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

transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py

@@ -37,12 +37,12 @@ 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, ModelOutput
+from ...modeling_outputs import BaseModelOutputWithPast, BaseModelOutputWithPooling, 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
-from ...utils.generic import maybe_autocast
+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 ..qwen2.modeling_qwen2 import Qwen2RMSNorm
 from .configuration_qwen2_5_vl import Qwen2_5_VLConfig, Qwen2_5_VLTextConfig, Qwen2_5_VLVisionConfig
 
@@ -214,11 +214,11 @@ class Qwen2_5_VLVisionAttention(nn.Module):
         key_states = key_states.transpose(0, 1).unsqueeze(0)
         value_states = value_states.transpose(0, 1).unsqueeze(0)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
-        if "flash" in self.config._attn_implementation:
+        if is_flash_attention_requested(self.config):
             # Flash Attention: Use cu_seqlens for variable length attention
             max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
             attn_output, _ = attention_interface(
@@ -316,6 +316,10 @@ 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)
@@ -415,7 +419,10 @@ class Qwen2_5_VisionTransformerPretrainedModel(Qwen2_5_VLPreTrainedModel):
 
         return window_index, cu_window_seqlens
 
-    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs) -> torch.Tensor:
+    @check_model_inputs
+    def forward(
+        self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs: Unpack[TransformersKwargs]
+    ) -> tuple | BaseModelOutputWithPooling:
         """
         Args:
             hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`):
@@ -469,11 +476,14 @@ class Qwen2_5_VisionTransformerPretrainedModel(Qwen2_5_VLPreTrainedModel):
                 **kwargs,
             )
 
-        hidden_states = self.merger(hidden_states)
+        merged_hidden_states = self.merger(hidden_states)
         reverse_indices = torch.argsort(window_index)
-        hidden_states = hidden_states[reverse_indices, :]
+        merged_hidden_states = merged_hidden_states[reverse_indices, :]
 
-        return hidden_states
+        return BaseModelOutputWithPooling(
+            last_hidden_state=hidden_states,
+            pooler_output=merged_hidden_states,
+        )
 
 
 @dataclass
@@ -697,9 +707,9 @@ class Qwen2_5_VLAttention(nn.Module):
             cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
             key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         attn_output, attn_weights = attention_interface(
             self,
@@ -724,7 +734,7 @@ class Qwen2_5_VLDecoderLayer(GradientCheckpointingLayer):
         super().__init__()
         self.hidden_size = config.hidden_size
 
-        if config.use_sliding_window and config._attn_implementation != "flash_attention_2":
+        if config.use_sliding_window and not is_flash_attention_requested(config):
             logger.warning_once(
                 f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
                 "unexpected results may be encountered."
@@ -1168,39 +1178,49 @@ 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
-    ):
-        """
-        Encodes videos into continuous embeddings that can be forwarded to the language model.
-
-        Args:
-            pixel_values_videos (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
-                The tensors corresponding to the input videos.
-            video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
-                The temporal, height and width of feature shape of each video in LLM.
+        self,
+        pixel_values_videos: torch.FloatTensor,
+        video_grid_thw: torch.LongTensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPooling:
+        r"""
+        pixel_values_videos (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
+            The tensors corresponding to the input videos.
+        video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
+            The temporal, height and width of feature shape of each video in LLM.
         """
         pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
-        video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
+        vision_outputs = self.visual(pixel_values_videos, grid_thw=video_grid_thw, return_dict=True, **kwargs)
         split_sizes = (video_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
-        video_embeds = torch.split(video_embeds, split_sizes)
-        return video_embeds
+        video_embeds = torch.split(vision_outputs.pooler_output, split_sizes)
+        vision_outputs.pooler_output = video_embeds
 
-    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: torch.LongTensor | None = None):
-        """
-        Encodes images into continuous embeddings that can be forwarded to the language model.
+        return vision_outputs
 
