Processors: don't default padding side (huggingface#33942)

* don't default padding side * fix
ylacombe · Oct 8, 2024 · 0dbc709 · 0dbc709
1 parent a3add29
commit 0dbc709
Show file tree

Hide file tree

Showing 4 changed files with 6 additions and 7 deletions.
diff --git a/src/transformers/models/idefics3/processing_idefics3.py b/src/transformers/models/idefics3/processing_idefics3.py
@@ -235,9 +235,6 @@ def __call__(
             **kwargs,
         )
 
-        # Temporary fix for "padding_side" in init_kwargs
-        output_kwargs["text_kwargs"].pop("padding_side", None)
-
         image_seq_len = image_seq_len if image_seq_len is not None else self.image_seq_len
 
         n_images_in_text = []

diff --git a/src/transformers/models/llava_onevision/processing_llava_onevision.py b/src/transformers/models/llava_onevision/processing_llava_onevision.py
@@ -172,8 +172,6 @@ def __call__(
             num_video_tokens = (num_frames * pooled_height_width * pooled_height_width) + 1  # +1 for newline token
             text = [sample.replace(self.video_token, self.video_token * num_video_tokens) for sample in text]
 
-        # Padding side can be in TextKwargs but is not accepted by the tokenizer
-        _ = output_kwargs["text_kwargs"].pop("padding_side", None)
         text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"])
         return BatchFeature(data={**text_inputs, **image_inputs, **video_inputs})
 

diff --git a/src/transformers/models/qwen2_vl/processing_qwen2_vl.py b/src/transformers/models/qwen2_vl/processing_qwen2_vl.py
@@ -150,7 +150,6 @@ def __call__(
                     index += 1
                 text[i] = text[i].replace("<|placeholder|>", "<|video_pad|>")
 
-        _ = output_kwargs["text_kwargs"].pop("padding_side", None)
         text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"])
 
         return BatchFeature(data={**text_inputs, **image_inputs, **videos_inputs})

diff --git a/src/transformers/processing_utils.py b/src/transformers/processing_utils.py
@@ -829,7 +829,12 @@ class MyProcessingKwargs(ProcessingKwargs, CommonKwargs, TextKwargs, ImagesKwarg
             for modality_key in ModelProcessorKwargs.__annotations__[modality].__annotations__.keys():
                 # init with tokenizer init kwargs if necessary
                 if modality_key in tokenizer_init_kwargs:
-                    default_kwargs[modality][modality_key] = tokenizer_init_kwargs[modality_key]
+                    value = (
+                        getattr(self.tokenizer, modality_key)
+                        if hasattr(self.tokenizer, modality_key)
+                        else tokenizer_init_kwargs[modality_key]
+                    )
+                    default_kwargs[modality][modality_key] = value
         # now defaults kwargs are updated with the tokenizers defaults.
         # pass defaults to output dictionary
         output_kwargs.update(default_kwargs)