transformers 4.53.1__tar.gz → 4.53.2__tar.gz

{transformers-4.53.1/src/transformers.egg-info → transformers-4.53.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: transformers
-Version: 4.53.1
+Version: 4.53.2
 Summary: State-of-the-art Machine Learning for JAX, PyTorch and TensorFlow
 Home-page: https://github.com/huggingface/transformers
 Author: The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/transformers/graphs/contributors)

{transformers-4.53.1 → transformers-4.53.2}/setup.py

@@ -457,7 +457,7 @@ install_requires = [
 
 setup(
     name="transformers",
-    version="4.53.1",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
+    version="4.53.2",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
     author="The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/transformers/graphs/contributors)",
     author_email="transformers@huggingface.co",
     description="State-of-the-art Machine Learning for JAX, PyTorch and TensorFlow",

{transformers-4.53.1 → transformers-4.53.2}/src/transformers/__init__.py

@@ -18,7 +18,7 @@
 # to defer the actual importing for when the objects are requested. This way `import transformers` provides the names
 # in the namespace without actually importing anything (and especially none of the backends).
 
-__version__ = "4.53.1"
+__version__ = "4.53.2"
 
 from pathlib import Path
 from typing import TYPE_CHECKING

{transformers-4.53.1 → transformers-4.53.2}/src/transformers/generation/continuous_batching.py

@@ -122,6 +122,11 @@ class RequestState:
         is_eos = token_id == self.eos_token_id and self.eos_token_id != -1
         is_max_len = self.generated_len() >= self.max_new_tokens
 
+        # Only add the token if we're not finishing due to max length
+        # (EOS tokens should still be added to the output)
+        if not (is_max_len and not is_eos):
+            self.static_outputs.extend([token_id])
+
         if is_eos or is_max_len:
             self.status = RequestStatus.FINISHED
             return True
@@ -1011,7 +1016,6 @@ class ContinuousBatchProcessor:
                 self.metrics.record_ttft_metric(state.created_time, state.request_id)
                 state.status = RequestStatus.DECODING
                 token = out_tokens[self.logits_indices[i]]
-                state.static_outputs.extend([token])
                 state.prompt_ids = [token]
                 if state.update_with_token(token):
                     self.metrics.record_request_completion(state.created_time, state.request_id)

{transformers-4.53.1 → transformers-4.53.2}/src/transformers/masking_utils.py

@@ -599,7 +599,7 @@ class AttentionMaskInterface(GeneralInterface):
 ALL_MASK_ATTENTION_FUNCTIONS: AttentionMaskInterface = AttentionMaskInterface()
 
 
-def find_packed_sequence_indices(position_ids: torch.Tensor) -> Optional[torch.Tensor]:
+def find_packed_sequence_indices(position_ids: torch.Tensor) -> torch.Tensor:
     """
     Find the indices of the sequence to which each new query token in the sequence belongs when using packed
     tensor format (i.e. several sequences packed in the same batch dimension).
@@ -713,7 +713,7 @@ def create_causal_mask(
     attention_mask: Optional[torch.Tensor],
     cache_position: torch.Tensor,
     past_key_values: Optional[Cache],
-    position_ids: Optional[torch.Tensor],
+    position_ids: Optional[torch.Tensor] = None,
     or_mask_function: Optional[Callable] = None,
     and_mask_function: Optional[Callable] = None,
 ) -> Optional[Union[torch.Tensor, BlockMask]]:
@@ -802,7 +802,7 @@ def create_sliding_window_causal_mask(
     attention_mask: Optional[torch.Tensor],
     cache_position: torch.Tensor,
     past_key_values: Optional[Cache],
-    position_ids: Optional[torch.Tensor],
+    position_ids: Optional[torch.Tensor] = None,
     or_mask_function: Optional[Callable] = None,
     and_mask_function: Optional[Callable] = None,
 ) -> Optional[Union[torch.Tensor, BlockMask]]:
@@ -897,7 +897,7 @@ def create_chunked_causal_mask(
     attention_mask: Optional[torch.Tensor],
     cache_position: torch.Tensor,
     past_key_values: Optional[Cache],
-    position_ids: Optional[torch.Tensor],
+    position_ids: Optional[torch.Tensor] = None,
     or_mask_function: Optional[Callable] = None,
     and_mask_function: Optional[Callable] = None,
 ) -> Optional[Union[torch.Tensor, BlockMask]]:
@@ -1006,7 +1006,7 @@ def create_masks_for_generate(
     attention_mask: Optional[torch.Tensor],
     cache_position: torch.Tensor,
     past_key_values: Optional[Cache],
-    position_ids: Optional[torch.Tensor],
+    position_ids: Optional[torch.Tensor] = None,
     or_mask_function: Optional[Callable] = None,
     and_mask_function: Optional[Callable] = None,
     **kwargs,

