transformers 5.0.0__py3-none-any.whl → 5.0.0rc0__py3-none-any.whl

transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py

@@ -4,6 +4,7 @@
 #             the file from the modular. If any change should be done, please apply the change to the
 #                          modular_qwen2_5_vl.py file directly. One of our CI enforces this.
 #                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
 # Copyright 2025 The Qwen Team and The HuggingFace Inc. team. All rights reserved.
 #
 # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
@@ -22,27 +23,26 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
 from collections.abc import Callable
 from dataclasses import dataclass
-from typing import Any, Optional
+from typing import Any, Optional, Union
 
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
 
-from ... import initialization as init
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
 from ...masking_utils import create_causal_mask, create_sliding_window_causal_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import GradientCheckpointingLayer
-from ...modeling_outputs import BaseModelOutputWithPast, BaseModelOutputWithPooling, ModelOutput
+from ...modeling_outputs import BaseModelOutputWithPast, ModelOutput
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging, torch_compilable_check
-from ...utils.generic import check_model_inputs, is_flash_attention_requested, maybe_autocast
+from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
 from ..qwen2.modeling_qwen2 import Qwen2RMSNorm
 from .configuration_qwen2_5_vl import Qwen2_5_VLConfig, Qwen2_5_VLTextConfig, Qwen2_5_VLVisionConfig
 
@@ -95,8 +95,6 @@ class Qwen2_5_VisionRotaryEmbedding(nn.Module):
 
     def __init__(self, dim: int, theta: float = 10000.0) -> None:
         super().__init__()
-        self.dim = dim
-        self.theta = theta
         inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
         self.register_buffer("inv_freq", inv_freq, persistent=False)
 
@@ -160,7 +158,7 @@ def eager_attention_forward(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attention_mask: torch.Tensor | None,
+    attention_mask: Optional[torch.Tensor],
     scaling: float,
     dropout: float = 0.0,
     **kwargs,
@@ -199,8 +197,8 @@ class Qwen2_5_VLVisionAttention(nn.Module):
         self,
         hidden_states: torch.Tensor,
         cu_seqlens: torch.Tensor,
-        rotary_pos_emb: torch.Tensor | None = None,
-        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
+        rotary_pos_emb: Optional[torch.Tensor] = None,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
         **kwargs,
     ) -> torch.Tensor:
         seq_length = hidden_states.shape[0]
@@ -218,8 +216,8 @@ class Qwen2_5_VLVisionAttention(nn.Module):
         if self.config._attn_implementation != "eager":
             attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
 
-        if is_flash_attention_requested(self.config):
-            # Flash Attention: Use cu_seqlens for variable length attention
+        if self.config._attn_implementation == "flash_attention_2":
+            # Flash Attention 2: Use cu_seqlens for variable length attention
             max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
             attn_output, _ = attention_interface(
                 self,
@@ -276,8 +274,8 @@ class Qwen2_5_VLVisionBlock(GradientCheckpointingLayer):
         self,
         hidden_states: torch.Tensor,
         cu_seqlens: torch.Tensor,
-        rotary_pos_emb: torch.Tensor | None = None,
-        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
+        rotary_pos_emb: Optional[torch.Tensor] = None,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
         **kwargs,
     ) -> torch.Tensor:
         hidden_states = hidden_states + self.attn(
@@ -305,21 +303,11 @@ class Qwen2_5_VLPreTrainedModel(PreTrainedModel):
     _can_compile_fullgraph = True
     _supports_attention_backend = True
 
-    def _init_weights(self, module):
-        super()._init_weights(module)
-        if isinstance(module, Qwen2_5_VisionRotaryEmbedding):
-            inv_freq = 1.0 / (module.theta ** (torch.arange(0, module.dim, 2, dtype=torch.float) / module.dim))
-            init.copy_(module.inv_freq, inv_freq)
-
 
 class Qwen2_5_VisionTransformerPretrainedModel(Qwen2_5_VLPreTrainedModel):
     config: Qwen2_5_VLVisionConfig
     _no_split_modules = ["Qwen2_5_VLVisionBlock"]
     _input_embed_layer = "patch_embed"
-    _can_record_outputs = {
-        "hidden_states": Qwen2_5_VLVisionBlock,
-        "attentions": Qwen2_5_VLVisionAttention,
-    }
 
     def __init__(self, config, *inputs, **kwargs) -> None:
         super().__init__(config, *inputs, **kwargs)
@@ -347,8 +335,6 @@ class Qwen2_5_VisionTransformerPretrainedModel(Qwen2_5_VLPreTrainedModel):
         )
         self.gradient_checkpointing = False
 
