audio_end_token_id = audio_end_token_id  # zf
+        self.pad_token_id = pad_token_id
 
         self.initializer_range = initializer_range
         self.spatial_merge_size = spatial_merge_size
+        self.tie_word_embeddings = tie_word_embeddings
 
         self.layer_types = layer_types
         if self.layer_types is None:
@@ -748,9 +764,7 @@ class Qwen2_5OmniTalkerConfig(PreTrainedConfig):
         layer_type_validation(self.layer_types, self.num_hidden_layers)
 
         self.rope_parameters = rope_parameters
-        super().__init__(
-            tie_word_embeddings=tie_word_embeddings, ignore_keys_at_rope_validation={"mrope_section"}, **kwargs
-        )
+        super().__init__(ignore_keys_at_rope_validation={"mrope_section"}, **kwargs)
 
 
 class Qwen2_5OmniDiTConfig(PreTrainedConfig):
@@ -808,7 +822,7 @@ class Qwen2_5OmniDiTConfig(PreTrainedConfig):
         ff_mult=2,
         emb_dim=512,
         head_dim=64,
-        rope_parameters: Optional[RopeParameters | dict[str, RopeParameters]] = None,
+        rope_parameters: RopeParameters | dict[str, RopeParameters] | None = None,
         max_position_embeddings=32768,
         block_size=24,
         look_ahead_layers=[10],
@@ -1190,13 +1204,13 @@ class Qwen2_5OmniPreTrainedModelForConditionalGeneration(Qwen2_5OmniPreTrainedMo
 
     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.
@@ -1504,12 +1518,12 @@ class Qwen2_5OmniThinkerCausalLMOutputWithPast(ModelOutput):
         The rope index difference between sequence length and multimodal rope.
     """
 
-    loss: Optional[torch.FloatTensor] = None
-    logits: Optional[torch.FloatTensor] = None
-    past_key_values: Optional[Cache] = None
-    hidden_states: Optional[tuple[torch.FloatTensor]] = None
-    attentions: Optional[tuple[torch.FloatTensor]] = None
-    rope_deltas: Optional[torch.LongTensor] = None
+    loss: torch.FloatTensor | None = None
+    logits: torch.FloatTensor | None = None
+    past_key_values: Cache | None = None
+    hidden_states: tuple[torch.FloatTensor] | None = None
+    attentions: tuple[torch.FloatTensor] | None = None
+    rope_deltas: torch.LongTensor | None = None
 
 
 class Qwen2_5OmniAudioAttention(nn.Module):
@@ -1545,10 +1559,10 @@ class Qwen2_5OmniAudioAttention(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()
@@ -1562,9 +1576,9 @@ class Qwen2_5OmniAudioAttention(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,
@@ -1597,7 +1611,7 @@ class Qwen2_5OmniAudioEncoderLayer(Qwen2AudioEncoderLayer):
         self,
         hidden_states: torch.Tensor,
         cu_seqlens: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
+        attention_mask: torch.Tensor | None = None,
         **kwargs,
     ) -> torch.Tensor:
         residual = hidden_states
@@ -1658,6 +1672,10 @@ class Qwen2_5OmniAudioEncoder(Qwen2_5OmniPreTrainedModel):
     input_modalities = "audio"
     _no_split_modules = ["Qwen2_5OmniAudioEncoderLayer"]
     _supports_sdpa = True
+    _can_record_outputs = {
+        "hidden_states": Qwen2_5OmniAudioEncoderLayer,
+        "attentions": Qwen2_5OmniAudioAttention,
+    }
 
     def __init__(self, config: Qwen2_5OmniAudioEncoderConfig):
         super().__init__(config)
@@ -1696,7 +1714,7 @@ class Qwen2_5OmniAudioEncoder(Qwen2_5OmniPreTrainedModel):
         # 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]
@@ -1710,14 +1728,9 @@ class Qwen2_5OmniAudioEncoder(Qwen2_5OmniPreTrainedModel):
             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,
-        input_features,
-        feature_lens=None,
-        aftercnn_lens=None,
-        **kwargs,
-    ):
+    def forward(self, input_features, feature_lens=None, aftercnn_lens=None, **kwargs: Unpack[TransformersKwargs]):
         r"""
         feature_lens (`torch.LongTensor` of shape `(batch_size,)`):
             mel length
@@ -1726,11 +1739,7 @@ class Qwen2_5OmniAudioEncoder(Qwen2_5OmniPreTrainedModel):
         """
         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 = torch.where(chunk_lengths == 0, self.n_window * 2, chunk_lengths)
@@ -1771,7 +1780,7 @@ class Qwen2_5OmniAudioEncoder(Qwen2_5OmniPreTrainedModel):
             each_audio_states = self.proj(each_audio_states)
             token_audio_list.append(each_audio_states)
         token_audio = torch.cat(token_audio_list, dim=0)
-        return BaseModelOutput(last_hidden_state=token_audio)
+        return BaseModelOutputWithPooling(last_hidden_state=token_audio)
 
