[icavan]

📌 Description

Reduce the default cuLA test-suite wall time while preserving representative coverage for each CUDA kernel family.

• add cula_fast, cula_slow, and cula_full test profiles while retaining compatibility with the existing KDA profiles
• keep representative dense/varlen, state/no-state, MHA/GVA, boundary-size, and architecture-specific cases in the default suite
• move broad CUDA parameter sweeps behind the explicit slow/full profiles
• collapse redundant Cartesian products in recurrence, packed decode, and decode coverage
• trim intracard-CP stress repetitions from 100 to 10 and remove redundant long-sequence combinations
• remove GVA cases from reference suites whose FLA/naive implementations require H == HV; dedicated decode/conv tests continue to cover GVA
• document the default and extended test commands

🔍 Related Issues

No linked issue.

🚀 Pull Request Checklist

✅ Pre-commit Checks

• I have installed pre-commit.
• I have installed the hooks.
• I have run pre-commit run --all-files and fixed any reported issues.

🧪 Tests

• Tests have been added or updated as needed.
• All default tests are passing.

Validation on aistudio-58650011-ssctl, one GB200, cold XDG_CACHE_HOME, rebased onto current origin/main:

101 passed, 808 skipped in 232.14s (0:03:52)
real 235.485s

Additional checks:

tests/test_test_suite_profile.py: 5 passed
pre-commit run --all-files: passed

⚡ Performance

• Before: more than 32 minutes wall time even with four-way GPU sharding; the slowest shard took 32:46.
• After: 3:55 wall time on one GB200 with a cold compilation cache.

Reviewer Notes

The default pytest -q run is time-bounded and representative. Use pytest -m cula_slow -q for extended sweeps and pytest -m cula_full -q for all retained cases. The cula_slow alias also includes the legacy kda_slow cases.

[zheyang0825]

LGTM

[Longxmas]

LGTM