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

Fix protocol and channel override when tuner is used (#1985)

Browse Source 
* Fix protocol and channel override when tuner is used

* Added comment

* Fix README for basic tuner implementation
This commit is contained in:
nawrinsu
2025-11-03 13:56:34 -08:00
committed by GitHub
parent caffd013f6
commit 166268d715
4 changed files with 11 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/enqueue.cc
+7 -1
View File
@@ -2055,6 +2055,9 @@ static ncclResult_t topoGetAlgoInfo(
info->protocol = protocol;
float time = minTime;
// Tuner plugin sets cost to 0.0 if it finds a match
bool isTunerMatchFound = (comm->tuner != NULL && minTime == 0.0);
// Yes, we are first assigning and then testing if protocol is sane, but that's OK in this case.
// coverity[check_after_sink]
if (info->algorithm == NCCL_ALGO_UNDEF || info->protocol == NCCL_PROTO_UNDEF) {
@@ -2071,7 +2074,10 @@ static ncclResult_t topoGetAlgoInfo(
WARN("Error : no algorithm/protocol available for function %s with datatype %s.%s%s", ncclFuncToString(info->func), ncclDatatypeToString(info->datatype), ncclAlgoEnvStr, ncclProtoEnvStr);
return (algoEnv || protoEnv) ? ncclInvalidUsage : ncclInternalError;
}
rcclUpdateCollectiveProtocol(comm, nBytes, info);
// Honor Tuner config if available
if (!isTunerMatchFound) {
rcclUpdateCollectiveProtocol(comm, nBytes, info);
}
rcclSetPipelining(comm, nBytes, info);
if (simInfo) simInfo->estimatedTime = time;
TRACE(NCCL_COLL, "%ld Bytes -> Algo %d proto %d time %f", nBytes, info->algorithm, info->protocol, time);