fix msccl stream usage (#717)

[ROCm/rccl commit: c8e33b1232]
Tento commit je obsažen v:
Ziyue Yang
2023-03-25 01:59:36 +08:00
odevzdal GitHub
rodič 1025c41f7f
revize aa605b0989
+10 -2
Zobrazit soubor
@@ -375,7 +375,11 @@ ncclResult_t mscclEnqueueCheck(
switch (threadLocalStatus.groupStatus) {
case mscclNoGroup:
if (comm->mscclCompatible) {
CUDACHECK(hipStreamGetCaptureInfo(stream, &captureStatus, &pid));
if (stream == (hipStream_t)0) {
captureStatus = hipStreamCaptureStatusNone;
} else {
CUDACHECK(hipStreamGetCaptureInfo(stream, &captureStatus, &pid));
}
if (captureStatus == hipStreamCaptureStatusNone) {
NCCLCHECK(mscclSchedulerSelectAlgo(&threadLocalStatus.savedSchedulerParams.back()));
if (threadLocalStatus.savedSchedulerParams.back().p.scheduled) {
@@ -388,7 +392,11 @@ ncclResult_t mscclEnqueueCheck(
break;
case mscclGroupSupportedOp:
if (comm->mscclCompatible) {
CUDACHECK(hipStreamGetCaptureInfo(stream, &captureStatus, &pid));
if (stream == (hipStream_t)0) {
captureStatus = hipStreamCaptureStatusNone;
} else {
CUDACHECK(hipStreamGetCaptureInfo(stream, &captureStatus, &pid));
}
if (captureStatus == hipStreamCaptureStatusNone) {
NCCLCHECK(mscclSchedulerSelectAlgo(&threadLocalStatus.savedSchedulerParams.back()));
if (threadLocalStatus.savedSchedulerParams.back().p.scheduled) {