PyTorch PR #189459 fixes cuBLASLt tunable op test failure on GB200
Timing noise caused Default TunableOp to always win—now autotuning picks the fastest candidate.
A new PyTorch pull request (PR #189459) addresses a tricky test failure on GB200 hardware. The test `test_cublaslt_candidate_tunableop_cuda_bfloat16` in `test_linalg.py` was failing internally because the Default TunableOp candidate could win every test case due to timing noise. Since cuBLAS heuristics already select good candidates, the default option often appeared fastest, masking the real performance of alternative cuBLASLt candidates.
The PR refactors the test to confirm two key behaviors: first, that autotuning actually evaluates cuBLASLt candidates instead of just the default; second, that autotuning correctly selects the fastest candidate regardless of whether it's the Default. This ensures the TunableOp system properly benchmarks all options on CUDA, not just default. The change was authored with Codex and approved by eqy.
- Fixes internal test failure on GB200 hardware due to timing noise
- Ensures autotuning runs cuBLASLt candidates, not just Default TunableOp
- Validates that the fastest candidate is selected, Default or not
Why It Matters
Ensures Tensor core performance tuning accurately picks the fastest kernel, critical for large-scale AI workloads.