[CI] Fix test suite names and add suite validation

[ispobock]

Motivation

Modifications

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.

[ispobock]

/tag-and-rerun-ci

[gemini-code-assist]

Code Review

This pull request renames the CPU test suite from 'stage-a-cpu-only' to 'stage-a-test-cpu' across several unit test files and introduces a validation mechanism in test/run_suite.py to ensure tests are registered to valid suites. Feedback was provided regarding the validation logic, noting that it currently uses a flat set of suite names which could allow a test to be incorrectly registered to a suite intended for a different hardware backend. It is recommended to refactor the validation to be backend-aware and to include AMD and NPU backends in the check.

[ispobock]

https://github.com/sgl-project/sglang/actions/runs/23933687537/job/69813982375?pr=21937#step:7:13701
test_fused_temperature_softmax.py test cannot pass, disable it first.
@Godmook @BBuf @DarkSharpness may have a check

[ispobock]

/rerun-stage stage-b-test-1-gpu-small

[github-actions]

✅ Triggered stage-b-test-1-gpu-small to run independently (skipping dependencies). View workflow run

[Godmook]

https://github.com/sgl-project/sglang/actions/runs/23933687537/job/69813982375?pr=21937#step:7:13701 test_fused_temperature_softmax.py test cannot pass, disable it first. @Godmook @BBuf @DarkSharpness may have a check

@ispobock @BBuf @DarkSharpness
The tolerance difference comes from intermediate precision: the reference path (div_ on bf16 tensor) truncates to bf16 before softmax, while the Triton kernel keeps fp32 throughout. This naturally produces ~2e-4 absolute / ~1e-2 relative difference on bf16 inputs.

Bumped to atol=5e-4, rtol=2e-2 that si above the observed max error and still tighter than most fused kernel tests in the repo (typically 1e-2 or looser). Does this range look reasonable to you?