-        Args:
-            pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
-                The tensors corresponding to the input images.
-            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
-                The temporal, height and width of feature shape of each image in LLM.
+    @can_return_tuple
+    @auto_docstring
+    def get_image_features(
+        self,
+        pixel_values: torch.FloatTensor,
+        image_grid_thw: torch.LongTensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPooling:
+        r"""
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
+            The tensors corresponding to the input images.
+        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+            The temporal, height and width of feature shape of each image in LLM.
         """
         pixel_values = pixel_values.type(self.visual.dtype)
-        image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
+        vision_outputs = self.visual(pixel_values, grid_thw=image_grid_thw, return_dict=True, **kwargs)
         split_sizes = (image_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
-        image_embeds = torch.split(image_embeds, split_sizes)
-        return image_embeds
+        image_embeds = torch.split(vision_outputs.pooler_output, split_sizes)
+        vision_outputs.pooler_output = image_embeds
+
+        return vision_outputs
 
     def get_placeholder_mask(
         self,
@@ -1228,18 +1248,19 @@ 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 and inputs_embeds[special_image_mask].numel() != image_features.numel():
-            raise ValueError(
-                f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {image_features.shape[0]}"
+        if image_features is not None:
+            torch_compilable_check(
+                inputs_embeds[special_image_mask].numel() == image_features.numel(),
+                f"Image features and image tokens do not match, tokens: {n_image_tokens}, features: {image_features.shape[0]}",
             )
 
         n_video_tokens = special_video_mask.sum()
         special_video_mask = special_video_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
-        if video_features is not None and inputs_embeds[special_video_mask].numel() != video_features.numel():
-            raise ValueError(
-                f"Videos features and video tokens do not match: tokens: {n_video_tokens}, features {video_features.shape[0]}"
+        if video_features is not None:
+            torch_compilable_check(
+                inputs_embeds[special_video_mask].numel() == video_features.numel(),
+                f"Video features and video tokens do not match, tokens: {n_video_tokens}, features: {video_features.shape[0]}",
             )
-
         return special_image_mask, special_video_mask
 
     @auto_docstring
@@ -1284,7 +1305,7 @@ 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)
+            image_embeds = self.get_image_features(pixel_values, image_grid_thw, return_dict=True).pooler_output
             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
@@ -1292,7 +1313,7 @@ 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)
+            video_embeds = self.get_video_features(pixel_values_videos, video_grid_thw, return_dict=True).pooler_output
             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
@@ -1393,13 +1414,37 @@ 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
-    ):
-        return self.model.get_video_features(pixel_values_videos, video_grid_thw)
+        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
+        )
 
-    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: torch.LongTensor | None = None):
-        return self.model.get_image_features(pixel_values, image_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)
 
     @can_return_tuple
     @auto_docstring

transformers/models/qwen2_5_vl/modular_qwen2_5_vl.py

@@ -30,10 +30,12 @@ from ...configuration_utils import PreTrainedConfig
 from ...feature_extraction_utils import BatchFeature
 from ...image_utils import ImageInput
 from ...modeling_layers import GradientCheckpointingLayer
+from ...modeling_outputs import BaseModelOutputWithPooling
 from ...modeling_utils import PreTrainedModel
 from ...processing_utils import MultiModalData, ProcessingKwargs, Unpack
 from ...tokenization_utils_base import PreTokenizedInput, TextInput
 from ...utils import logging
+from ...utils.generic import check_model_inputs
 from ...video_utils import VideoInput
 from ..qwen2_vl.configuration_qwen2_vl import Qwen2VLConfig, Qwen2VLTextConfig
 from ..qwen2_vl.modeling_qwen2_vl import (
@@ -177,6 +179,10 @@ 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)
@@ -276,7 +282,10 @@ class Qwen2_5_VisionTransformerPretrainedModel(Qwen2_5_VLPreTrainedModel):
 
         return window_index, cu_window_seqlens
 
-    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs) -> torch.Tensor:
+    @check_model_inputs
+    def forward(
+        self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs: Unpack[TransformersKwargs]
+    ) -> tuple | BaseModelOutputWithPooling:
         """
         Args:
             hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`):
@@ -330,11 +339,14 @@ class Qwen2_5_VisionTransformerPretrainedModel(Qwen2_5_VLPreTrainedModel):
                 **kwargs,
             )
 
