[Bug]: Order-dependent mismatch in test_nvfp4_gather_grouped_gemm_act_fusion_rubin on SM107

Summary

tests/unittest/_torch/thop/parallel/test_cute_dsl_moe.py::test_nvfp4_gather_grouped_gemm_act_fusion_rubin[1024-1-32-128-relu2] fails on SM107 only when the whole file runs in one pytest process:

Mismatch percentage 0.05% for rtol 1e-4

It passes 3/3 in isolation and when run with only its sibling params. Reproduced in two separate full-file batches on 4x SM107 against main.

Notes

Order dependence points at state carried across tests (autotuner cache, workspace reuse, or a missing synchronization in the Rubin gather/act-fusion grouped GEMM). It may share a root cause with #18335 (bulk-async reduction not awaited before sC reuse in the Rubin fused-MoE finalize kernel), which shows up as an order-dependent illegal memory access in the BF16 finalize path.

Ask

Root-cause with compute-sanitizer --tool memcheck / racecheck on the full file, and add a per-test reset or the missing sync.