Processors do not pass on `return_tensors` to tokenizers properly.

[shuheng-liu]

System Info

• transformers version: 4.52.3
• Platform: macOS-15.4.1-arm64-arm-64bit-Mach-O
• Python version: 3.13.0
• Huggingface_hub version: 0.32.0
• Safetensors version: 0.5.3
• Accelerate version: not installed
• Accelerate config: not found
• DeepSpeed version: not installed
• PyTorch version (GPU?): 2.7.0 (False)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using distributed or parallel set-up in script?:

Who can help?

@ArthurZucker @zucchini-nlp

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

1. Install pytorch, pillow, and transformers=4.52.3 using pip.
2. Execute the following script:

import torch
from transformers import AutoProcessor
processor = AutoProcessor.from_pretrained("google/paligemma-3b-pt-224")
batch_features = processor(
    text="<image> What's in this image?",
    images=torch.zeros(3, 224, 224),
    suffix="Nothing",
    return_tensors="pt"
)

This yields an AttributeError with transformers==4.52.3

  File "/private/tmp/venv/lib/python3.13/site-packages/transformers/models/paligemma/processing_paligemma.py", line 313, in __call__
    labels = inputs["input_ids"].masked_fill(inputs["token_type_ids"] == 0, -100)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'list' object has no attribute 'masked_fill'

Expected behavior

The batch_features should be created without error.

There seems to be a recent bug in the __call__ method of many processors, including, e.g., PaliGemmaProcessor
This is likely caused by

transformers/src/transformers/models/paligemma/processing_paligemma.py

Lines 301 to 307 in 31f8a0f

	return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)
	inputs = self.tokenizer(
	input_strings,
	text_pair=suffix,
	return_token_type_ids=return_token_type_ids,
	**output_kwargs["text_kwargs"],
	)

which was changed in commit 32eca71

I believe the intention was to call .get() instead of .pop() on text_kwargs on line 301. Calling .pop() modifies text_kwargs in-place and hence the tokenizer would return inputs["input_ids"] as list instead of pytorch tensors. The masked_fill call below will fail when it's a list.

transformers/src/transformers/models/paligemma/processing_paligemma.py

Lines 312 to 313 in 31f8a0f

	if return_token_type_ids:
	labels = inputs["input_ids"].masked_fill(inputs["token_type_ids"] == 0, -100)

[zucchini-nlp]

Indeed, this happens only when a suffix is passed and apparently we didn't have tests for it 🙃 I'll make a fix, thanks for reporting