Simplify test_chunked_prefill; remove redundant tests

[hnyls2002]

Summary

Remove redundant tests in test_chunked_prefill.py to reduce CI time:

• test_chunked_prefill — default path (radix_cache=True, mixed_chunk=False), implicitly covered by all other MMLU integration tests
• test_chunked_prefill_without_radix_cache — disable_radix_cache path is already covered by test_no_overlap_scheduler.py
• test_mixed_chunked_prefill_multi_requests — no accuracy assertion (just response.json()); chunked_prefill_size=2048 rarely triggers chunking on short prompts

Keep:

• test_mixed_chunked_prefill — unique coverage of enable_mixed_chunk=True on default path
• test_mixed_chunked_prefill_without_radix_cache — unique coverage of mixed_chunk + disable_radix_cache combo

Test plan

• CI: test_chunked_prefill.py passes with remaining 2 tests
• May need to adjust est_time (currently 360) after CI run

[gemini-code-assist]

Code Review

This pull request simplifies the chunked prefill test suite by removing several test cases, including those for multi-request scenarios. Feedback indicates that the remaining tests should explicitly define "chunked_prefill_size" to ensure the chunked prefill logic is properly exercised. Furthermore, it is recommended to restore and adjust the multi-request test case rather than removing it, as it provides critical coverage for concurrent request scheduling.

I am having trouble creating individual review comments. Click here to see my feedback.

test/registered/scheduler/test_chunked_prefill.py (16)

The test test_mixed_chunked_prefill does not explicitly set chunked_prefill_size. In the scheduler implementation, is_mixed_chunk is only enabled if chunked_prefill_size is not None. Without this parameter, the test may not be exercising the chunked prefill logic. Consider explicitly passing a chunked_prefill_size (e.g., 128) to ensure the intended code path is tested.

        run_mmlu_test(disable_radix_cache=False, enable_mixed_chunk=True, chunked_prefill_size=128)

test/registered/scheduler/test_chunked_prefill.py (19)

Similar to the test above, chunked_prefill_size should be explicitly set here to ensure that chunked prefill is actually active during the test.

        run_mmlu_test(disable_radix_cache=True, enable_mixed_chunk=True, chunked_prefill_size=128)

test/registered/scheduler/test_chunked_prefill.py (27-31)

Removing test_mixed_chunked_prefill_multi_requests eliminates coverage for concurrent requests with chunked prefill. While the current implementation might not trigger chunking frequently with a size of 2048 on short prompts, it is better to fix the test by reducing the chunked_prefill_size (e.g., to 128) rather than removing it. This provides valuable coverage for the scheduler's state management under load, which sequential MMLU tests may miss.

[hnyls2002]

/rerun-test test_mixed_chunked_prefill.py

[github-actions]

✅ 1-gpu-5090 (1 test): View workflow run

cd test/ && python3 registered/scheduler/test_mixed_chunked_prefill.py

[hnyls2002]

/tag-and-rerun-ci