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

transformers/models/qwen3_vl/modeling_qwen3_vl.py

@@ -4,6 +4,7 @@
 #             the file from the modular. If any change should be done, please apply the change to the
 #                          modular_qwen3_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.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -20,40 +21,28 @@
 
 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 ...integrations import use_kernel_forward_from_hub, use_kernel_func_from_hub, use_kernelized_func
+from ...integrations import use_kernel_forward_from_hub, use_kernel_func_from_hub
 from ...masking_utils import create_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, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, torch_compilable_check
-from ...utils.generic import check_model_inputs, is_flash_attention_requested, maybe_autocast
+from ...utils import TransformersKwargs, auto_docstring, is_torchdynamo_compiling
+from ...utils.generic import check_model_inputs
 from .configuration_qwen3_vl import Qwen3VLConfig, Qwen3VLTextConfig, Qwen3VLVisionConfig
 
 
-@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 Qwen3VLVisionMLP(nn.Module):
     def __init__(self, config):
         super().__init__()
@@ -92,8 +81,6 @@ class Qwen3VLVisionRotaryEmbedding(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)
 
@@ -157,7 +144,7 @@ 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: Unpack[TransformersKwargs],
@@ -196,8 +183,8 @@ class Qwen3VLVisionAttention(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]
@@ -215,8 +202,8 @@ class Qwen3VLVisionAttention(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,
@@ -273,8 +260,8 @@ class Qwen3VLVisionBlock(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(
@@ -305,15 +292,15 @@ class Qwen3VLTextRotaryEmbedding(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
 
         self.mrope_section = config.rope_parameters.get("mrope_section", [24, 20, 20])
 
     @staticmethod
     def compute_default_rope_parameters(
-        config: Qwen3VLTextConfig | None = None,
+        config: Optional[Qwen3VLTextConfig] = 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
@@ -350,7 +337,7 @@ class Qwen3VLTextRotaryEmbedding(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)
             freqs = self.apply_interleaved_mrope(freqs, self.mrope_section)
             emb = torch.cat((freqs, freqs), dim=-1)
@@ -399,7 +386,7 @@ class Qwen3VLTextRMSNorm(nn.Module):
 
 
 @use_kernel_func_from_hub("rotary_pos_emb")
-def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
     """Applies Rotary Position Embedding to the query and key tensors.
 
     Args:
@@ -407,6 +394,8 @@ def apply_rotary_pos_emb(q, k, cos, sin, 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
@@ -424,7 +413,6 @@ def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
     return q_embed, k_embed
 
 
-@use_kernelized_func(apply_rotary_pos_emb)
 class Qwen3VLTextAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
@@ -451,6 +439,7 @@ class Qwen3VLTextAttention(nn.Module):
         self.o_proj = nn.Linear(
             config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
         )
+        self.rotary_fn = apply_rotary_pos_emb
         self.q_norm = Qwen3VLTextRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # unlike olmo, only on the head dim!
         self.k_norm = Qwen3VLTextRMSNorm(
             self.head_dim, eps=config.rms_norm_eps
@@ -460,11 +449,11 @@ class Qwen3VLTextAttention(nn.Module):
         self,
         hidden_states: torch.Tensor,
         position_embeddings: tuple[torch.Tensor, torch.Tensor],
-        attention_mask: torch.Tensor | None,
-        past_key_values: Cache | None = None,
-        cache_position: torch.LongTensor | None = None,
+        attention_mask: Optional[torch.Tensor],
+        past_key_values: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
         **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
         input_shape = hidden_states.shape[:-1]
         hidden_shape = (*input_shape, -1, self.head_dim)
 
@@ -531,11 +520,11 @@ class Qwen3VLTextDecoderLayer(GradientCheckpointingLayer):
         self,
         hidden_states: torch.Tensor,
         position_embeddings: tuple[torch.Tensor, torch.Tensor],
-        attention_mask: torch.Tensor | None = None,
-        position_ids: torch.LongTensor | None = None,
-        past_key_values: Cache | None = None,
-        use_cache: bool | None = False,
-        cache_position: torch.LongTensor | None = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
         **kwargs: Unpack[TransformersKwargs],
     ) -> torch.Tensor:
         residual = hidden_states
@@ -578,11 +567,11 @@ class Qwen3VLModelOutputWithPast(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
 