     def padded_and_mask_function(self, tensor_list, tensor_len, padding_value=0, padding_side="right"):
         """
@@ -1853,7 +1862,7 @@ class Qwen2_5OmniVisionAttention(nn.Module):
         self,
         hidden_states: torch.Tensor,
         cu_seqlens: torch.Tensor,
-        rotary_pos_emb: Optional[torch.Tensor] = None,
+        rotary_pos_emb: torch.Tensor | None = None,
         **kwargs,
     ) -> torch.Tensor:
         seq_length = hidden_states.shape[0]
@@ -1867,11 +1876,11 @@ class Qwen2_5OmniVisionAttention(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 self.config._attn_implementation == "flash_attention_2":
+        if is_flash_attention_requested(self.config):
             # Flash Attention 2: Use cu_seqlens for variable length attention
             max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
             attn_output, _ = attention_interface(
@@ -1926,7 +1935,7 @@ class Qwen2_5OmniVisionBlock(Qwen2_5_VLVisionBlock):
         self,
         hidden_states: torch.Tensor,
         cu_seqlens: torch.Tensor,
-        rotary_pos_emb: Optional[torch.Tensor] = None,
+        rotary_pos_emb: torch.Tensor | None = None,
         **kwargs,
     ) -> torch.Tensor:
         hidden_states = hidden_states + self.attn(
@@ -1948,12 +1957,19 @@ class Qwen2_5OmniVisionEncoder(Qwen2_5_VisionTransformerPretrainedModel):
     input_modalities = ("image", "video")
     _no_split_modules = ["Qwen2_5OmniVisionBlock"]
     _input_embed_layer = "patch_embed"
+    _can_record_outputs = {
+        "hidden_states": Qwen2_5OmniVisionBlock,
+        "attentions": Qwen2_5OmniVisionAttention,
+    }
 
     def __init__(self, config: Qwen2_5OmniVisionEncoderConfig, *inputs, **kwargs) -> None:
         super().__init__(config, *inputs, **kwargs)
         self.blocks = nn.ModuleList([Qwen2_5OmniVisionBlock(config) for _ in range(config.depth)])
 
-    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 | BaseModelOutputWithPooling:
         """
         Args:
             hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`):
@@ -2006,11 +2022,15 @@ class Qwen2_5OmniVisionEncoder(Qwen2_5_VisionTransformerPretrainedModel):
                 rotary_pos_emb=rotary_pos_emb,
                 **kwargs,
             )
-        hidden_states = self.merger(hidden_states)
+
+        merged_hidden_states = self.merger(hidden_states)
         reverse_indices = torch.argsort(window_index)
-        hidden_states = hidden_states[reverse_indices, :]
+        merged_hidden_states = merged_hidden_states[reverse_indices, :]
 
-        return hidden_states
+        return BaseModelOutputWithPooling(
+            last_hidden_state=hidden_states,
+            pooler_output=merged_hidden_states,
+        )
 
 
 class Qwen2_5OmniRotaryEmbedding(Qwen2VLRotaryEmbedding):
@@ -2021,7 +2041,7 @@ class Qwen2_5OmniRotaryEmbedding(Qwen2VLRotaryEmbedding):
 # It's same as `Qwen2_5_VLAttention`, but talker model's hidden_size isn't divisible by num_heads.
 # Removes the value error as a workaround.
 class Qwen2_5OmniAttention(Qwen2_5_VLAttention):
-    def __init__(self, config: Qwen2_5OmniConfig, layer_idx: Optional[int] = None):
+    def __init__(self, config: Qwen2_5OmniConfig, layer_idx: int | None = None):
         nn.Module.__init__(self)
         self.config = config
         self.layer_idx = layer_idx
@@ -2080,7 +2100,6 @@ class Qwen2_5OmniThinkerForConditionalGeneration(Qwen2_5OmniPreTrainedModelForCo
         self.vocab_size = config.text_config.vocab_size
         self.model = Qwen2_5OmniThinkerTextModel._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.post_init()
@@ -2091,52 +2110,56 @@ class Qwen2_5OmniThinkerForConditionalGeneration(Qwen2_5OmniPreTrainedModelForCo
     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 | BaseModelOutputWithPooling:
+        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
+        return self.visual(pixel_values_videos, grid_thw=video_grid_thw, **kwargs)
 