{transformers-4.53.1 → transformers-4.53.2}/src/transformers/modeling_flash_attention_utils.py

@@ -103,6 +103,16 @@ if is_flash_attn_2_available():
     from flash_attn.bert_padding import unpad_input as unpad_input_fa2
     from flash_attn.layers.rotary import apply_rotary_emb
 
+    HAS_FA2 = True
+    FA_VERSION = 2
+elif is_torch_npu_available():
+    # patch functions in package `flash-attn` when using flash-attention on Ascend NPU.
+    from .integrations.npu_flash_attention import npu_apply_rotary_emb as apply_rotary_emb  # noqa: F401
+    from .integrations.npu_flash_attention import npu_flash_attn_func as flash_attn_2_func
+    from .integrations.npu_flash_attention import npu_flash_attn_varlen_func as flash_attn_2_varlen_func
+    from .integrations.npu_flash_attention import pad_input as pad_input_fa2
+    from .integrations.npu_flash_attention import unpad_input as unpad_input_fa2
+
     HAS_FA2 = True
     FA_VERSION = 2
 else:
@@ -136,22 +146,6 @@ if FA_VERSION:
     unpad_input = globals()[f"unpad_input_fa{FA_VERSION}"]
     pad_input = globals()[f"pad_input_fa{FA_VERSION}"]
 
-# patch functions in package `flash-attn` when using flash-attention on Ascend NPU.
-if is_torch_npu_available():
-    from .integrations.npu_flash_attention import (
-        npu_apply_rotary_emb as apply_rotary_emb,  # noqa: F401
-    )
-    from .integrations.npu_flash_attention import (
-        npu_flash_attn_func as flash_attn_func,
-    )
-    from .integrations.npu_flash_attention import (
-        npu_flash_attn_varlen_func as flash_attn_varlen_func,
-    )
-    from .integrations.npu_flash_attention import (
-        pad_input,
-        unpad_input,
-    )
-
 
 _flash_supports_window_size = False
 

{transformers-4.53.1 → transformers-4.53.2}/src/transformers/models/auto/tokenization_auto.py

@@ -573,6 +573,7 @@ TOKENIZER_MAPPING_NAMES = OrderedDict[str, tuple[Optional[str], Optional[str]]](
                 "GemmaTokenizerFast" if is_tokenizers_available() else None,
             ),
         ),
+        ("smollm3", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)),
         ("speech_to_text", ("Speech2TextTokenizer" if is_sentencepiece_available() else None, None)),
         ("speech_to_text_2", ("Speech2Text2Tokenizer", None)),
         ("speecht5", ("SpeechT5Tokenizer" if is_sentencepiece_available() else None, None)),

{transformers-4.53.1 → transformers-4.53.2}/src/transformers/models/cohere2/configuration_cohere2.py

@@ -19,6 +19,8 @@
 # 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.
+import warnings
+
 from ...configuration_utils import PretrainedConfig, layer_type_validation
 from ...modeling_rope_utils import rope_config_validation
 
@@ -216,14 +218,29 @@ class Cohere2Config(PretrainedConfig):
             **kwargs,
         )
 
+        # BC -> the pattern used to be a simple int, and it's still present in configs on the Hub
+        self._sliding_window_pattern = kwargs.get("sliding_window_pattern", 4)
+
         if self.layer_types is None:
             # BC -> the pattern used to be a simple int, and it's still present in configs on the Hub
-            sliding_window_pattern = getattr(self, "sliding_window_pattern", 4)
+            self._sliding_window_pattern = getattr(self, "sliding_window_pattern", 4)
             self.layer_types = [
-                "sliding_attention" if bool((i + 1) % sliding_window_pattern) else "full_attention"
+                "sliding_attention" if bool((i + 1) % self._sliding_window_pattern) else "full_attention"
                 for i in range(self.num_hidden_layers)
             ]
         layer_type_validation(self.layer_types)
 