-        hidden_states = self.merger(hidden_states)
+        merged_hidden_states = self.merger(hidden_states)
         reverse_indices = torch.argsort(window_index)
-        hidden_states = hidden_states[reverse_indices, :]
+        merged_hidden_states = merged_hidden_states[reverse_indices, :]
 
-        return hidden_states
+        return BaseModelOutputWithPooling(
+            last_hidden_state=hidden_states,
+            pooler_output=merged_hidden_states,
+        )
 
 
 class Qwen2_5_VLModelOutputWithPast(Qwen2VLModelOutputWithPast):
@@ -578,7 +590,7 @@ class Qwen2_5_VLModel(Qwen2VLModel):
             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)
+            image_embeds = self.get_image_features(pixel_values, image_grid_thw, return_dict=True).pooler_output
             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
@@ -586,7 +598,7 @@ class Qwen2_5_VLModel(Qwen2VLModel):
             inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
 
         if pixel_values_videos is not None:
-            video_embeds = self.get_video_features(pixel_values_videos, video_grid_thw)
+            video_embeds = self.get_video_features(pixel_values_videos, video_grid_thw, return_dict=True).pooler_output
             video_embeds = 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

transformers/models/qwen2_audio/modeling_qwen2_audio.py

@@ -27,7 +27,7 @@ from ...masking_utils import create_bidirectional_mask
 from ...modeling_layers import GradientCheckpointingLayer
 from ...modeling_outputs import BaseModelOutput, ModelOutput
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
-from ...utils import auto_docstring, logging
+from ...utils import auto_docstring, is_torchdynamo_compiling, logging, torch_compilable_check
 from ..auto import AutoModel, AutoModelForCausalLM
 from .configuration_qwen2_audio import Qwen2AudioConfig, Qwen2AudioEncoderConfig
 
@@ -162,9 +162,9 @@ class Qwen2AudioAttention(nn.Module):
         key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
         value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         attn_output, attn_weights = attention_interface(
             self,
@@ -284,7 +284,6 @@ class Qwen2AudioEncoder(Qwen2AudioPreTrainedModel):
 
         embed_dim = config.d_model
         self.num_mel_bins = config.num_mel_bins
-        self.padding_idx = config.pad_token_id
         self.max_source_positions = config.max_source_positions
         self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
 
@@ -439,7 +438,9 @@ class Qwen2AudioForConditionalGeneration(Qwen2AudioPreTrainedModel, GenerationMi
         self.vocab_size = config.text_config.vocab_size
         self.language_model = AutoModelForCausalLM.from_config(config.text_config)
 
-        self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
+        self.pad_token_id = (
+            self.config.text_config.pad_token_id if self.config.text_config.pad_token_id is not None else -1
+        )
         self._padding_side = "left"  # set it to left by default, user can use setter to change padding_sides
         self.post_init()
 
@@ -774,7 +775,7 @@ class Qwen2AudioForConditionalGeneration(Qwen2AudioPreTrainedModel, GenerationMi
                 audio_tokens = input_ids == self.config.audio_token_id
                 legacy_processing = (audio_tokens[:, :-1] & audio_tokens[:, 1:]).sum() == 0
 
-                if legacy_processing:
+                if not is_torchdynamo_compiling() and legacy_processing:
                     logger.warning_once(
                         "Expanding inputs for audio tokens in Qwen2Audio should be done in processing."
                     )
@@ -789,11 +790,10 @@ class Qwen2AudioForConditionalGeneration(Qwen2AudioPreTrainedModel, GenerationMi
 
                     n_audio_tokens = (input_ids == self.config.audio_token_id).sum().item()
                     n_audio_features = audio_features.shape[0]
-
-                    if n_audio_tokens != n_audio_features:
-                        raise ValueError(
-                            f"Audio features and audio tokens do not match: tokens: {n_audio_tokens}, features {n_audio_features}"
-                        )
+                    torch_compilable_check(
+                        n_audio_tokens == n_audio_features,
+                        f"Audio features and audio tokens do not match, tokens: {n_audio_tokens}, features: {n_audio_features}",
+                    )
                     special_audio_mask = (input_ids == self.config.audio_token_id).to(inputs_embeds.device)
                     special_audio_mask = special_audio_mask.unsqueeze(-1).expand_as(inputs_embeds)
                     audio_features = audio_features.to(inputs_embeds.device, inputs_embeds.dtype)