-        self.post_init()
-
     def rot_pos_emb(self, grid_thw):
         pos_ids = []
         for t, h, w in grid_thw:
@@ -419,10 +405,7 @@ class Qwen2_5_VisionTransformerPretrainedModel(Qwen2_5_VLPreTrainedModel):
 
         return window_index, cu_window_seqlens
 
-    @check_model_inputs
-    def forward(
-        self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs: Unpack[TransformersKwargs]
-    ) -> tuple | BaseModelOutputWithPooling:
+    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs) -> torch.Tensor:
         """
         Args:
             hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`):
@@ -476,14 +459,11 @@ class Qwen2_5_VisionTransformerPretrainedModel(Qwen2_5_VLPreTrainedModel):
                 **kwargs,
             )
 
-        merged_hidden_states = self.merger(hidden_states)
+        hidden_states = self.merger(hidden_states)
         reverse_indices = torch.argsort(window_index)
-        merged_hidden_states = merged_hidden_states[reverse_indices, :]
+        hidden_states = hidden_states[reverse_indices, :]
 
-        return BaseModelOutputWithPooling(
-            last_hidden_state=hidden_states,
-            pooler_output=merged_hidden_states,
-        )
+        return hidden_states
 
 
 @dataclass
@@ -503,11 +483,11 @@ class Qwen2_5_VLModelOutputWithPast(ModelOutput):
         The rope index difference between sequence length and multimodal rope.
     """
 
-    last_hidden_state: 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
+    last_hidden_state: 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
 
 
 class Qwen2_5_VLRotaryEmbedding(nn.Module):
@@ -527,13 +507,13 @@ class Qwen2_5_VLRotaryEmbedding(nn.Module):
         inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
 
         self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
+        self.original_inv_freq = inv_freq
 
     @staticmethod
     def compute_default_rope_parameters(
-        config: Qwen2_5_VLConfig | None = None,
+        config: Optional[Qwen2_5_VLConfig] = None,
         device: Optional["torch.device"] = None,
-        seq_len: int | None = None,
+        seq_len: Optional[int] = None,
     ) -> tuple["torch.Tensor", float]:
         """
         Computes the inverse frequencies according to the original RoPE implementation
@@ -567,7 +547,7 @@ class Qwen2_5_VLRotaryEmbedding(nn.Module):
         position_ids_expanded = position_ids[:, :, None, :].float()  # shape (3, bs, 1, positions)
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(2, 3)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
@@ -643,7 +623,7 @@ class Qwen2_5_VLAttention(nn.Module):
     and "Generating Long Sequences with Sparse Transformers".
     """
 
-    def __init__(self, config: Qwen2_5_VLTextConfig, layer_idx: int | None = None):
+    def __init__(self, config: Qwen2_5_VLTextConfig, layer_idx: Optional[int] = None):
         super().__init__()
         self.config = config
         self.layer_idx = layer_idx
@@ -679,15 +659,15 @@ class Qwen2_5_VLAttention(nn.Module):
     def forward(
         self,
         hidden_states: torch.Tensor,
-        attention_mask: torch.Tensor | None = None,
-        position_ids: torch.LongTensor | None = None,
-        past_key_values: Cache | None = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
         output_attentions: bool = False,
         use_cache: bool = False,
-        cache_position: torch.LongTensor | None = None,
-        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
         **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
         bsz, q_len, _ = hidden_states.size()
 
         query_states = self.q_proj(hidden_states)
@@ -734,7 +714,7 @@ class Qwen2_5_VLDecoderLayer(GradientCheckpointingLayer):
         super().__init__()
         self.hidden_size = config.hidden_size
 