+    @property
+    def sliding_window_pattern(self):
+        warnings.warn(
+            "The `sliding_window_pattern` attribute is deprecated and will be removed in v4.55.0.",
+            FutureWarning,
+        )
+        return self._sliding_window_pattern
+
+    @sliding_window_pattern.setter
+    def sliding_window_pattern(self, value):
+        self._sliding_window_pattern = value
+
 
 __all__ = ["Cohere2Config"]

{transformers-4.53.1 → transformers-4.53.2}/src/transformers/models/cohere2/modular_cohere2.py

@@ -13,6 +13,7 @@
 # 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.
+import warnings
 from typing import Callable, Optional
 
 import torch
@@ -238,15 +239,30 @@ class Cohere2Config(PretrainedConfig):
             **kwargs,
         )
 
+        # BC -> the pattern used to be a simple int, and it's still present in configs on the Hub
+        self._sliding_window_pattern = kwargs.get("sliding_window_pattern", 4)
+
         if self.layer_types is None:
             # BC -> the pattern used to be a simple int, and it's still present in configs on the Hub
-            sliding_window_pattern = getattr(self, "sliding_window_pattern", 4)
+            self._sliding_window_pattern = getattr(self, "sliding_window_pattern", 4)
             self.layer_types = [
-                "sliding_attention" if bool((i + 1) % sliding_window_pattern) else "full_attention"
+                "sliding_attention" if bool((i + 1) % self._sliding_window_pattern) else "full_attention"
                 for i in range(self.num_hidden_layers)
             ]
         layer_type_validation(self.layer_types)
 
+    @property
+    def sliding_window_pattern(self):
+        warnings.warn(
+            "The `sliding_window_pattern` attribute is deprecated and will be removed in v4.55.0.",
+            FutureWarning,
+        )
+        return self._sliding_window_pattern
+
+    @sliding_window_pattern.setter
+    def sliding_window_pattern(self, value):
+        self._sliding_window_pattern = value
+
 
 class Cohere2RotaryEmbedding(CohereRotaryEmbedding):
     pass

{transformers-4.53.1 → transformers-4.53.2}/src/transformers/models/gemma3/configuration_gemma3.py

@@ -19,6 +19,7 @@
 # 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.
+import warnings
 from typing import Any, Optional, Union
 
 from ...configuration_utils import PretrainedConfig, layer_type_validation
@@ -145,10 +146,6 @@ class Gemma3TextConfig(PretrainedConfig):
     >>> # Accessing the model configuration
     >>> configuration = model.config
     ```
-        rope_local_base_freq (float, *optional*, defaults to 10000.0):
-            The base period of the RoPE embeddings for local attention.
-        sliding_window_pattern (`int`, *optional*, defaults to 6):
-            Pattern for the sliding window attention.
     """
 
     model_type = "gemma3_text"
@@ -230,15 +227,28 @@ class Gemma3TextConfig(PretrainedConfig):
         self.rope_scaling = rope_scaling
         rope_config_validation(self)
 
+        # BC -> the pattern used to be a simple int, and it's still present in configs on the Hub
+        self._sliding_window_pattern = kwargs.get("sliding_window_pattern", 6)
+
         if self.layer_types is None:
-            # BC -> the pattern used to be a simple int, and it's still present in configs on the Hub
-            sliding_window_pattern = getattr(self, "sliding_window_pattern", 6)
             self.layer_types = [
-                "sliding_attention" if bool((i + 1) % sliding_window_pattern) else "full_attention"
+                "sliding_attention" if bool((i + 1) % self._sliding_window_pattern) else "full_attention"
                 for i in range(self.num_hidden_layers)
             ]
         layer_type_validation(self.layer_types)
 
+    @property
+    def sliding_window_pattern(self):
+        warnings.warn(
+            "The `sliding_window_pattern` attribute is deprecated and will be removed in v4.55.0.",
+            FutureWarning,
+        )
+        return self._sliding_window_pattern
+
+    @sliding_window_pattern.setter
+    def sliding_window_pattern(self, value):
+        self._sliding_window_pattern = value
+
 
 class Gemma3Config(PretrainedConfig):
     r"""

