Fix arguments (#40605)

* Fix invalid arguments

Signed-off-by: cyy <[email protected]>

* Fix typing

Signed-off-by: cyy <[email protected]>

* Add missing self

Signed-off-by: cyy <[email protected]>

* Add missing self and other fixes

Signed-off-by: cyy <[email protected]>

*  More fixes

Signed-off-by: cyy <[email protected]>

*  More fixes

Signed-off-by: cyy <[email protected]>

---------

Signed-off-by: cyy <[email protected]>

Author: cyyever

src/transformers/models/evolla/modeling_evolla.py

@@ -637,7 +637,11 @@ def forward(
         )
 
     def get_extended_attention_mask(
-        self, attention_mask: Tensor, input_shape: tuple[int], device: torch.device = None, dtype: torch.float = None
+        self,
+        attention_mask: Tensor,
+        input_shape: tuple[int],
+        device: torch.device = None,
+        dtype: Optional[torch.dtype] = None,
     ) -> Tensor:
         """
         Makes broadcastable attention and causal masks so that future and masked tokens are ignored.

src/transformers/models/evolla/modular_evolla.py

@@ -269,7 +269,11 @@ def forward(
         )
 
     def get_extended_attention_mask(
-        self, attention_mask: Tensor, input_shape: tuple[int], device: torch.device = None, dtype: torch.float = None
+        self,
+        attention_mask: Tensor,
+        input_shape: tuple[int],
+        device: torch.device = None,
+        dtype: Optional[torch.dtype] = None,
     ) -> Tensor:
         """
         Makes broadcastable attention and causal masks so that future and masked tokens are ignored.

src/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py

@@ -1565,13 +1565,12 @@ def prepare_inputs_for_generation(
             # 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
+            # models currently cannot do assisted decoding
             if cache_position[0] == 0 or self.model.rope_deltas is None:
                 vision_positions, rope_deltas = self.model.get_rope_index(
                     model_inputs.get("input_ids", None),
                     image_grid_thw=image_grid_thw,
                     video_grid_thw=video_grid_thw,
-                    second_per_grid_ts=second_per_grid_ts,
                     attention_mask=attention_mask,
                 )
                 self.model.rope_deltas = rope_deltas

src/transformers/models/qwen2_5_vl/modular_qwen2_5_vl.py

@@ -809,13 +809,12 @@ def prepare_inputs_for_generation(
             # 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
+            # models currently cannot do assisted decoding
             if cache_position[0] == 0 or self.model.rope_deltas is None:
                 vision_positions, rope_deltas = self.model.get_rope_index(
                     model_inputs.get("input_ids", None),
                     image_grid_thw=image_grid_thw,
                     video_grid_thw=video_grid_thw,
-                    second_per_grid_ts=second_per_grid_ts,
                     attention_mask=attention_mask,
                 )
                 self.model.rope_deltas = rope_deltas

src/transformers/models/sam2/modular_sam2.py

@@ -108,13 +108,13 @@ class Sam2ImageProcessorFast(SamImageProcessorFast):
     def __init__(self, **kwargs: Unpack[Sam2FastImageProcessorKwargs]):
         BaseImageProcessorFast.__init__(self, **kwargs)
 
-    def pad_image():
+    def pad_image(self):
         raise NotImplementedError("No pad_image for SAM 2.")
 
-    def _get_preprocess_shape():
+    def _get_preprocess_shape(self):
         raise NotImplementedError("No _get_preprocess_shape for SAM 2.")
 