-    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.
-
-        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 | BaseModelOutputWithPooling:
+        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)
@@ -2152,20 +2175,20 @@ class Qwen2_5OmniThinkerForConditionalGeneration(Qwen2_5OmniPreTrainedModelForCo
             input_features,
             feature_lens=feature_lens,
             aftercnn_lens=audio_feat_lengths,
+            return_dict=True,
+            **kwargs,
         )
-        audio_features = audio_outputs.last_hidden_state
-
-        if audio_features.shape[0] != sum(audio_output_lengths.tolist()):
+        if audio_outputs.last_hidden_state.shape[0] != sum(audio_output_lengths.tolist()):
             raise ValueError("length of audio_features should match audio_output_lengths")
 
-        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
@@ -2193,16 +2216,18 @@ class Qwen2_5OmniThinkerForConditionalGeneration(Qwen2_5OmniPreTrainedModelForCo
 
         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)
@@ -2211,29 +2236,29 @@ class Qwen2_5OmniThinkerForConditionalGeneration(Qwen2_5OmniPreTrainedModelForCo
     @auto_docstring
     def forward(
         self,
-        input_ids: Optional[torch.LongTensor] = None,
-        input_features: Optional[torch.FloatTensor] = None,
-        pixel_values: Optional[torch.FloatTensor] = None,
-        pixel_values_videos: Optional[torch.FloatTensor] = None,
-        image_grid_thw: Optional[torch.LongTensor] = None,
-        video_grid_thw: Optional[torch.LongTensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        feature_attention_mask: Optional[torch.Tensor] = None,
-        audio_feature_lengths: Optional[torch.LongTensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[Cache] = None,
-        inputs_embeds: Optional[torch.FloatTensor] = None,
-        rope_deltas: Optional[torch.LongTensor] = None,
-        labels: Optional[torch.LongTensor] = None,
-        use_cache: Optional[bool] = None,
-        output_attentions: Optional[bool] = None,
-        output_hidden_states: Optional[bool] = None,
-        return_dict: Optional[bool] = None,
-        use_audio_in_video: Optional[bool] = None,
-        cache_position: Optional[torch.LongTensor] = None,
-        video_second_per_grid: Optional[torch.LongTensor] = None,
+        input_ids: torch.LongTensor | None = None,
+        input_features: torch.FloatTensor | None = None,
+        pixel_values: torch.FloatTensor | None = None,
+        pixel_values_videos: torch.FloatTensor | None = None,
+        image_grid_thw: torch.LongTensor | None = None,
+        video_grid_thw: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        feature_attention_mask: torch.Tensor | None = None,
+        audio_feature_lengths: torch.LongTensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        rope_deltas: torch.LongTensor | None = None,
+        labels: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        return_dict: bool | None = None,
+        use_audio_in_video: bool | None = None,
+        cache_position: torch.LongTensor | None = None,
+        video_second_per_grid: torch.LongTensor | None = None,
         **kwargs: Unpack[TransformersKwargs],
-    ) -> Union[tuple, Qwen2_5OmniThinkerCausalLMOutputWithPast]:
+    ) -> tuple | Qwen2_5OmniThinkerCausalLMOutputWithPast:
         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.
@@ -2306,13 +2331,14 @@ class Qwen2_5OmniThinkerForConditionalGeneration(Qwen2_5OmniPreTrainedModelForCo
                 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 = self.get_image_features(pixel_values, image_grid_thw)
+            image_embeds = self.get_image_features(pixel_values, image_grid_thw, return_dict=True).pooler_output
             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
@@ -2320,7 +2346,7 @@ class Qwen2_5OmniThinkerForConditionalGeneration(Qwen2_5OmniPreTrainedModelForCo
             inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
 
         if pixel_values_videos is not None:
-            video_embeds = self.get_video_features(pixel_values_videos, video_grid_thw)
+            video_embeds = self.get_video_features(pixel_values_videos, video_grid_thw, return_dict=True).pooler_output
             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
@@ -2469,13 +2495,13 @@ class Qwen2_5OmniTalkerCausalLMOutputWithPast(ModelOutput):
         response that the talker model will use to generate speech tokens.
     """
 