transformers/models/qwen2_moe/configuration_qwen2_moe.py

@@ -101,6 +101,12 @@ class Qwen2MoeConfig(PreTrainedConfig):
             Whether to add a bias to the queries, keys and values.
         layer_types (`dict[int, str]`, *optional*): a dictionarry that explicitly maps layer index with
             the attention type. The attention type is one of `sliding_attention`, `full_attention`.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        bos_token_id (`int`, *optional*):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*):
+            End of stream token id.
     ```python
     >>> from transformers import Qwen2MoeModel, Qwen2MoeConfig
 
@@ -163,6 +169,9 @@ class Qwen2MoeConfig(PreTrainedConfig):
         mlp_only_layers: bool | None = None,
         qkv_bias: bool | None = True,
         layer_types: list[str] | None = None,
+        pad_token_id: int | None = None,
+        bos_token_id: int | None = None,
+        eos_token_id: int | None = None,
         **kwargs,
     ):
         self.layer_types = layer_types
@@ -202,13 +211,14 @@ class Qwen2MoeConfig(PreTrainedConfig):
                 for i in range(self.num_hidden_layers)
             ]
         layer_type_validation(self.layer_types)
+        self.pad_token_id = pad_token_id
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.tie_word_embeddings = tie_word_embeddings
 
         self.rope_parameters = rope_parameters
 
-        super().__init__(
-            tie_word_embeddings=tie_word_embeddings,
-            **kwargs,
-        )
+        super().__init__(**kwargs)
 
 
 __all__ = ["Qwen2MoeConfig"]

transformers/models/qwen2_moe/modeling_qwen2_moe.py

@@ -274,9 +274,9 @@ class Qwen2MoeAttention(nn.Module):
             cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
             key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         attn_output, attn_weights = attention_interface(
             self,
@@ -629,7 +629,7 @@ def load_balancing_loss_func(
 @auto_docstring
 class Qwen2MoeForCausalLM(Qwen2MoePreTrainedModel, GenerationMixin):
     _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
-    _tp_plan = {"lm_head": "colwise_rep"}
+    _tp_plan = {"lm_head": "colwise_gather_output"}
     _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
 
     def __init__(self, config):

transformers/models/qwen2_moe/modular_qwen2_moe.py

@@ -129,9 +129,9 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
         return expert_output
 
 
-class Qwen2MoeDecoderLayer(LlamaDecoderLayer, nn.Module):
+class Qwen2MoeDecoderLayer(LlamaDecoderLayer):
     def __init__(self, config: Qwen2MoeConfig, layer_idx: int):
-        nn.Module.__init__()
+        nn.Module.__init__(self)
         self.self_attn = Qwen2MoeAttention(config, layer_idx)
         if (layer_idx not in config.mlp_only_layers) and (
             config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
@@ -235,7 +235,7 @@ class Qwen2MoeModel(MixtralModel):
 
 class Qwen2MoeForCausalLM(MixtralForCausalLM, GenerationMixin):
     _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
-    _tp_plan = {"lm_head": "colwise_rep"}
+    _tp_plan = {"lm_head": "colwise_gather_output"}
     _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
 
     def __init__(self, config):

transformers/models/qwen2_vl/configuration_qwen2_vl.py

@@ -97,8 +97,6 @@ class Qwen2VLTextConfig(PreTrainedConfig):
         use_cache (`bool`, *optional*, defaults to `True`):
             Whether or not the model should return the last key/values attentions (not used by all models). Only
             relevant if `config.is_decoder=True`.
-        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
-            Whether the model's input and output word embeddings should be tied.
         use_sliding_window (`bool`, *optional*, defaults to `False`):
             Whether to use sliding window attention.
         sliding_window (`int`, *optional*, defaults to 4096):
@@ -167,7 +165,6 @@ class Qwen2VLTextConfig(PreTrainedConfig):
         initializer_range: float | None = 0.02,
         rms_norm_eps: int | None = 1e-05,
         use_cache: bool | None = True,
-        tie_word_embeddings: bool | None = False,
         use_sliding_window: bool | None = False,
         sliding_window: int | None = 4096,
         max_window_layers: int | None = 80,
@@ -211,11 +208,10 @@ class Qwen2VLTextConfig(PreTrainedConfig):
         layer_type_validation(self.layer_types, self.num_hidden_layers)
 