 
 @auto_docstring
@@ -603,20 +592,10 @@ class Qwen3VLPreTrainedModel(PreTrainedModel):
         "attentions": Qwen3VLTextAttention,
     }
 
-    def _init_weights(self, module):
-        super()._init_weights(module)
-        if isinstance(module, Qwen3VLVisionRotaryEmbedding):
-            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 Qwen3VLVisionModel(Qwen3VLPreTrainedModel):
     config: Qwen3VLVisionConfig
     _no_split_modules = ["Qwen3VLVisionBlock"]
-    _can_record_outputs = {
-        "hidden_states": Qwen3VLVisionBlock,
-        "attentions": Qwen3VLVisionAttention,
-    }
 
     def __init__(self, config, *inputs, **kwargs) -> None:
         super().__init__(config, *inputs, **kwargs)
@@ -653,8 +632,6 @@ class Qwen3VLVisionModel(Qwen3VLPreTrainedModel):
 
         self.gradient_checkpointing = False
 
-        self.post_init()
-
     def rot_pos_emb(self, grid_thw: torch.Tensor) -> torch.Tensor:
         merge_size = self.spatial_merge_size
 
@@ -754,10 +731,7 @@ class Qwen3VLVisionModel(Qwen3VLPreTrainedModel):
         patch_pos_embeds = torch.cat(patch_pos_embeds_permute)
         return patch_pos_embeds
 
-    @check_model_inputs
-    def forward(
-        self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs: Unpack[TransformersKwargs]
-    ) -> tuple | BaseModelOutputWithDeepstackFeatures:
+    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)`):
@@ -805,13 +779,9 @@ class Qwen3VLVisionModel(Qwen3VLPreTrainedModel):
                 )
                 deepstack_feature_lists.append(deepstack_feature)
 
-        merged_hidden_states = self.merger(hidden_states)
+        hidden_states = self.merger(hidden_states)
 
-        return BaseModelOutputWithDeepstackFeatures(
-            last_hidden_state=hidden_states,
-            pooler_output=merged_hidden_states,
-            deepstack_features=deepstack_feature_lists,
-        )
+        return hidden_states, deepstack_feature_lists
 
 
 @auto_docstring(
@@ -844,18 +814,18 @@ class Qwen3VLTextModel(Qwen3VLPreTrainedModel):
     @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,
-        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,
+        cache_position: Optional[torch.LongTensor] = None,
         # args for deepstack
-        visual_pos_masks: torch.Tensor | None = None,
-        deepstack_visual_embeds: list[torch.Tensor] | None = None,
+        visual_pos_masks: Optional[torch.Tensor] = None,
+        deepstack_visual_embeds: Optional[list[torch.Tensor]] = None,
         **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple | BaseModelOutputWithPast:
+    ) -> Union[tuple, BaseModelOutputWithPast]:
         r"""
         visual_pos_masks (`torch.Tensor` of shape `(batch_size, seqlen)`, *optional*):
             The mask of the visual positions.
@@ -971,10 +941,10 @@ class Qwen3VLModel(Qwen3VLPreTrainedModel):
 
     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,
-        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,
+        attention_mask: Optional[torch.Tensor] = None,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         """Different from the original implementation, Qwen3VL use timestamps rather than absolute time position ids."""
 