-        if config.use_sliding_window and not is_flash_attention_requested(config):
+        if config.use_sliding_window and config._attn_implementation != "flash_attention_2":
             logger.warning_once(
                 f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
                 "unexpected results may be encountered."
@@ -749,15 +729,15 @@ class Qwen2_5_VLDecoderLayer(GradientCheckpointingLayer):
     def forward(
         self,
         hidden_states: torch.Tensor,
-        attention_mask: torch.Tensor | None = None,
-        position_ids: torch.LongTensor | None = None,
-        past_key_values: Cache | None = None,
-        output_attentions: bool | None = False,
-        use_cache: bool | None = False,
-        cache_position: torch.LongTensor | None = None,
-        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
         **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple[torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]:
+    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
         """
         Args:
             hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
@@ -838,18 +818,18 @@ class Qwen2_5_VLTextModel(Qwen2_5_VLPreTrainedModel):
     @auto_docstring
     def forward(
         self,
-        input_ids: torch.LongTensor | None = None,
-        attention_mask: torch.Tensor | None = None,
-        position_ids: torch.LongTensor | None = None,
-        past_key_values: Cache | None = None,
-        inputs_embeds: torch.FloatTensor | None = None,
-        use_cache: bool | None = None,
-        output_attentions: bool | None = None,
-        output_hidden_states: bool | None = None,
-        return_dict: bool | None = None,
-        cache_position: torch.LongTensor | None = None,
+        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,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
         **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple | BaseModelOutputWithPast:
+    ) -> Union[tuple, BaseModelOutputWithPast]:
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
@@ -995,11 +975,11 @@ class Qwen2_5_VLModel(Qwen2_5_VLPreTrainedModel):
 
     def get_rope_index(
         self,
-        input_ids: torch.LongTensor | None = None,
-        image_grid_thw: torch.LongTensor | None = None,
-        video_grid_thw: torch.LongTensor | None = None,
-        second_per_grid_ts: torch.Tensor | None = None,
-        attention_mask: torch.Tensor | None = None,
+        input_ids: Optional[torch.LongTensor] = None,
+        image_grid_thw: Optional[torch.LongTensor] = None,
+        video_grid_thw: Optional[torch.LongTensor] = None,
+        second_per_grid_ts: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         """
         Calculate the 3D rope index based on image and video's temporal, height and width in LLM.
@@ -1178,56 +1158,46 @@ class Qwen2_5_VLModel(Qwen2_5_VLPreTrainedModel):
 
             return position_ids, mrope_position_deltas
 
-    @can_return_tuple
-    @auto_docstring
     def get_video_features(
-        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.
+        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.
         """
         pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
-        vision_outputs = self.visual(pixel_values_videos, grid_thw=video_grid_thw, return_dict=True, **kwargs)
+        video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
         split_sizes = (video_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
-        video_embeds = torch.split(vision_outputs.pooler_output, split_sizes)
-        vision_outputs.pooler_output = video_embeds
+        video_embeds = torch.split(video_embeds, split_sizes)
+        return video_embeds
 
-        return vision_outputs
+    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.
 
-    @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.
+        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.
         """
         pixel_values = pixel_values.type(self.visual.dtype)
-        vision_outputs = self.visual(pixel_values, grid_thw=image_grid_thw, return_dict=True, **kwargs)
+        image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
         split_sizes = (image_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
-        image_embeds = torch.split(vision_outputs.pooler_output, split_sizes)
-        vision_outputs.pooler_output = image_embeds
-
-        return vision_outputs
+        image_embeds = torch.split(image_embeds, split_sizes)
+        return image_embeds
 
     def get_placeholder_mask(
         self,
         input_ids: torch.LongTensor,
         inputs_embeds: torch.FloatTensor,
-        image_features: torch.FloatTensor | None = None,
-        video_features: torch.FloatTensor | None = None,
+        image_features: Optional[torch.FloatTensor] = None,
+        video_features: Optional[torch.FloatTensor] = None,
     ):
         """
         Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is
@@ -1248,42 +1218,41 @@ class Qwen2_5_VLModel(Qwen2_5_VLPreTrainedModel):
 
         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:
-            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]}",
+        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]}"
             )
 
         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:
-            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]}",
+        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]}"
             )
+
         return special_image_mask, special_video_mask
 