{transformers-4.53.1 → transformers-4.53.2}/src/transformers/models/gemma3/modular_gemma3.py

@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import copy
+import warnings
 from collections.abc import Callable
 from typing import Any, Optional, Union
 
@@ -171,10 +172,6 @@ class Gemma3TextConfig(Gemma2Config, PretrainedConfig):
     >>> # Accessing the model configuration
     >>> configuration = model.config
     ```
-        rope_local_base_freq (float, *optional*, defaults to 10000.0):
-            The base period of the RoPE embeddings for local attention.
-        sliding_window_pattern (`int`, *optional*, defaults to 6):
-            Pattern for the sliding window attention.
     """
 
     model_type = "gemma3_text"
@@ -241,15 +238,28 @@ class Gemma3TextConfig(Gemma2Config, PretrainedConfig):
         self.rope_scaling = rope_scaling
         rope_config_validation(self)
 
+        # BC -> the pattern used to be a simple int, and it's still present in configs on the Hub
+        self._sliding_window_pattern = kwargs.get("sliding_window_pattern", 6)
+
         if self.layer_types is None:
-            # BC -> the pattern used to be a simple int, and it's still present in configs on the Hub
-            sliding_window_pattern = getattr(self, "sliding_window_pattern", 6)
             self.layer_types = [
-                "sliding_attention" if bool((i + 1) % sliding_window_pattern) else "full_attention"
+                "sliding_attention" if bool((i + 1) % self._sliding_window_pattern) else "full_attention"
                 for i in range(self.num_hidden_layers)
             ]
         layer_type_validation(self.layer_types)
 
+    @property
+    def sliding_window_pattern(self):
+        warnings.warn(
+            "The `sliding_window_pattern` attribute is deprecated and will be removed in v4.55.0.",
+            FutureWarning,
+        )
+        return self._sliding_window_pattern
+
+    @sliding_window_pattern.setter
+    def sliding_window_pattern(self, value):
+        self._sliding_window_pattern = value
+
 
 class Gemma3Config(PretrainedConfig):
     r"""

{transformers-4.53.1 → transformers-4.53.2}/src/transformers/models/glm4v/image_processing_glm4v_fast.py

@@ -121,6 +121,7 @@ class Glm4vImageProcessorFast(BaseImageProcessorFast):
         do_convert_rgb: bool,
         input_data_format: Optional[Union[str, ChannelDimension]],
         device: Optional[Union[str, torch.device]],
+        disable_grouping: Optional[bool],
     ):
         """
         Preprocess an image or batch of images. Copy of the `preprocess` method from `CLIPImageProcessor`.
@@ -173,7 +174,7 @@ class Glm4vImageProcessorFast(BaseImageProcessorFast):
         resized_height, resized_width = height, width
 
         # Group images by size for batched resizing
-        grouped_images, grouped_images_index = group_images_by_shape(images)
+        grouped_images, grouped_images_index = group_images_by_shape(images, disable_grouping=disable_grouping)
         resized_images_grouped = {}
         for shape, stacked_images in grouped_images.items():
             if do_resize:
@@ -191,7 +192,7 @@ class Glm4vImageProcessorFast(BaseImageProcessorFast):
         resized_images = reorder_images(resized_images_grouped, grouped_images_index)
         # Group images by size for further processing
         # Needed in case do_resize is False, or resize returns images with different sizes
-        grouped_images, grouped_images_index = group_images_by_shape(resized_images)
+        grouped_images, grouped_images_index = group_images_by_shape(resized_images, disable_grouping=disable_grouping)
         processed_images_grouped = {}
         for shape, stacked_images in grouped_images.items():
             # Fused rescale and normalize
@@ -249,6 +250,7 @@ class Glm4vImageProcessorFast(BaseImageProcessorFast):
         data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
         input_data_format: Optional[Union[str, ChannelDimension]] = None,
         device: Optional["torch.device"] = None,
+        disable_grouping: Optional[bool] = False,
         **kwargs,
     ):
         r"""
@@ -323,6 +325,7 @@ class Glm4vImageProcessorFast(BaseImageProcessorFast):
                     do_convert_rgb=do_convert_rgb,
                     input_data_format=input_data_format,
                     device=device,
+                    disable_grouping=disable_grouping,
                 )
                 pixel_values.extend(patches)
                 vision_grid_thws.append(image_grid_thw)
@@ -351,11 +354,11 @@ class Glm4vImageProcessorFast(BaseImageProcessorFast):
 