@@ -1088,54 +1058,43 @@ class Qwen3VLModel(Qwen3VLPreTrainedModel):
 
             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 | 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.
+        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. The deepstack visual features are also returned.
+
+        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.
         """
         # Same implementation as for images
-        return self.get_image_features(pixel_values_videos, video_grid_thw, **kwargs)
+        return self.get_image_features(pixel_values_videos, video_grid_thw)
 
-    @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.
+    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. The deepstack visual features are also returned.
+
+        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_output: BaseModelOutputWithDeepstackFeatures = self.visual(
-            pixel_values, grid_thw=image_grid_thw, return_dict=True, **kwargs
-        )
-        image_embeds = vision_output.pooler_output
+        image_embeds, deepstack_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(image_embeds, split_sizes)
-        vision_output.pooler_output = image_embeds
-
-        return vision_output
+        return image_embeds, deepstack_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
@@ -1156,19 +1115,18 @@ class Qwen3VLModel(Qwen3VLPreTrainedModel):
 
         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
@@ -1176,17 +1134,17 @@ class Qwen3VLModel(Qwen3VLPreTrainedModel):
     def forward(
         self,
         input_ids: torch.LongTensor = 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,
-        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,
-        cache_position: torch.LongTensor | None = 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,
+        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,
+        cache_position: Optional[torch.LongTensor] = None,
         **kwargs: Unpack[TransformersKwargs],
-    ) -> tuple | Qwen3VLModelOutputWithPast:
+    ) -> Union[tuple, Qwen3VLModelOutputWithPast]:
         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.
@@ -1203,11 +1161,7 @@ class Qwen3VLModel(Qwen3VLPreTrainedModel):
         video_mask = None
 
         if pixel_values is not None:
-            image_outputs: BaseModelOutputWithDeepstackFeatures = self.get_image_features(
-                pixel_values, image_grid_thw, return_dict=True
-            )
-            image_embeds = image_outputs.pooler_output
-            deepstack_image_embeds = image_outputs.deepstack_features
+            image_embeds, deepstack_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
@@ -1215,11 +1169,7 @@ class Qwen3VLModel(Qwen3VLPreTrainedModel):
             inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
 
         if pixel_values_videos is not None:
-            video_outputs: BaseModelOutputWithDeepstackFeatures = self.get_video_features(
-                pixel_values_videos, video_grid_thw, return_dict=True
-            )
-            video_embeds = video_outputs.pooler_output
-            deepstack_video_embeds = video_outputs.deepstack_features
+            video_embeds, deepstack_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
@@ -1251,19 +1201,44 @@ class Qwen3VLModel(Qwen3VLPreTrainedModel):
             deepstack_visual_embeds = deepstack_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:
+            attention_mask_tensor = (
+                attention_mask if not isinstance(attention_mask, dict) else attention_mask["full_attention"]
+            )
+            if attention_mask_tensor is not None and attention_mask_tensor.ndim == 4:
+                attention_mask_tensor = torch.diagonal(attention_mask_tensor[:, 0], dim1=1, dim2=2)
+                # Only apply conversion for floating point tensors (inverted masks)
+                if attention_mask_tensor.dtype.is_floating_point:
+                    attention_mask_tensor = attention_mask_tensor / torch.finfo(attention_mask_tensor.dtype).min
+                    attention_mask_tensor = (1.0 - attention_mask_tensor).int()
+
+            # Calculate RoPE index once per generation in the pre-fill stage only.
+            # 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 asssisted decoding
+            prefill_compiled_stage = is_torchdynamo_compiling() and (
+                (input_ids is not None and input_ids.shape[1] != 1)
+                or (inputs_embeds is not None and inputs_embeds.shape[1] != 1)
+            )
+            prefill_noncompiled_stage = not is_torchdynamo_compiling() and (
+                (cache_position is not None and cache_position[0] == 0)
+                or (past_key_values is None or past_key_values.get_seq_length() == 0)
+            )
+            if (prefill_compiled_stage or prefill_noncompiled_stage) or self.rope_deltas is None:
                 position_ids, rope_deltas = self.get_rope_index(
                     input_ids,
                     image_grid_thw,
                     video_grid_thw,
-                    attention_mask=attention_mask,
+                    attention_mask=attention_mask_tensor,
                 )
                 self.rope_deltas = rope_deltas
             # then use the prev pre-calculated rope-deltas to get the correct position ids
             else:
                 batch_size, seq_length, _ = inputs_embeds.shape
-                delta = (past_key_values_length + self.rope_deltas).to(inputs_embeds.device)
+                delta = (
+                    (cache_position[0] + self.rope_deltas).to(inputs_embeds.device)
+                    if cache_position is not None
+                    else 0
+                )
                 position_ids = torch.arange(seq_length, device=inputs_embeds.device)
                 position_ids = position_ids.view(1, -1).expand(batch_size, -1)
                 if cache_position is not None:  # otherwise `deltas` is an int `0`
@@ -1311,12 +1286,12 @@ class Qwen3VLCausalLMOutputWithPast(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 Qwen3VLForConditionalGeneration(Qwen3VLPreTrainedModel, GenerationMixin):
@@ -1339,55 +1314,31 @@ class Qwen3VLForConditionalGeneration(Qwen3VLPreTrainedModel, GenerationMixin):
     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 | 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.
-        """
-        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 | 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.
-        """
-        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
+    @check_model_inputs
     def forward(
         self,
         input_ids: torch.LongTensor = 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,
-        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,
-        cache_position: torch.LongTensor | None = None,
-        logits_to_keep: int | torch.Tensor = 0,
+        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,
+        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,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
         **kwargs: Unpack[TransformersKwargs],
-    ) -> tuple | Qwen3VLCausalLMOutputWithPast:
+    ) -> Union[tuple, Qwen3VLCausalLMOutputWithPast]:
         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, ...,
@@ -1463,8 +1414,6 @@ class Qwen3VLForConditionalGeneration(Qwen3VLPreTrainedModel, GenerationMixin):
             loss=loss,
             logits=logits,
             past_key_values=outputs.past_key_values,
-            hidden_states=outputs.hidden_states,
-            attentions=outputs.attentions,
             rope_deltas=outputs.rope_deltas,
         )
 
