[CI] Fix PaddleOCR-VL HF test failure due to create_causal_mask API rename

[AndreasKaratzas]

• The PaddleOCR-VL HF model calls create_causal_mask(inputs_embeds=...) but transformers renamed the parameter to input_embeds (singular), causing a TypeError in the HF runner
• Adds paddleocr_vl_patch_hf_runner that monkey-patches create_causal_mask in the model's module to accept the old parameter name

Test plan

• pytest -s -v tests/models/multimodal/generation/test_common.py::test_single_image_models[paddleocr_vl-test_case34]

cc @kenroche

[gemini-code-assist]

Code Review

This pull request fixes a TypeError in the PaddleOCR-VL HF test by monkey-patching the create_causal_mask function to support a renamed parameter. The fix is correct, but the implementation of the monkey-patch has a potential issue with test isolation as it modifies a global module without reverting the change. I've added a comment with a suggestion for a more robust implementation to prevent side effects on other tests.

[gemini-code-assist]

This monkey-patch modifies a module in sys.modules, which is a global state. This can break test isolation and cause side effects in other tests running in the same process. For instance, a different test might rely on the original behavior of create_causal_mask and would fail in an unexpected way.

To ensure test reliability, this patch should be reverted after the test finishes. The standard way to achieve this is by using a context manager or a pytest fixture that handles the setup and teardown of the patch.

A possible approach would be to refactor this into a context manager that can be used within the test execution logic:

from contextlib import contextmanager

@contextmanager
def patched_create_causal_mask_for_paddleocr_vl(hf_model: HfRunner):
    import sys
    model_module = sys.modules.get(type(hf_model.model.model).__module__)
    
    original_create_causal_mask = getattr(model_module, "create_causal_mask", None)
    if original_create_causal_mask is None:
        yield
        return

    def patched_create_causal_mask(*args, **kwargs):
        if "inputs_embeds" in kwargs:
            kwargs["input_embeds"] = kwargs.pop("inputs_embeds")
        return original_create_causal_mask(*args, **kwargs)

    model_module.create_causal_mask = patched_create_causal_mask
    try:
        yield
    finally:
        model_module.create_causal_mask = original_create_causal_mask

This would require changes in the test runner to use the context manager, but it would make the test suite more robust.

[AndreasKaratzas]

Global state is technically valid but not practically relevant here. The module itself is test-specific and only renames inputs_embeds.