         factor = patch_size * merge_size
         resized_height, resized_width = smart_resize(
-            t=self.temporal_patch_size,
+            num_frames=self.temporal_patch_size,
             height=height,
             width=width,
+            temporal_factor=self.temporal_patch_size,
             factor=factor,
-            t_factor=self.temporal_patch_size,
         )
         grid_h, grid_w = resized_height // patch_size, resized_width // patch_size
         return grid_h * grid_w

{transformers-4.53.1 → transformers-4.53.2}/src/transformers/models/glm4v/modeling_glm4v.py

@@ -952,7 +952,7 @@ class Glm4vTextModel(Glm4vPreTrainedModel):
 @auto_docstring
 class Glm4vModel(Glm4vPreTrainedModel):
     base_model_prefix = ""
-    _checkpoint_conversion_mapping = None
+    _checkpoint_conversion_mapping = {}
     config_class = Glm4vConfig
     _no_split_modules = ["Glm4vTextDecoderLayer", "Glm4vVisionBlock"]
 
@@ -1053,7 +1053,8 @@ class Glm4vModel(Glm4vPreTrainedModel):
                 dtype=input_ids.dtype,
                 device=input_ids.device,
             )
-
+            image_index, video_index = 0, 0
+            video_group_index = 0
             attention_mask = attention_mask.to(total_input_ids.device)
             for i, input_ids in enumerate(total_input_ids):
                 input_ids = input_ids[attention_mask[i] == 1]
@@ -1083,8 +1084,6 @@ class Glm4vModel(Glm4vPreTrainedModel):
 
                 llm_pos_ids_list = []
                 video_frame_num = 1
-                image_index, video_index = 0, 0
-
                 for modality_type, start_idx, end_idx in input_type_group:
                     st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
 
@@ -1125,12 +1124,14 @@ class Glm4vModel(Glm4vPreTrainedModel):
                             t_index = torch.tensor(t_idx).view(-1, 1).expand(-1, llm_grid_h * llm_grid_w).flatten()
 
                             h_index = torch.arange(llm_grid_h).view(1, -1, 1).expand(1, -1, llm_grid_w).flatten()
-
                             w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(1, llm_grid_h, -1).flatten()
-
                             llm_pos_ids_list.append(torch.stack([t_index, h_index, w_index]) + st_idx)
 
-                        video_index += 1
+                        video_group_index += 1
+
+                        if video_group_index >= video_grid_thw[video_index][0]:
+                            video_index += 1
+                            video_group_index = 0
 
                         video_frame_num += 1
 
@@ -1179,7 +1180,13 @@ class Glm4vModel(Glm4vPreTrainedModel):
                 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)
+        # reshape video_grid_thw -> [b, 3] -> [1, h, w] * frames
+        temp_frames_hw = []
+        for t, h, w in video_grid_thw:
+            repeated_row = torch.tensor([1, h.item(), w.item()]).unsqueeze(0).repeat(t, 1)
+            temp_frames_hw.append(repeated_row)
+        flattened_video_grid_thw = torch.cat(temp_frames_hw, dim=0)
+        video_embeds = self.visual(pixel_values_videos, grid_thw=flattened_video_grid_thw)
         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
@@ -1379,7 +1386,7 @@ class Glm4vCausalLMOutputWithPast(ModelOutput):
 
 
 class Glm4vForConditionalGeneration(Glm4vPreTrainedModel, GenerationMixin):
-    _checkpoint_conversion_mapping = None
+    _checkpoint_conversion_mapping = {}
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):

{transformers-4.53.1 → transformers-4.53.2}/src/transformers/models/glm4v/modular_glm4v.py

