make some ut cases pass on xpu w/ latest torch

[yao-matrix]

make below 12 cases pass on XPU w/ latest torch.

tests/models/llava/test_modeling_llava.py::LlavaForConditionalGenerationIntegrationTest::test_pixtral
tests/models/llava/test_modeling_llava.py::LlavaForConditionalGenerationIntegrationTest::test_pixtral_4bit
tests/models/llava/test_modeling_llava.py::LlavaForConditionalGenerationIntegrationTest::test_pixtral_batched
tests/models/aya_vision/test_modeling_aya_vision.py::AyaVisionIntegrationTest::test_small_model_integration_batched_generate
tests/models/gemma3/test_modeling_gemma3.py::Gemma3IntegrationTest::test_model_4b_batch
tests/models/gemma3/test_modeling_gemma3.py::Gemma3IntegrationTest::test_model_4b_batch_crops
tests/models/gemma3/test_modeling_gemma3.py::Gemma3IntegrationTest::test_model_4b_multiimage
tests/models/glm4v/test_processor_glm4v.py::Glm4vProcessorTest::test_apply_chat_template_video_frame_sampling
tests/models/internvl/test_modeling_internvl.py::InternVLLlamaIntegrationTest::test_llama_small_model_integration_batched_generate
tests/models/internvl/test_modeling_internvl.py::InternVLQwen2IntegrationTest::test_qwen2_small_model_integration_batched_generate
tests/models/llava_onevision/test_modeling_llava_onevision.py::LlavaOnevisionForConditionalGenerationIntegrationTest::test_small_model_integration_test_multi_image
tests/models/llava_onevision/test_modeling_llava_onevision.py::LlavaOnevisionForConditionalGenerationIntegrationTest::test_small_model_integration_test_multi_image_nested

@SunMarc @ydshieh , pls help review, thx very much.

[SunMarc]

Thanks ! Any idea what caused the outputs to change for xpu ?

[yao-matrix]

Thanks ! Any idea what caused the outputs to change for xpu ?

@SunMarc, XPU is still stabilizing the op implementations, Hugging Face, vLLM and other top-of-stack libraries are keeping report founded bugs or issues to them, and Intel PyTorch team is keeping fixing bugs and improving numerical behaviors, so the op implementations are keep enhancing rather than keeping still, which will impact model-level output. The good thing is: the output is more and more CUDA alike, which means the op enhancement is towards to CUDA-numerical-compatible way.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: aya_vision, gemma3, llava, llava_onevision

[SunMarc]

@bot /style

[github-actions]

Style bot fixed some files and pushed the changes.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.