[jiqing-feng]

Fixes 13 slow integration tests that were failing on Intel XPU. Verified on 8x Intel Arc Pro B60 (get_device_properties() returns ("xpu", 5)), torch 2.14.0+xpu: 22/22 pass, 9/22 before.

Expected values for XPU gen-5

These tests already had expectations for other devices (and in a few cases for gen-3 Arc, whose values differ from B60). Added an ("xpu", 5) entry; no existing key was modified, so no other device is affected.

• emu3 — test_model_generation_batched (needs torch >= 2.16 nightly; on 2.14/2.15 the fp32 SDPA kernel hits a device-side fault at this sequence length)
• ernie4_5_vl_moe — test_small_model_integration_test_batch
• falcon_h1 — test_falcon_h1_hard
• fast_vlm — test_model_integration
• internvl — test_llama_small_model_integration_generate
• kosmos2_5 — test_sdpa
• llava — test_pixtral_batched
• llava_next_video — test_small_model_integration_test_batch_matches_single
• llava_onevision — test_small_model_integration_test_multi_image
• qwen2_5_vl — test_small_model_integration_test_with_video
• voxtral_realtime — test_batched_longform
• phi4_multimodal — test_vision_text_generation: the existing ("xpu", 5) value was recorded on different XPU hardware and does not reproduce on B60; updated to the B60 output.

Non-deterministic tests

Two tests produced a different output on nearly every run on XPU. Added the existing @require_deterministic_for_xpu decorator (already used by higgs_audio_v2 and elsewhere in tests/) to make them reproducible, then recorded the resulting stable output.

• gemma3n — test_model_4b_bf16: three runs without the decorator gave three different outputs; three runs with it gave identical output. Note that ROCm already expects turquoise here instead of CUDA's clear blue, so this token is a known cross-device divergence point rather than an XPU defect.
• deepseek_v2 — test_batch_fa2: the model is sharded with device_map="auto" and the test asserted a single hardcoded list. Converted to Expectations and added the decorator.

deepseek_v3::test_eager_matches_sdpa_generate

This test overrides the common one with a strict torch.testing.assert_close on token ids. The tiny fp16 checkpoint produces exactly tied top-2 logits for one of the prompts (eager sees 17.0 vs 16.98438, sdpa sees 17.0 vs 17.0), so argmax tie-breaking is arbitrary even though both kernels agree to within a single fp16 ULP — over non-pad positions max|eager - sdpa| is 0.015625, exactly 1 ULP at that magnitude. One flipped token at position 14 then cascades through the remaining 30. The large raw difference seen before masking comes entirely from fully-masked pad rows, which never reach the output.

This is not a new relaxation: the common test_eager_matches_sdpa_generate in GenerationTesterMixin already compares with assert_similar_generate_outputs, which accepts a token mismatch when the scores at the first mismatch are numerically equivalent. The override had opted out of it. Switched back to the shared helper (return_dict_in_generate / output_scores are needed to give it the scores), with atol sized for fp16 rather than the fp32-oriented 1e-3 default — a genuine kernel discrepancy is far above that. Result: 8/8 passes, vs 3/6 before.

Testing

RUN_SLOW=1 pytest tests/models/{deepseek_v2,deepseek_v3,ernie4_5_vl_moe,falcon_h1,fast_vlm,gemma3n,internvl,kosmos2_5,llava,llava_next_video,llava_onevision,phi4_multimodal,qwen2_5_vl,voxtral_realtime}

[github-actions]

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

run-slow: deepseek_v2, deepseek_v3, emu3, ernie4_5_vl_moe, falcon_h1, fast_vlm, gemma3n, internvl, kosmos2_5, llava, llava_next_video, llava_onevision, phi4_multimodal, qwen2_5_vl, voxtral_realtime