-    def resize():
+    def resize(self):
         raise NotImplementedError("No need to override resize for SAM 2.")
 
     def _preprocess(

src/transformers/models/wavlm/modeling_wavlm.py

@@ -191,7 +191,7 @@ def torch_multi_head_self_attention(
         attention_mask: Union[torch.LongTensor, torch.BoolTensor],
         gated_position_bias: torch.FloatTensor,
         output_attentions: bool,
-    ) -> (torch.FloatTensor, torch.FloatTensor):
+    ) -> tuple[torch.FloatTensor, torch.FloatTensor]:
         """simple wrapper around torch's multi_head_attention_forward function"""
         # self-attention assumes q = k = v
         query = key = value = hidden_states.transpose(0, 1)

src/transformers/models/wavlm/modular_wavlm.py

@@ -122,7 +122,7 @@ def torch_multi_head_self_attention(
         attention_mask: Union[torch.LongTensor, torch.BoolTensor],
         gated_position_bias: torch.FloatTensor,
         output_attentions: bool,
-    ) -> (torch.FloatTensor, torch.FloatTensor):
+    ) -> tuple[torch.FloatTensor, torch.FloatTensor]:
         """simple wrapper around torch's multi_head_attention_forward function"""
         # self-attention assumes q = k = v
         query = key = value = hidden_states.transpose(0, 1)

src/transformers/models/yolos/modular_yolos.py

@@ -171,22 +171,22 @@ def post_process_object_detection(
 
         return results
 
-    def post_process_segmentation():
+    def post_process_segmentation(self):
         raise NotImplementedError("Segmentation post-processing is not implemented for Deformable DETR yet.")
 
-    def post_process_instance():
+    def post_process_instance(self):
         raise NotImplementedError("Instance post-processing is not implemented for Deformable DETR yet.")
 
-    def post_process_panoptic():
+    def post_process_panoptic(self):
         raise NotImplementedError("Panoptic post-processing is not implemented for Deformable DETR yet.")
 
-    def post_process_instance_segmentation():
+    def post_process_instance_segmentation(self):
         raise NotImplementedError("Segmentation post-processing is not implemented for Deformable DETR yet.")
 
-    def post_process_semantic_segmentation():
+    def post_process_semantic_segmentation(self):
         raise NotImplementedError("Semantic segmentation post-processing is not implemented for Deformable DETR yet.")
 
-    def post_process_panoptic_segmentation():
+    def post_process_panoptic_segmentation(self):
         raise NotImplementedError("Panoptic segmentation post-processing is not implemented for Deformable DETR yet.")
 
 

src/transformers/trainer.py

@@ -294,9 +294,6 @@ def safe_globals():
 if TYPE_CHECKING:
     import optuna
 
-    if is_datasets_available():
-        import datasets
-
 logger = logging.get_logger(__name__)
 
 
@@ -418,14 +415,14 @@ class Trainer:
     def __init__(
         self,
         model: Union[PreTrainedModel, nn.Module, None] = None,
-        args: TrainingArguments = None,
+        args: Optional[TrainingArguments] = None,
         data_collator: Optional[DataCollator] = None,
         train_dataset: Optional[Union[Dataset, IterableDataset, "datasets.Dataset"]] = None,
         eval_dataset: Optional[Union[Dataset, dict[str, Dataset], "datasets.Dataset"]] = None,
         processing_class: Optional[
             Union[PreTrainedTokenizerBase, BaseImageProcessor, FeatureExtractionMixin, ProcessorMixin]
         ] = None,
-        model_init: Optional[Callable[[], PreTrainedModel]] = None,
+        model_init: Optional[Callable[..., PreTrainedModel]] = None,
         compute_loss_func: Optional[Callable] = None,
         compute_metrics: Optional[Callable[[EvalPrediction], dict]] = None,
         callbacks: Optional[list[TrainerCallback]] = None,

src/transformers/video_processing_utils.py

@@ -681,7 +681,7 @@ def get_video_processor_dict(
         else:
             video_processor_file = VIDEO_PROCESSOR_NAME
             try:
-                # Try to load with a new config name first and if not successfull try with the old file name
+                # Try to load with a new config name first and if not successful try with the old file name
                 # NOTE: we will gradually change to saving all processor configs as nested dict in PROCESSOR_NAME
                 resolved_video_processor_files = cached_files(
                     pretrained_model_name_or_path,