         self.rope_parameters = rope_parameters
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.pad_token_id = pad_token_id
         super().__init__(
-            tie_word_embeddings=tie_word_embeddings,
-            bos_token_id=bos_token_id,
-            eos_token_id=eos_token_id,
-            pad_token_id=pad_token_id,
             ignore_keys_at_rope_validation={"mrope_section"},
             **kwargs,
         )
@@ -258,6 +254,8 @@ class Qwen2VLConfig(PreTrainedConfig):
             The token index to denote start of vision input.
         vision_end_token_id (`int`, *optional*, defaults to 151653):
             The token index to denote end of vision input.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
 
     ```python
     >>> from transformers import Qwen2VLForConditionalGeneration, Qwen2VLConfig
@@ -284,6 +282,7 @@ class Qwen2VLConfig(PreTrainedConfig):
         video_token_id=151656,
         vision_start_token_id=151652,
         vision_end_token_id=151653,
+        tie_word_embeddings=False,
         **kwargs,
     ):
         if isinstance(vision_config, dict):
@@ -305,9 +304,7 @@ class Qwen2VLConfig(PreTrainedConfig):
         self.video_token_id = video_token_id
         self.vision_start_token_id = vision_start_token_id
         self.vision_end_token_id = vision_end_token_id
-
-        # FIXME: arthur/cyril - tying has to be used from the text config
-        kwargs["tie_word_embeddings"] = self.text_config.tie_word_embeddings
+        self.tie_word_embeddings = tie_word_embeddings
         super().__init__(**kwargs)
 
 

transformers/models/qwen2_vl/image_processing_qwen2_vl_fast.py

@@ -21,7 +21,7 @@
 from typing import Optional, Union
 
 import torch
-from torchvision.transforms.v2 import functional as F
+import torchvision.transforms.v2.functional as tvF
 
 from ...image_processing_utils import BatchFeature
 from ...image_processing_utils_fast import (
@@ -62,8 +62,6 @@ class Qwen2VLImageProcessorFast(BaseImageProcessorFast):
     patch_size = 14
     temporal_patch_size = 2
     merge_size = 2
-    min_pixels = None
-    max_pixels = None
     valid_kwargs = Qwen2VLImageProcessorKwargs
     model_input_names = ["pixel_values", "image_grid_thw"]
 
@@ -82,7 +80,7 @@ class Qwen2VLImageProcessorFast(BaseImageProcessorFast):
         if "shortest_edge" not in size or "longest_edge" not in size:
             raise ValueError("size must contain 'shortest_edge' and 'longest_edge' keys.")
 
-        super().__init__(size=size, min_pixels=min_pixels, max_pixels=max_pixels, **kwargs)
+        super().__init__(size=size, **kwargs)
 
     def _further_process_kwargs(
         self,
@@ -105,7 +103,7 @@ class Qwen2VLImageProcessorFast(BaseImageProcessorFast):
         else:
             size = {**self.size}
 
-        return super()._further_process_kwargs(size=size, min_pixels=min_pixels, max_pixels=max_pixels, **kwargs)
+        return super()._further_process_kwargs(size=size, **kwargs)
 
     @auto_docstring
     def preprocess(
@@ -141,7 +139,7 @@ class Qwen2VLImageProcessorFast(BaseImageProcessorFast):
         images: list["torch.Tensor"],
         do_resize: bool,
         size: SizeDict,
-        interpolation: Optional["F.InterpolationMode"],
+        interpolation: Optional["tvF.InterpolationMode"],
         do_rescale: bool,
         rescale_factor: float,
         do_normalize: bool,