H800/H100 fixes and tuning.
Re-enable intra-process direct pointer buffer access when CUMEM is
enabled.
This commit is contained in:
Sylvain Jeaugey
2023-09-26 05:56:27 -07:00
parent 3435178b6c
commit 8c6c595185
10 changed files with 96 additions and 74 deletions
+1 -1
View File
@@ -1174,7 +1174,7 @@ static ncclResult_t initTransportsRank(struct ncclComm* comm, struct ncclComm* p
int sendNode = (node+delta)%nNodes;
for (int step=0; step < steps; step++) {
int recvIndex = (localRank-step+steps)%steps;
int recvRank = recvIndex < nodeRanks[recvNode].localRanks ? nodeRanks[recvNode].localRankToRank[recvIndex] : -1;
int recvRank = recvIndex < nodeRanks[recvNode].localRanks ? nodeRanks[recvNode].localRankToRank[recvIndex] : -1;
tasks->p2pRecvOrder[i] = recvRank;
int sendIndex = (localRank+step)%steps;
int sendRank = sendIndex < nodeRanks[sendNode].localRanks ? nodeRanks[sendNode].localRankToRank[sendIndex] : -1;