-    loss: Optional[torch.FloatTensor] = None
-    logits: Optional[torch.FloatTensor] = None
-    past_key_values: Optional[Cache] = None
-    hidden_states: Optional[tuple[torch.FloatTensor]] = None
-    attentions: Optional[tuple[torch.FloatTensor]] = None
-    rope_deltas: Optional[torch.LongTensor] = None
-    thinker_reply_part: Optional[torch.FloatTensor] = None
+    loss: torch.FloatTensor | None = None
+    logits: torch.FloatTensor | None = None
+    past_key_values: Cache | None = None
+    hidden_states: tuple[torch.FloatTensor] | None = None
+    attentions: tuple[torch.FloatTensor] | None = None
+    rope_deltas: torch.LongTensor | None = None
+    thinker_reply_part: torch.FloatTensor | None = None
 
 
 class Qwen2_5OmniTalkerModel(Qwen2_5_VLTextModel):
@@ -2526,26 +2552,26 @@ class Qwen2_5OmniTalkerForConditionalGeneration(Qwen2_5OmniPreTrainedModelForCon
     @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,
-        thinker_reply_part: Optional[torch.FloatTensor] = None,
-        inputs_embeds: Optional[torch.FloatTensor] = None,
-        rope_deltas: Optional[torch.LongTensor] = None,
-        use_cache: Optional[bool] = None,
-        cache_position: Optional[torch.LongTensor] = None,
-        input_text_ids: Optional[torch.LongTensor] = None,
-        image_grid_thw: Optional[torch.LongTensor] = None,
-        video_grid_thw: Optional[torch.LongTensor] = None,
-        use_audio_in_video: Optional[bool] = None,
-        audio_feature_lengths: Optional[torch.LongTensor] = None,
-        video_second_per_grid: Optional[torch.LongTensor] = None,
-        output_attentions: Optional[bool] = None,
-        output_hidden_states: Optional[bool] = None,
-        return_dict: Optional[bool] = 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,
+        thinker_reply_part: torch.FloatTensor | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        rope_deltas: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        cache_position: torch.LongTensor | None = None,
+        input_text_ids: torch.LongTensor | None = None,
+        image_grid_thw: torch.LongTensor | None = None,
+        video_grid_thw: torch.LongTensor | None = None,
+        use_audio_in_video: bool | None = None,
+        audio_feature_lengths: torch.LongTensor | None = None,
+        video_second_per_grid: torch.LongTensor | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        return_dict: bool | None = None,
         **kwargs,
-    ) -> Union[tuple, Qwen2_5OmniTalkerCausalLMOutputWithPast]:
+    ) -> tuple | Qwen2_5OmniTalkerCausalLMOutputWithPast:
         r"""
         thinker_reply_part (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
             Hidden states from the thinker model's output that represent the text reply part to be processed.
@@ -2606,12 +2632,13 @@ class Qwen2_5OmniTalkerForConditionalGeneration(Qwen2_5OmniPreTrainedModelForCon
                     video_second_per_grid,
                 )
 
-                inputs_embeds[:, -1, :] += self.get_input_embeddings()(
-                    torch.tensor([self.codec_bos_token], dtype=torch.long, device=inputs_embeds.device)
-                )
-                inputs_embeds[:, -2, :] += self.get_input_embeddings()(
-                    torch.tensor([self.codec_pad_token], dtype=torch.long, device=inputs_embeds.device)
-                )
+                if inputs_embeds is not None:
+                    inputs_embeds[:, -1, :] += self.get_input_embeddings()(
+                        torch.tensor([self.codec_bos_token], dtype=torch.long, device=inputs_embeds.device)
+                    )
+                    inputs_embeds[:, -2, :] += self.get_input_embeddings()(
+                        torch.tensor([self.codec_pad_token], dtype=torch.long, device=inputs_embeds.device)
+                    )
                 self.rope_deltas = rope_deltas
 
             else:
@@ -2748,9 +2775,9 @@ class Qwen2_5OmniDiTRotaryEmbedding(LlamaRotaryEmbedding):
 
     @staticmethod
     def compute_default_rope_parameters(
-        config: Optional[Qwen2_5OmniDiTConfig] = None,
+        config: Qwen2_5OmniDiTConfig | None = None,
         device: Optional["torch.device"] = None,
-        seq_len: Optional[int] = None,
+        seq_len: int | None = None,
     ) -> tuple["torch.Tensor", float]:
         return super().compute_default_rope_parameters(
             config,
@@ -2760,7 +2787,7 @@ class Qwen2_5OmniDiTRotaryEmbedding(LlamaRotaryEmbedding):
 
 
 # Modified from Llama with a different rotate function, will fixed in next release
-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:
@@ -2768,8 +2795,6 @@ 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
@@ -3112,9 +3137,9 @@ class DiTInputEmbedding(nn.Module):
         speaker_embedding: torch.Tensor,
         condition_vector: torch.Tensor,
         code_embed: torch.Tensor,
-        drop_audio_cond: Optional[bool] = False,
-        code_embed_uncond: Optional[bool] = None,
-        apply_cfg: Optional[bool] = True,
+        drop_audio_cond: bool | None = False,
+        code_embed_uncond: bool | None = None,
+        apply_cfg: bool | None = True,
     ):
         if apply_cfg:
             hidden_states = torch.cat([hidden_states, hidden_states], dim=0)
@@ -3836,15 +3861,24 @@ class Qwen2_5OmniToken2WavDiTModel(Qwen2_5OmniPreTrainedModel):
         guidance_scale=0.5,
         sway_coefficient=-1.0,
     ):
-        noise_initialization = torch.randn([1, 30000, self.mel_dim], dtype=reference_mel_spectrogram.dtype)
         maximum_duration = quantized_code.shape[1] * self.repeats
-        initial_state = noise_initialization[:, :maximum_duration].to(quantized_code.device)
         batch_size = reference_mel_spectrogram.shape[0]
-        conditioning_vector = conditioning_vector.unsqueeze(1).repeat(1, maximum_duration, 1)
-
         if batch_size != 1:
             raise ValueError("Only batch size = 1 is currently supported")
 
+        if maximum_duration > self.config.max_position_embeddings:
+            raise ValueError(
+                f"Requested mel length ({maximum_duration}) exceeds `dit_config.max_position_embeddings` "
+                f"({self.config.max_position_embeddings}). Provide shorter `quantized_code`."
+            )
+
+        initial_state = torch.randn(
+            [batch_size, maximum_duration, self.mel_dim],
+            dtype=reference_mel_spectrogram.dtype,
+            device=quantized_code.device,
+        )
+        conditioning_vector = conditioning_vector.unsqueeze(1).repeat(1, maximum_duration, 1)
+
         def ode_function(time_step, hidden_states):
             if guidance_scale < 1e-5:
                 prediction = self(
@@ -3855,6 +3889,7 @@ class Qwen2_5OmniToken2WavDiTModel(Qwen2_5OmniPreTrainedModel):
                     time_step=time_step,
                     drop_audio_conditioning=False,
                     drop_code=False,
+                    apply_cfg=False,
                 )
                 return prediction
 
@@ -3899,9 +3934,9 @@ class Qwen2_5OmniToken2WavModel(Qwen2_5OmniPreTrainedModel):
     def __init__(self, config: Qwen2_5OmniToken2WavConfig):
         super().__init__(config)
         attn_impl = config._attn_implementation
-        if config._attn_implementation == "flash_attention_2":
+        if is_flash_attention_requested(config):
             logger.warning_once(
-                "Qwen2_5OmniToken2WavModel must inference with fp32, but flash_attention_2 only supports fp16 and bf16, "
+                "Qwen2_5OmniToken2WavModel must inference with fp32, but Flash Attention only supports fp16 and bf16, "
                 "attention implementation of Qwen2_5OmniToken2WavModel will fallback to sdpa."
             )
             attn_impl = "sdpa"
@@ -4051,7 +4086,7 @@ class Qwen2_5OmniForConditionalGeneration(Qwen2_5OmniPreTrainedModel, Generation
     # TODO: raushan, defaults should be saved in generation config
     def generate(
         self,
-        input_ids: Optional[torch.Tensor] = None,
+        input_ids: torch.Tensor | None = None,
         speaker: str = "Chelsie",
         use_audio_in_video: bool = False,
         thinker_max_new_tokens: int = 1024,
@@ -4087,7 +4122,7 @@ class Qwen2_5OmniForConditionalGeneration(Qwen2_5OmniPreTrainedModel, Generation
                 - **Text** (`torch.Tensor`): Generated text token sequence.
                 - **Audio waveform** (`torch.Tensor`): Generated audio waveform.
         """
-        # check `False` on purpose because the paramter can be `str/bool`. This is needed for BC
+        # check `False` on purpose because the parameter can be `str/bool`. This is needed for BC
         generation_mode = kwargs.pop("generation_mode", None)
         return_audio = generation_mode != "text" and generation_mode is not False