     @auto_docstring
     def forward(
         self,
-        input_ids: torch.LongTensor | None = None,
-        attention_mask: torch.Tensor | None = None,
-        position_ids: torch.LongTensor | None = None,
-        past_key_values: Cache | None = None,
-        inputs_embeds: torch.FloatTensor | None = None,
-        use_cache: bool | None = None,
-        output_attentions: bool | None = None,
-        output_hidden_states: bool | None = None,
-        return_dict: bool | None = None,
-        pixel_values: torch.Tensor | None = None,
-        pixel_values_videos: torch.FloatTensor | None = None,
-        image_grid_thw: torch.LongTensor | None = None,
-        video_grid_thw: torch.LongTensor | None = None,
-        rope_deltas: torch.LongTensor | None = None,
-        cache_position: torch.LongTensor | None = None,
-        second_per_grid_ts: torch.Tensor | None = None,
+        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,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        pixel_values: Optional[torch.Tensor] = None,
+        pixel_values_videos: Optional[torch.FloatTensor] = None,
+        image_grid_thw: Optional[torch.LongTensor] = None,
+        video_grid_thw: Optional[torch.LongTensor] = None,
+        rope_deltas: Optional[torch.LongTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        second_per_grid_ts: Optional[torch.Tensor] = None,
         **kwargs: Unpack[TransformersKwargs],
-    ) -> tuple | Qwen2_5_VLModelOutputWithPast:
+    ) -> Union[tuple, Qwen2_5_VLModelOutputWithPast]:
         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.
@@ -1305,7 +1274,7 @@ class Qwen2_5_VLModel(Qwen2_5_VLPreTrainedModel):
             inputs_embeds = self.get_input_embeddings()(input_ids)
 
         if pixel_values is not None:
-            image_embeds = self.get_image_features(pixel_values, image_grid_thw, return_dict=True).pooler_output
+            image_embeds = self.get_image_features(pixel_values, image_grid_thw)
             image_embeds = torch.cat(image_embeds, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
             image_mask, _ = self.get_placeholder_mask(
                 input_ids, inputs_embeds=inputs_embeds, image_features=image_embeds
@@ -1313,7 +1282,7 @@ class Qwen2_5_VLModel(Qwen2_5_VLPreTrainedModel):
             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, return_dict=True).pooler_output
+            video_embeds = self.get_video_features(pixel_values_videos, video_grid_thw)
             video_embeds = torch.cat(video_embeds, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
             _, video_mask = self.get_placeholder_mask(
                 input_ids, inputs_embeds=inputs_embeds, video_features=video_embeds
@@ -1321,8 +1290,7 @@ class Qwen2_5_VLModel(Qwen2_5_VLPreTrainedModel):
             inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)
 
         if position_ids is None:
-            past_key_values_length = 0 if past_key_values is None else past_key_values.get_seq_length()
-            if self.rope_deltas is None or past_key_values_length == 0:
+            if self.rope_deltas is None or cache_position is None or cache_position[0] == 0:
                 position_ids, rope_deltas = self.get_rope_index(
                     input_ids,
                     image_grid_thw,
@@ -1335,7 +1303,10 @@ class Qwen2_5_VLModel(Qwen2_5_VLPreTrainedModel):
                 batch_size, seq_length, _ = inputs_embeds.shape
                 position_ids = torch.arange(seq_length, device=inputs_embeds.device)
                 position_ids = position_ids.view(1, 1, -1).expand(3, batch_size, -1)
-                delta = (past_key_values_length + self.rope_deltas).to(inputs_embeds.device)
+                if cache_position is not None:
+                    delta = (cache_position[0] + self.rope_deltas).to(inputs_embeds.device)
+                else:
+                    delta = torch.zeros((batch_size, seq_length), device=inputs_embeds.device)
                 delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=1)
                 position_ids = position_ids + delta.to(position_ids.device)
 
@@ -1384,12 +1355,12 @@ class Qwen2_5_VLCausalLMOutputWithPast(ModelOutput):
         The rope index difference between sequence length and multimodal rope.
     """
 
-    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
+    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
 
 
 class Qwen2_5_VLForConditionalGeneration(Qwen2_5_VLPreTrainedModel, GenerationMixin):
@@ -1414,61 +1385,37 @@ class Qwen2_5_VLForConditionalGeneration(Qwen2_5_VLPreTrainedModel, GenerationMi
     def set_input_embeddings(self, value):
         self.model.set_input_embeddings(value)
 
-    @auto_docstring
     def get_video_features(
-        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.
-        """
-        return self.model.get_video_features(
-            pixel_values_videos=pixel_values_videos, video_grid_thw=video_grid_thw, **kwargs
-        )
+        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: Optional[torch.LongTensor] = None
+    ):
+        return self.model.get_video_features(pixel_values_videos, video_grid_thw)
 
-    @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.
-        """
-        return self.model.get_image_features(pixel_values=pixel_values, image_grid_thw=image_grid_thw, **kwargs)
+    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: Optional[torch.LongTensor] = None):
+        return self.model.get_image_features(pixel_values, image_grid_thw)
 