@@ -1481,7 +1430,6 @@ class Qwen3VLForConditionalGeneration(Qwen3VLPreTrainedModel, GenerationMixin):
         pixel_values_videos=None,
         image_grid_thw=None,
         video_grid_thw=None,
-        is_first_iteration=False,
         **kwargs,
     ):
         # Overwritten -- in specific circumstances we don't want to forward image inputs to the model
@@ -1498,39 +1446,13 @@ class Qwen3VLForConditionalGeneration(Qwen3VLPreTrainedModel, GenerationMixin):
             image_grid_thw=image_grid_thw,
             video_grid_thw=video_grid_thw,
             use_cache=use_cache,
-            is_first_iteration=is_first_iteration,
             **kwargs,
         )
 
-        # Qwen3VL position_ids are prepared with rope_deltas
-        if position_ids is None:
-            # Calculate RoPE index once per generation in the pre-fill stage only.
-            # 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 asssisted decoding
-            if model_inputs["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,
-                    video_grid_thw=video_grid_thw,
-                    attention_mask=attention_mask,
-                )
-                self.model.rope_deltas = rope_deltas
-            # then use the prev pre-calculated rope-deltas to get the correct position ids
-            elif "position_ids" in model_inputs:
-                batch_size, seq_length = model_inputs["position_ids"].shape
-                device = model_inputs["position_ids"].device
-                position_ids = torch.arange(seq_length, device=device)
-                position_ids = position_ids.view(1, 1, -1).expand(3, batch_size, -1)
-                delta = cache_position[0] + self.model.rope_deltas
-                delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)
-                vision_positions = position_ids + delta.expand_as(position_ids)
-
-            # Concatenate "text + vision" positions into [4, bs, seq-len]
-            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:
+        # Qwen3VL position_ids are prepareed with rope_deltas in forward
+        model_inputs["position_ids"] = None
+
+        if cache_position[0] != 0:
             model_inputs["pixel_values"] = None
             model_inputs["pixel_values_videos"] = None
 
@@ -1538,8 +1460,8 @@ class Qwen3VLForConditionalGeneration(Qwen3VLPreTrainedModel, GenerationMixin):
 
     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.
@@ -1591,7 +1513,7 @@ class Qwen3VLForConditionalGeneration(Qwen3VLPreTrainedModel, GenerationMixin):
         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 -- Qwen3VL use timestamps and remove second_per_grid_ts