transformers 5.0.0rc0__py3-none-any.whl → 5.0.0rc2__py3-none-any.whl

transformers/models/qwen3_vl/modeling_qwen3_vl.py

@@ -27,10 +27,11 @@ 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
+from ...integrations import use_kernel_forward_from_hub, use_kernel_func_from_hub, use_kernelized_func
 from ...masking_utils import create_causal_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import GradientCheckpointingLayer
@@ -38,8 +39,8 @@ 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, is_torchdynamo_compiling
-from ...utils.generic import check_model_inputs
+from ...utils import TransformersKwargs, auto_docstring, can_return_tuple
+from ...utils.generic import check_model_inputs, maybe_autocast
 from .configuration_qwen3_vl import Qwen3VLConfig, Qwen3VLTextConfig, Qwen3VLVisionConfig
 
 
@@ -81,6 +82,8 @@ 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)
 
@@ -202,8 +205,8 @@ class Qwen3VLVisionAttention(nn.Module):
         if self.config._attn_implementation != "eager":
             attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
 
-        if self.config._attn_implementation == "flash_attention_2":
-            # Flash Attention 2: Use cu_seqlens for variable length attention
+        if "flash" in self.config._attn_implementation:
+            # Flash Attention: Use cu_seqlens for variable length attention
             max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
             attn_output, _ = attention_interface(
                 self,
@@ -292,7 +295,7 @@ 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.original_inv_freq = inv_freq
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
 
         self.mrope_section = config.rope_parameters.get("mrope_section", [24, 20, 20])
 
@@ -337,7 +340,7 @@ 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 torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        with maybe_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)
@@ -413,6 +416,7 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, 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"""
 
@@ -439,7 +443,6 @@ 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
@@ -592,6 +595,12 @@ 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
@@ -632,6 +641,8 @@ 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
 
@@ -1201,44 +1212,19 @@ class Qwen3VLModel(Qwen3VLPreTrainedModel):
             deepstack_visual_embeds = deepstack_video_embeds
 
         if position_ids is None:
-            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:
+            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:
                 position_ids, rope_deltas = self.get_rope_index(
                     input_ids,
                     image_grid_thw,
                     video_grid_thw,
-                    attention_mask=attention_mask_tensor,
+                    attention_mask=attention_mask,
                 )
                 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 = (
-                    (cache_position[0] + self.rope_deltas).to(inputs_embeds.device)
-                    if cache_position is not None
-                    else 0
-                )
+                delta = (past_key_values_length + self.rope_deltas).to(inputs_embeds.device)
                 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`
@@ -1322,7 +1308,7 @@ class Qwen3VLForConditionalGeneration(Qwen3VLPreTrainedModel, GenerationMixin):
     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)
 
-    @check_model_inputs
+    @can_return_tuple
     def forward(
         self,
         input_ids: torch.LongTensor = None,
@@ -1414,6 +1400,8 @@ 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,
         )
 
@@ -1430,6 +1418,7 @@ 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
@@ -1446,13 +1435,39 @@ 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 prepareed with rope_deltas in forward
-        model_inputs["position_ids"] = None
-
-        if cache_position[0] != 0:
+        # 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:
             model_inputs["pixel_values"] = None
             model_inputs["pixel_values_videos"] = None
 

transformers/models/qwen3_vl/modular_qwen3_vl.py

@@ -22,6 +22,7 @@ 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 ...configuration_utils import PreTrainedConfig
@@ -31,11 +32,11 @@ from ...masking_utils import create_causal_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_outputs import BaseModelOutputWithPast
 from ...modeling_rope_utils import RopeParameters, dynamic_rope_update
-from ...modeling_utils import ALL_ATTENTION_FUNCTIONS
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import ProcessingKwargs, Unpack
 from ...tokenization_utils_base import PreTokenizedInput, TextInput
-from ...utils import auto_docstring, is_torchdynamo_compiling, logging
-from ...utils.generic import check_model_inputs
+from ...utils import auto_docstring, can_return_tuple, logging
+from ...utils.generic import check_model_inputs, maybe_autocast
 from ...video_utils import VideoInput
 from ..llama.modeling_llama import LlamaRotaryEmbedding
 from ..qwen2_5_vl.modeling_qwen2_5_vl import (
@@ -151,7 +152,7 @@ class Qwen3VLTextConfig(PreTrainedConfig):
             Dictionary containing the configuration parameters for the RoPE embeddings. The dictionary should contain
             a value for `rope_theta` and optionally parameters used for scaling in case you want to use RoPE
             with longer `max_position_embeddings`.
-        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
+        attention_bias (`bool`, *optional*, defaults to `False`):
             Whether to use a bias in the query, key, value and output projection layers during self-attention.
         attention_dropout (`float`, *optional*, defaults to 0.0):
             The dropout ratio for the attention probabilities.
@@ -238,13 +239,13 @@ class Qwen3VLConfig(PreTrainedConfig):
         vision_config (`Union[PreTrainedConfig, dict]`,  *optional*, defaults to `Qwen3VLVisionConfig`):
             The config object or dictionary of the vision backbone.
         image_token_id (`int`, *optional*, defaults to 151655):
-            The image token index to encode the image prompt.
+            The token id used as the placeholder for image inputs.
         video_token_id (`int`, *optional*, defaults to 151656):
-            The video token index to encode the image prompt.
+            The token id used as the placeholder for video inputs.
         vision_start_token_id (`int`, *optional*, defaults to 151652):
-            The start token index to encode the image prompt.
+            The token id that marks the start of a vision segment (image or video).
         vision_end_token_id (`int`, *optional*, defaults to 151653):
-            The end token index to encode the image prompt.
+            The token id that marks the end of a vision segment (image or video).
         tie_word_embeddings (`bool`, *optional*, defaults to `False`):
             Whether to tie the word embeddings.
 
