AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix MSCCL proxy number of chunks calculation (#821)

Browse Source 
Current number of transmissions parsed from MSCCL algorithm is 1-based value,
but when calculating proxy number of chunks, it's taken as 0-based value.
This commit fixes this issue.
This commit is contained in:
Ziyue Yang
2023-07-27 04:24:49 +08:00
committed by GitHub
parent 420f8af6a0
commit f7dc7b7e6a
2 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/misc/msccl/msccl_setup.cc
+2 -2
View File
@@ -118,7 +118,7 @@ ncclResult_t mscclSetupProxy(struct mscclAlgo* hostAlgo, ncclComm_t comm) {
int nRecvs = 0;
for (int j = 0; j < recvPeer->nExistingCounts; j++){
int c = recvPeer->existingCounts[j];
int nStepsInCount = DIVUP(c+1, status.maxAllowedCount);
int nStepsInCount = DIVUP(c, status.maxAllowedCount);
nRecvs += recvPeer->nTransmissionsOfCount[c] * nStepsInCount;
}
proxyOp.nsteps = nLoopsChunkSteps * nRecvs;
@@ -131,7 +131,7 @@ ncclResult_t mscclSetupProxy(struct mscclAlgo* hostAlgo, ncclComm_t comm) {
int nSends = 0;
for (int j = 0; j < sendPeer->nExistingCounts; j++){
int c = sendPeer->existingCounts[j];
int nStepsInCount = DIVUP(c+1, status.maxAllowedCount);
int nStepsInCount = DIVUP(c, status.maxAllowedCount);
nSends += sendPeer->nTransmissionsOfCount[c] * nStepsInCount;
}
proxyOp.nsteps = nLoopsChunkSteps * nSends;