@@ -1004,7 +1004,7 @@ class Glm4vTextModel(Qwen2_5_VLTextModel):
 
 
 class Glm4vModel(Qwen2_5_VLModel):
-    _checkpoint_conversion_mapping = None
+    _checkpoint_conversion_mapping = {}
     _no_split_modules = ["Glm4vTextDecoderLayer", "Glm4vVisionBlock"]
 
     def __init__(self, config):
@@ -1087,7 +1087,8 @@ class Glm4vModel(Qwen2_5_VLModel):
                 dtype=input_ids.dtype,
                 device=input_ids.device,
             )
-
+            image_index, video_index = 0, 0
+            video_group_index = 0
             attention_mask = attention_mask.to(total_input_ids.device)
             for i, input_ids in enumerate(total_input_ids):
                 input_ids = input_ids[attention_mask[i] == 1]
@@ -1117,8 +1118,6 @@ class Glm4vModel(Qwen2_5_VLModel):
 
                 llm_pos_ids_list = []
                 video_frame_num = 1
-                image_index, video_index = 0, 0
-
                 for modality_type, start_idx, end_idx in input_type_group:
                     st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
 
@@ -1159,12 +1158,14 @@ class Glm4vModel(Qwen2_5_VLModel):
                             t_index = torch.tensor(t_idx).view(-1, 1).expand(-1, llm_grid_h * llm_grid_w).flatten()
 
                             h_index = torch.arange(llm_grid_h).view(1, -1, 1).expand(1, -1, llm_grid_w).flatten()
-
                             w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(1, llm_grid_h, -1).flatten()
-
                             llm_pos_ids_list.append(torch.stack([t_index, h_index, w_index]) + st_idx)
 
-                        video_index += 1
+                        video_group_index += 1
+
+                        if video_group_index >= video_grid_thw[video_index][0]:
+                            video_index += 1
+                            video_group_index = 0
 
                         video_frame_num += 1
 
@@ -1200,6 +1201,30 @@ class Glm4vModel(Qwen2_5_VLModel):
 
             return position_ids, mrope_position_deltas
 
+    def get_video_features(
+        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: Optional[torch.LongTensor] = None
+    ):
+        """
+        Encodes videos into continuous embeddings that can be forwarded to the language model.
+
+        Args:
+            pixel_values_videos (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
+                The tensors corresponding to the input videos.
+            video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
+                The temporal, height and width of feature shape of each video in LLM.
+        """
+        pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
+        # reshape video_grid_thw -> [b, 3] -> [1, h, w] * frames
+        temp_frames_hw = []
+        for t, h, w in video_grid_thw:
+            repeated_row = torch.tensor([1, h.item(), w.item()]).unsqueeze(0).repeat(t, 1)
+            temp_frames_hw.append(repeated_row)
+        flattened_video_grid_thw = torch.cat(temp_frames_hw, dim=0)
+        video_embeds = self.visual(pixel_values_videos, grid_thw=flattened_video_grid_thw)
+        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
+
     @auto_docstring
     @can_return_tuple
     def forward(
@@ -1353,7 +1378,7 @@ class Glm4vCausalLMOutputWithPast(Qwen2_5_VLCausalLMOutputWithPast):
 
 
 class Glm4vForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
-    _checkpoint_conversion_mapping = None
+    _checkpoint_conversion_mapping = {}
 
     def forward(
         self,
@@ -1661,32 +1686,38 @@ class Glm4vProcessor(Qwen2_5_VLProcessor):
             video_index = 0
             for i in range(len(text)):
                 while self.video_token in text[i]:
-                    num_frames = len(video_grid_thw)
+                    num_frames = video_grid_thw[video_index][0]
                     video_structure = ""
 
                     if hasattr(timestamps, "tolist"):
                         timestamps_list = timestamps.tolist()[0]
                     else:
                         timestamps_list = timestamps[0] if isinstance(timestamps[0], list) else timestamps
+
                     unique_timestamps = []
                     for idx in range(0, len(timestamps_list)):
                         unique_timestamps.append(timestamps_list[idx])
+
                     selected_timestamps = unique_timestamps[:num_frames]
                     while len(selected_timestamps) < num_frames:
                         selected_timestamps.append(selected_timestamps[-1] if selected_timestamps else 0)
+
                     for frame_idx in range(num_frames):
                         timestamp_sec = selected_timestamps[frame_idx]
                         frame_structure = f"<|begin_of_image|>{self.image_token}<|end_of_image|>{timestamp_sec}"
                         video_structure += frame_structure
+
                     text[i] = text[i].replace(self.video_token, video_structure, 1)
+                    num_image_tokens = (
+                        video_grid_thw[video_index].prod() // merge_length // video_grid_thw[video_index][0]
+                    )
+                    for frame_idx in range(num_frames):
+                        if self.image_token in text[i]:
+                            text[i] = text[i].replace(self.image_token, "<|placeholder|>" * num_image_tokens, 1)
+
                     video_index += 1
 