@@ -389,7 +390,7 @@ class Qwen3VLTextRotaryEmbedding(LlamaRotaryEmbedding):
         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 torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        with maybe_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)
@@ -488,6 +489,12 @@ class Qwen3VLPreTrainedModel(Qwen2VLPreTrainedModel):
         "attentions": Qwen3VLTextAttention,
     }
 
+    def _init_weights(self, module):
+        PreTrainedModel._init_weights(self, 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
@@ -528,6 +535,8 @@ 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
 
@@ -1033,44 +1042,19 @@ class Qwen3VLModel(Qwen2_5_VLModel):
             deepstack_visual_embeds = deepstack_video_embeds
 
         if position_ids is None:
-            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:
+            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:
                 position_ids, rope_deltas = self.get_rope_index(
                     input_ids,
                     image_grid_thw,
                     video_grid_thw,
-                    attention_mask=attention_mask_tensor,
+                    attention_mask=attention_mask,
                 )
                 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 = (
-                    (cache_position[0] + self.rope_deltas).to(inputs_embeds.device)
-                    if cache_position is not None
-                    else 0
-                )
+                delta = (past_key_values_length + self.rope_deltas).to(inputs_embeds.device)
                 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`
@@ -1105,7 +1089,7 @@ class Qwen3VLForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
     config: Qwen3VLConfig
     _checkpoint_conversion_mapping = {}
 
-    @check_model_inputs
+    @can_return_tuple
     def forward(
         self,
         input_ids: torch.LongTensor = None,
@@ -1197,6 +1181,8 @@ class Qwen3VLForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
             loss=loss,
             logits=logits,
             past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
             rope_deltas=outputs.rope_deltas,
         )
 
@@ -1213,6 +1199,7 @@ class Qwen3VLForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
         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
@@ -1229,13 +1216,39 @@ class Qwen3VLForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
             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 prepareed with rope_deltas in forward
-        model_inputs["position_ids"] = None
-
-        if cache_position[0] != 0:
+        # 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:
             model_inputs["pixel_values"] = None
             model_inputs["pixel_values_videos"] = None
 
@@ -1391,9 +1404,9 @@ class Qwen3VLProcessor(Qwen2VLProcessor):
         **kwargs: Unpack[Qwen3VLProcessorKwargs],
     ) -> BatchFeature:
         """
-        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        Main method to prepare for the model one or several sequence(s) and image(s). This method forwards the `text`
         and `kwargs` arguments to Qwen2TokenizerFast's [`~Qwen2TokenizerFast.__call__`] if `text` is not `None` to encode
-        the text. To prepare the vision inputs, this method forwards the `vision_infos` and `kwrags` arguments to
+        the text. To prepare the vision inputs, this method forwards the `vision_infos` and `kwargs` arguments to
         Qwen2VLImageProcessor's [`~Qwen2VLImageProcessor.__call__`] if `vision_infos` is not `None`.
 
         Args:
@@ -1405,7 +1418,7 @@ class Qwen3VLProcessor(Qwen2VLProcessor):
                 (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
                 `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
             videos (`np.ndarray`, `torch.Tensor`, `list[np.ndarray]`, `list[torch.Tensor]`):
-                The image or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch
+                The video or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch
                 tensor, or a nested list of 3D frames. Both channels-first and channels-last formats are supported.
             return_tensors (`str` or [`~utils.TensorType`], *optional*):
                 If set, will return tensors of a particular framework. Acceptable values are:

transformers/models/qwen3_vl/processing_qwen3_vl.py

@@ -99,9 +99,9 @@ class Qwen3VLProcessor(ProcessorMixin):
         **kwargs: Unpack[Qwen3VLProcessorKwargs],
     ) -> BatchFeature:
         """
-        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        Main method to prepare for the model one or several sequence(s) and image(s). This method forwards the `text`
         and `kwargs` arguments to Qwen2TokenizerFast's [`~Qwen2TokenizerFast.__call__`] if `text` is not `None` to encode
-        the text. To prepare the vision inputs, this method forwards the `vision_infos` and `kwrags` arguments to
+        the text. To prepare the vision inputs, this method forwards the `vision_infos` and `kwargs` arguments to
         Qwen2VLImageProcessor's [`~Qwen2VLImageProcessor.__call__`] if `vision_infos` is not `None`.
 
         Args:
@@ -113,7 +113,7 @@ class Qwen3VLProcessor(ProcessorMixin):
                 (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
                 `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
             videos (`np.ndarray`, `torch.Tensor`, `list[np.ndarray]`, `list[torch.Tensor]`):
-                The image or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch
+                The video or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch
                 tensor, or a nested list of 3D frames. Both channels-first and channels-last formats are supported.
             return_tensors (`str` or [`~utils.TensorType`], *optional*):
                 If set, will return tensors of a particular framework. Acceptable values are: