Reapplying PR #1641 [AG and RS channel tuning] Add thread work threshold to tuning models and precompute reg index in LL128 (#1713)

* Reapply "[AG and RS channel tuning] Add thread work threshold to tuning models and precompute reg index in LL128 (#1641)"

This reverts commit 943ad6f7820739385a0b54e81f823d0df1dbf71c.

* Decreasing NCCL_LL128_SHMEM_ELEMS_PER_THREAD from 16 to 8
This commit is contained in:
Pedram Alizadeh
2025-06-04 13:22:11 -04:00
کامیت شده توسط GitHub
والد e94b360246
کامیت 3f7c08648f
7فایلهای تغییر یافته به همراه50 افزوده شده و 17 حذف شده
+1
مشاهده پرونده
@@ -1773,6 +1773,7 @@ static ncclResult_t topoGetAlgoInfo(
// NVLS should not need more than 16 channels to get peak BW.
nc = comm->nvlsChannels;
} else {
rcclUpdateThreadThreshold(comm, nBytes, info, threadThreshold);
// Ring/Tree channel tuning
while (nBytes < nc * nt * threadThreshold) {
if (nc >= 2) nc--;