[Perf] Tune FlashInfer all-reduce thresholds for single-node TP8 on SM103 (#53606)

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
This commit is contained in:
Jee Jee Li
2026-08-26 01:46:56 +00:00
committed by GitHub
parent d125b540b7
commit 88e1b11313
3 changed files with 3 additions and 2 deletions
+1 -1
View File
@@ -326,7 +326,7 @@ def test_cuda_communicator_checkpoints_flashinfer_workspaces(
("mnnvl", 103, 16, 4, 8 * flashinfer_all_reduce.MiB - 1),
("mnnvl", 103, 2, 1, None),
("mnnvl", 103, 12, 3, None),
("mnnvl", 103, 8, 1, None),
("mnnvl", 103, 8, 1, 64 * flashinfer_all_reduce.MiB - 1),
("trtllm", 103, 8, 2, None),
("mnnvl", 90, 8, 2, None),
],
@@ -116,7 +116,7 @@ FI_ALLREDUCE_FUSION_MAX_SIZE_MB: dict[int, dict[int, float]] = {
103: {
2: 64, # 64MB
4: 64, # 64MB
8: 2, # 2MB
8: 4, # 4MB
16: 64, # 64MB (mnnvl multi-node)
},
107: {
@@ -86,6 +86,7 @@ SYMM_MEM_ALL_REDUCE_MAX_SIZES = {
# 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, 1): 64,
(103, 8, 2): 64,
(103, 16, 4): 8,
}