[XPU] Add diffusion server test config for Intel XPU B60

[Amrutha-M05]

Motivation

Let the multimodal_gen diffusion server tests run on Intel XPU with correct expectations.

The diffusion server tests compare generated images against committed ground-truth images and check stage timings against committed perf baselines. Both sets of reference values are CUDA-derived, and neither is reachable from XPU today:

• get_consistency_platform() has no XPU branch, so an XPU run silently falls through to the h100 thresholds.
• BASELINE_CONFIG composes per-platform perf overrides for Ascend and MUSA, but there is no XPU entry.

The image mismatch is not a correctness bug. XPU seeds its initial latent noise from a different RNG stream than CUDA, so the pixels cannot match bit-for-bit: the image stays semantically identical (CLIP remains high) while ssim / psnr / mean_abs_diff degrade. Comparing XPU output against CUDA thresholds therefore fails on a difference that carries no signal.

This PR adds XPU as a first-class platform in both registries, following the existing Ascend and MUSA pattern.

Modifications

Purely additive: 4 files, 56 insertions, 0 deletions.

• test/test_utils.py — register intel_xpu_b60 in CONSISTENCY_THRESHOLD_FILE_BY_PLATFORM, add the intelxpub60 alias, and add the current_platform.is_xpu() branch to get_consistency_platform().
• test/server/consistency_thresholds/intel_xpu_b60.json — XPU consistency thresholds for qwen_image_t2i_2_gpus and fsdp-inference.
• test/server/xpu/perf_baselines_xpu.json — XPU perf baselines for fsdp-inference, measured on Intel Arc Pro B60 (22.71 GiB per card).
• test/server/testcase_configs.py — chain the XPU baseline file into BASELINE_CONFIG, alongside the existing Ascend and MUSA .update() calls.

No CUDA path changes: both new files are read only when the resolved platform is XPU, and the two registry additions are new keys.

Thresholds

case	clip	ssim	psnr	mean_abs_diff
fsdp-inference	0.97	0.6	11.0	37.0
qwen_image_t2i_2_gpus	0.98	0.79	15.7	17.2

Only these two cases were measured on XPU hardware. Every other case still inherits from h100.json; that is called out in the _comment field of the threshold file so the remaining cases can be triaged the same way rather than being assumed correct.

Accuracy Tests

Verified on 2x Intel Arc Pro B60 (22.7 GiB per card):

python/sglang/multimodal_gen/test/server/test_server_2_gpu.py::TestDiffusionServerTwoGpu::test_diffusion_generation[fsdp-inference]

Model Tongyi-MAI/Z-Image-Turbo. Passes end to end in ~157 s, including the consistency check against the thresholds above. Measured CLIP 0.9779 (passes); SSIM 0.6609 / PSNR 12.0434 / mean-abs-diff 33.3489, which fail the H100 thresholds and pass the XPU ones. XPU output is deterministic across 4 runs (identical md5).

The perf baselines are single-host measurements, not an averaged multi-run distribution, so they are starting reference values rather than tight regression bounds.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests. — no new test case is introduced; this PR adds the platform configuration that existing diffusion server tests read on XPU.
• Update documentation according to Write documentations. — N/A, no user-facing behaviour change.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

cc: @siju-samuel @rbabukv

---

CI States

Latest PR Test (Base): ❌ Run #33606366411
Latest PR Test (Extra): ❌ Run #33606364529
Latest PR Test (AMD ROCm 7.2): ❌ Run #33606366341

[siju-samuel]

/tag-run-ci-label

[Amrutha-M05]

Closing, as #40664 merged this PR's changes into main.