-                for frame_idx in range(len(video_grid_thw)):
-                    if self.image_token in text[i]:
-                        num_image_tokens = video_grid_thw[frame_idx].prod() // merge_length
-                        text[i] = text[i].replace(self.image_token, "<|placeholder|>" * num_image_tokens, 1)
                 text[i] = text[i].replace("<|placeholder|>", self.image_token)
-
         return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)
         text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"])
         self._check_special_mm_tokens(text, text_inputs, modalities=["image", "video"])

{transformers-4.53.1 → transformers-4.53.2}/src/transformers/models/glm4v/processing_glm4v.py

@@ -167,32 +167,38 @@ class Glm4vProcessor(ProcessorMixin):
             video_index = 0
             for i in range(len(text)):
                 while self.video_token in text[i]:
-                    num_frames = len(video_grid_thw)
+                    num_frames = video_grid_thw[video_index][0]
                     video_structure = ""
 
                     if hasattr(timestamps, "tolist"):
                         timestamps_list = timestamps.tolist()[0]
                     else:
                         timestamps_list = timestamps[0] if isinstance(timestamps[0], list) else timestamps
+
                     unique_timestamps = []
                     for idx in range(0, len(timestamps_list)):
                         unique_timestamps.append(timestamps_list[idx])
+
                     selected_timestamps = unique_timestamps[:num_frames]
                     while len(selected_timestamps) < num_frames:
                         selected_timestamps.append(selected_timestamps[-1] if selected_timestamps else 0)
+
                     for frame_idx in range(num_frames):
                         timestamp_sec = selected_timestamps[frame_idx]
                         frame_structure = f"<|begin_of_image|>{self.image_token}<|end_of_image|>{timestamp_sec}"
                         video_structure += frame_structure
+
                     text[i] = text[i].replace(self.video_token, video_structure, 1)
+                    num_image_tokens = (
+                        video_grid_thw[video_index].prod() // merge_length // video_grid_thw[video_index][0]
+                    )
+                    for frame_idx in range(num_frames):
+                        if self.image_token in text[i]:
+                            text[i] = text[i].replace(self.image_token, "<|placeholder|>" * num_image_tokens, 1)
+
                     video_index += 1
 
-                for frame_idx in range(len(video_grid_thw)):
-                    if self.image_token in text[i]:
-                        num_image_tokens = video_grid_thw[frame_idx].prod() // merge_length
-                        text[i] = text[i].replace(self.image_token, "<|placeholder|>" * num_image_tokens, 1)
                 text[i] = text[i].replace("<|placeholder|>", self.image_token)
-
         return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)
         text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"])
         self._check_special_mm_tokens(text, text_inputs, modalities=["image", "video"])

{transformers-4.53.1 → transformers-4.53.2}/src/transformers/models/glm4v/video_processing_glm4v.py

@@ -246,10 +246,6 @@ class Glm4vVideoProcessor(BaseVideoProcessor):
         processed_grids = reorder_videos(processed_grids, grouped_videos_index)
         pixel_values_videos = torch.cat(processed_videos, dim=0)
         video_grid_thw = torch.tensor(processed_grids)
-        total_frames = video_grid_thw[0][0].item()
-        h = video_grid_thw[0][1].item()
-        w = video_grid_thw[0][2].item()
-        video_grid_thw = [[1, h, w] for _ in range(total_frames)]
         data = {
             "pixel_values_videos": pixel_values_videos,
             "video_grid_thw": video_grid_thw,

{transformers-4.53.1 → transformers-4.53.2/src/transformers.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: transformers
-Version: 4.53.1
+Version: 4.53.2
 Summary: State-of-the-art Machine Learning for JAX, PyTorch and TensorFlow
 Home-page: https://github.com/huggingface/transformers
 Author: The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/transformers/graphs/contributors)