[Perf] [KDA] cookbook: enable custom all-reduce for GB300 BF16 Muse Spark (#45429)

Co-authored-by: kernel-design-agents <kernel-design-agents@users.noreply.github.com>
This commit is contained in:
Song Bian
2026-10-22 20:46:28 +08:00
committed by GitHub
co-authored by kernel-design-agents
parent e43f6128a0
commit 0ec3fee429
@@ -390,7 +390,9 @@ export const config = {
{
match: { hw: "gb300", quant: "bf16" },
nnodes: 2,
env: [],
// Expandable segments admit the two-node NVL72 TP8 group to custom
// all-reduce v2; without them decode all-reduce falls back to NCCL.
env: ["PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True"],
flags: [
"--model-path {{MODEL_NAME}}",
"--tp-size 8",