     @can_return_tuple
     @auto_docstring
     def forward(
         self,
-        input_ids: torch.LongTensor | None = None,
-        attention_mask: torch.Tensor | None = None,
-        position_ids: torch.LongTensor | None = None,
-        past_key_values: Cache | None = None,
-        inputs_embeds: torch.FloatTensor | None = None,
-        labels: torch.LongTensor | None = None,
-        use_cache: bool | None = None,
-        output_attentions: bool | None = None,
-        output_hidden_states: bool | None = None,
-        pixel_values: torch.Tensor | None = None,
-        pixel_values_videos: torch.FloatTensor | None = None,
-        image_grid_thw: torch.LongTensor | None = None,
-        video_grid_thw: torch.LongTensor | None = None,
-        rope_deltas: torch.LongTensor | None = None,
-        cache_position: torch.LongTensor | None = None,
-        second_per_grid_ts: torch.Tensor | None = None,
-        logits_to_keep: int | torch.Tensor = 0,
+        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,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        pixel_values: Optional[torch.Tensor] = None,
+        pixel_values_videos: Optional[torch.FloatTensor] = None,
+        image_grid_thw: Optional[torch.LongTensor] = None,
+        video_grid_thw: Optional[torch.LongTensor] = None,
+        rope_deltas: Optional[torch.LongTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        second_per_grid_ts: Optional[torch.Tensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
         **kwargs: Unpack[TransformersKwargs],
-    ) -> tuple | Qwen2_5_VLCausalLMOutputWithPast:
+    ) -> Union[tuple, Qwen2_5_VLCausalLMOutputWithPast]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
             Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
@@ -1579,7 +1526,6 @@ class Qwen2_5_VLForConditionalGeneration(Qwen2_5_VLPreTrainedModel, GenerationMi
         image_grid_thw=None,
         video_grid_thw=None,
         second_per_grid_ts=None,
-        is_first_iteration=False,
         **kwargs,
     ):
         # Overwritten -- in specific circumstances we don't want to forward image inputs to the model
@@ -1597,7 +1543,6 @@ class Qwen2_5_VLForConditionalGeneration(Qwen2_5_VLPreTrainedModel, GenerationMi
             video_grid_thw=video_grid_thw,
             second_per_grid_ts=second_per_grid_ts,
             use_cache=use_cache,
-            is_first_iteration=is_first_iteration,
             **kwargs,
         )
 
@@ -1607,7 +1552,7 @@ class Qwen2_5_VLForConditionalGeneration(Qwen2_5_VLPreTrainedModel, GenerationMi
             # When compiling, we can't check tensor values thus we check only input length
             # It is safe to assume that `length!=1` means we're in pre-fill because compiled
             # models currently cannot do assisted decoding
-            if (cache_position[0] == 0 or not use_cache) or self.model.rope_deltas is None:
+            if cache_position[0] == 0 or self.model.rope_deltas is None:
                 vision_positions, rope_deltas = self.model.get_rope_index(
                     model_inputs.get("input_ids", None),
                     image_grid_thw=image_grid_thw,
@@ -1630,7 +1575,7 @@ class Qwen2_5_VLForConditionalGeneration(Qwen2_5_VLPreTrainedModel, GenerationMi
             text_positions = model_inputs["position_ids"][None, ...]
             model_inputs["position_ids"] = torch.cat([text_positions, vision_positions], dim=0)
 
-        if not is_first_iteration and use_cache:
+        if cache_position[0] != 0:
             model_inputs["pixel_values"] = None
             model_inputs["pixel_values_videos"] = None
 
@@ -1638,8 +1583,8 @@ class Qwen2_5_VLForConditionalGeneration(Qwen2_5_VLPreTrainedModel, GenerationMi
 
     def _get_image_nums_and_video_nums(
         self,
-        input_ids: torch.LongTensor | None,
-        inputs_embeds: torch.Tensor | None = None,
+        input_ids: Optional[torch.LongTensor],
+        inputs_embeds: Optional[torch.Tensor] = None,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         """
         Get the number of images and videos for each sample to calculate the separation length of the sample tensor.
@@ -1691,7 +1636,7 @@ class Qwen2_5_VLForConditionalGeneration(Qwen2_5_VLPreTrainedModel, GenerationMi
         self,
         expand_size: int = 1,
         is_encoder_decoder: bool = False,
-        input_ids: torch.LongTensor | None = None,
+        input_ids: Optional[torch.LongTensor] = None,
         **model_kwargs,
     ) -> tuple[torch.LongTensor, dict[str, Any]]:
         # Overwritten -- Support for expanding tensors without a batch size dimension