[Perf] Tune FlashInfer all-reduce selection on SM103 (#53318)

Signed-off-by: Summer Yang <girasoleyang@gmail.com>
This commit is contained in:
Summer Yang
2026-08-24 05:19:59 +00:00
committed by GitHub
parent f94666b60d
commit e6e1af4ca1
4 changed files with 164 additions and 21 deletions
@@ -82,6 +82,14 @@ SYMM_MEM_ALL_REDUCE_MAX_SIZES = {
},
}
# Per-rank input limit for standalone FlashInfer MNNVL all-reduce.
# The key is (compute capability, world size, node count).
FI_MNNVL_ALLREDUCE_MAX_SIZE_MB: dict[tuple[int, int, int], float] = {
(103, 4, 1): 80,
(103, 8, 2): 64,
(103, 16, 4): 8,
}
# NCCL symmetric memory allreduce configuration based on H100 and GB200 benchmarks.
# PyNCCL-symm outperforms custom_AR for small and large tensor sizes,
# while custom_AR wins for mid-range sizes.