[Perf] Tune FlashInfer all-reduce thresholds for single-node TP8 on SM103 (#53606)
Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
Jee Jee Li committed
Aug 26, 2026 at 09:46 UTC
88e1b113131903936c4ee333a21d77deaacd50ee
3 files changed
+3
-2
tests/distributed/test_comm_ops.py
+1
-1
@@ -326,7 +326,7 @@ def test_cuda_communicator_checkpoints_flashinfer_workspaces(
326
("mnnvl", 103, 16, 4, 8 * flashinfer_all_reduce.MiB - 1),
327
("mnnvl", 103, 2, 1, None),
328
("mnnvl", 103, 12, 3, None),
329
- ("mnnvl", 103, 8, 1, None),
329
+ ("mnnvl", 103, 8, 1, 64 * flashinfer_all_reduce.MiB - 1),
330
("trtllm", 103, 8, 2, None),
331
("mnnvl", 90, 8, 2, None),
332
],
vllm/compilation/passes/fusion/allreduce_rms_fusion.py
+1
-1
@@ -116,7 +116,7 @@ FI_ALLREDUCE_FUSION_MAX_SIZE_MB: dict[int, dict[int, float]] = {
116
103: {
117
2: 64, # 64MB
118
4: 64, # 64MB
119
- 8: 2, # 2MB
119
+ 8: 4, # 4MB
120
16: 64, # 64MB (mnnvl multi-node)
121
},
122
107: {
vllm/distributed/device_communicators/all_reduce_utils.py
+1
@@ -86,6 +86,7 @@ SYMM_MEM_ALL_REDUCE_MAX_SIZES = {
86
# The key is (compute capability, world size, node count).
87
FI_MNNVL_ALLREDUCE_MAX_SIZE_MB: dict[tuple[int, int, int], float] = {
88
(103, 4, 1): 80,
89
+ (103, 8, 1): 64,
90
(103, 8, 2): 64,
91
(103, 16, 4): 8,
92
}