Fix MSCCLPP accuracy issue for allreduce7 (#1634)

* ext-src: fix a graph-mode bug in allreduce7

* change MSCCLPP threshold to 16MB

* ext-src: change message size threshold for allreduce7

* ext-src: address review comments
This commit is contained in:
Nusrat Islam authored and GitHub committed 2025-04-18 08:54:32 -05:00
1 förälder 7a84c5dbb0
incheckning f20c33effd
3 files changed
+246 -38

No files matched your search

+1 -1
Visa fil
@@ -114,7 +114,7 @@ bool operator ==(const ncclUniqueId& a, const ncclUniqueId& b) {
return memcmp(a.internal, b.internal, NCCL_UNIQUE_ID_BYTES) == 0;
}
RCCL_PARAM(MscclppThreshold, "MSCCLPP_THRESHOLD", (size_t)(1024*1024));
RCCL_PARAM(MscclppThreshold, "MSCCLPP_THRESHOLD", (size_t)(16*1024*1024));
#endif
static constexpr int64_t defaultEnableMscclpp = 0;