CI: synchronize prefill graph test fixtures

[mickqian]

Summary

Synchronize the lightweight PrefillCudaGraphRunner test fixtures with the state now read by can_run_graph():

• initialize capture_num_tokens in the multimodal piecewise CUDA graph fixture;
• initialize prefill_backend_name and has_mha_companion_layers in the prefill-padding fixture.

Root cause

#31391 and #31487 were developed from the same older main revision and each added runner state plus a lightweight fixture built with __new__():

• Enable Kimi multimodal breakable prefill CUDA graph replay #31391 added the backend/MHA-prefix state and multimodal graph tests;
• perf: avoid excessive prefill CUDA graph padding #31487 added the capture-bucket padding guard and padding tests.

Each PR's CPU tests passed in isolation because its fixture only needed the state present on that shared base. After both changes landed, can_run_graph() read both sets of fields, exposing symmetric fixture drift: the multimodal fixture lacked capture_num_tokens, while the padding fixture lacked the backend/MHA state.

The fixtures now explicitly model tc_piecewise behavior and the [4, 16] capture buckets used by these tests.

Impact

Test-only; runtime behavior is unchanged. This unblocks the affected base-a-test-cpu shard on current main and dependent PRs.

Validation

• pre-commit run --files test/registered/unit/configs/test_multimodal_piecewise_cuda_graph.py test/registered/unit/model_executor/runner/test_prefill_cuda_graph_padding.py
• git diff --check
• targeted runtime test deferred to CI because the local macOS environment lacks the SGLang Python dependencies (numpy is unavailable)

---

CI States

Latest PR Test (Base): ⏳ Run #29641154932
Latest PR Test (Extra): ⏳ Run #29641154868

[gemini-code-assist]

Code Review

This pull request updates the _make_prefill_runner method in test_multimodal_piecewise_cuda_graph.py by setting the capture_num_tokens attribute of the runner to [4, 16]. There are no review comments, and I have no additional feedback to provide.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[mickqian]

/tag-and-rerun-ci