[v0.5.10][3] Fix processor return_tensors duplicate kwarg for transformers >=5.0

[yueming-yuan]

ci-sglang-pr: sglang-miles-v0.5.10

[gemini-code-assist]

Code Review

This pull request updates the build_processor_kwargs function in miles/utils/processing_utils.py to move the return_tensors parameter from the top-level dictionary into modality-specific sub-dictionaries, specifically text_kwargs. This change is intended to prevent duplicate keyword argument errors in transformers >= 5.0. Feedback suggests explicitly removing any top-level return_tensors key from the result dictionary to ensure no conflicts occur if the key was included in the original input.

[gemini-code-assist]

To fully resolve the duplicate keyword argument issue in transformers >= 5.0, any existing top-level return_tensors should be removed from the result dictionary. Since the logic now explicitly sets return_tensors within modality-specific dictionaries (like text_kwargs), leaving it at the top level—if it was provided in the input multimodal_inputs—will still trigger a ValueError due to the conflict between the global and nested settings.

    result = dict(multimodal_inputs) if multimodal_inputs else {}
    result.pop("return_tensors", None)

    # return_tensors=None for text (input_ids), "pt" for modality-specific outputs.
    # Use per-modality dicts to avoid transformers >=5.0 duplicate kwarg error.
    result["text_kwargs"] = {**result.get("text_kwargs", {}), "return_tensors": None}

[maocheng23]

This makes sense to me

[maocheng23]

LGTM but nit

[maocheng23]

This makes sense to me

[maocheng23]

Suggested change

	result["text_kwargs"] = {**result.get("text_kwargs", {}), "return_tensors": None}
	result.pop("return_tensors", None)
	result["text_kwargs"] = {**result.get("text_kwargs", {}), "return_tensors": None}