[CPU] Fix NUMA/core binding for DP ranks

[chunyuan-w]

Motivation

Before this PR, ModelRunner.init_threads_binding picked the NUMA node / core group for a worker using tp_rank. Under DP, DP rank 0 and DP rank 1 both resolve to tp_rank=0 and get bound to the same NUMA node/cores instead of being spread across the machine.
This PR fixes this incorrect binding.

Modifications

1. init_threads_binding now takes numa_index (the worker's global device id) and world_size = dp_size * tp_size * pp_size, instead of tp_rank and tp_size, so every rank across every DP replica maps to a distinct NUMA node / core group.

   Removed the world_size == len(threads_bind_list) equality assert for the explicit SGLANG_CPU_OMP_THREADS_BIND path, replacing it with only the per-rank bounds check. This permits the router-mode case (local world_size=1, multi-group bind string) while still catching a bind list that's too short.

2. test/registered/cpu/test_subblock_sparse_attention.py, added in [diffusion][Minimax H3]support subblock sparse attention on SM90 #34680, fails in Xeon CPU CI with ModuleNotFoundError: No module named 'imageio': importing it pulls in sglang.multimodal_gen, which needs diffusers/imageio/addict from the diffusion extra.

   Install .[diffusion] instead of . in docker/xeon.Dockerfile — the extra is already declared in python/pyproject_cpu.toml. Note this grows the CI image.

Accuracy Tests

Speed Tests and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

---

CI States

Latest PR Test (Base): ❌ Run #32679423999
Latest PR Test (Extra): ❌ Run #32679423888
Latest PR Test (AMD ROCm 7.2): ⏳ Run #32679424006

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[chunyuan-w]

/tag-run-ci-label

[chunyuan-w]

This change is needed to fix Xeon CI failure. Please check details in PR description (Modifications 2)

[mingfeima]

i am wondering whether does test/registered/cpu/test_subblock_sparse_attention.py really work?
this is registered under cpu tests, but uses cuda device.

maybe we should just remove this file.

[chunyuan-w]

I verified locally on a CPU-only machine and the test passes after installing diffusion dependencies. I can wait for the xeon CI to finish to double confirm this.

test_dispatch_is_resolved_once_per_device (__main__.TestSubBlockSparseAttentionDispatch) ... ok
test_dispatches_sm100 (__main__.TestSubBlockSparseAttentionDispatch) ... ok
test_rejects_unsupported_compute_capability (__main__.TestSubBlockSparseAttentionDispatch) ... ok

----------------------------------------------------------------------
Ran 3 tests in 0.002s

The test won't create tensor on CUDA, and it patches the torch.cuda.get_device_capability.
From what I understand, the test covers the dispatch logic of _get_subblock_sparse_attention_runner:
capability 9.0 → _sm90_sparse_attention, 10.0 → _sm100_sparse_attention, anything else → RuntimeError.

Shall we keep it and add the diffusion installation into Dockerfile?

[mingfeima]

yes, we are going to add it in futuer after all.