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

misc/msccl: force use of mscclpp (#1581)

Browse Source 
[ROCm/rccl commit: ac823818aa]
This commit is contained in:
Nusrat Islam
2025-03-04 12:48:59 -06:00
committed by GitHub
parent d1247bbf2a
commit e7c90e0a46
3 changed files with 21 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/misc/msccl/msccl_lifecycle.cc
+14 -10
View File
@@ -528,22 +528,24 @@ ncclResult_t mscclEnqueueCheck(
const bool buffsRegistered = sendBuffRegistered && recvBuffRegistered;
/* check if one rank per GPU and graph mode is enabled */
if ((graphMode || buffsRegistered) && comm->mscclCompatible && nBytes > 0 && (nBytes & 31) == 0) {
if ((graphMode || buffsRegistered || comm->mscclppForceEnable) && comm->mscclCompatible && nBytes > 0 && (nBytes & 31) == 0) {
bool isManagedBuffer = false;
if (sendBuff) CUDACHECK(hipPointerGetAttribute(&isManagedBuffer, HIP_POINTER_ATTRIBUTE_IS_MANAGED, const_cast<void*>(sendBuff)));
if (!isManagedBuffer && recvBuff) CUDACHECK(hipPointerGetAttribute(&isManagedBuffer, HIP_POINTER_ATTRIBUTE_IS_MANAGED, const_cast<void*>(recvBuff)));
if (isManagedBuffer) { /* MSCCL++ not enabled for managed memory buffers */ }
else if (func == mscclFuncAllReduce && nBytes <= comm->mscclpp_threshold && isMscclppAllReduceSupported(dataType, op)) {
INFO(NCCL_COLL,"%s: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p graphMode %d",
"mscclpp_ncclAllReduce", comm->opCount, sendBuff, recvBuff, count, dataType, op, root, comm, comm->nRanks, stream, graphMode);
INFO(NCCL_COLL,"%s: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p graphMode %d ubr %d",
"mscclpp_ncclAllReduce", comm->opCount, sendBuff, recvBuff, count, dataType, op, root, comm, comm->nRanks, stream,
graphMode, buffsRegistered);
NCCLCHECK(mscclpp_ncclAllReduce(sendBuff, recvBuff, count, dataType, op, comm->mscclpp_comm, stream));
threadLocalStatus.savedSchedulerParams.clear();
break;
}
else if (func == mscclFuncAllGather && nBytes * comm->nRanks <= comm->mscclpp_threshold) {
INFO(NCCL_COLL,"%s: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p graphMode %d",
"mscclpp_ncclAllGather", comm->opCount, sendBuff, recvBuff, count, dataType, op, root, comm, comm->nRanks, stream, graphMode);
INFO(NCCL_COLL,"%s: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p graphMode %d ubr %d",
"mscclpp_ncclAllGather", comm->opCount, sendBuff, recvBuff, count, dataType, op, root, comm, comm->nRanks, stream,
graphMode, buffsRegistered);
NCCLCHECK(mscclpp_ncclAllGather(sendBuff, recvBuff, count, dataType, comm->mscclpp_comm, stream));
threadLocalStatus.savedSchedulerParams.clear();
break;
@@ -572,22 +574,24 @@ ncclResult_t mscclEnqueueCheck(
const bool buffsRegistered = sendBuffRegistered && recvBuffRegistered;
/* check if one rank per GPU and graph mode is enabled */
if ((graphMode || buffsRegistered) && comm->mscclCompatible && nBytes > 0 && (nBytes & 31) == 0) {
if ((graphMode || buffsRegistered || comm->mscclppForceEnable) && comm->mscclCompatible && nBytes > 0 && (nBytes & 31) == 0) {
bool isManagedBuffer = false;
if (sendBuff) CUDACHECK(hipPointerGetAttribute(&isManagedBuffer, HIP_POINTER_ATTRIBUTE_IS_MANAGED, const_cast<void*>(sendBuff)));
if (!isManagedBuffer && recvBuff) CUDACHECK(hipPointerGetAttribute(&isManagedBuffer, HIP_POINTER_ATTRIBUTE_IS_MANAGED, const_cast<void*>(recvBuff)));
if (isManagedBuffer) { /* MSCCL++ not enabled for managed memory buffers */ }
else if (func == mscclFuncAllReduce && nBytes <= comm->mscclpp_threshold && isMscclppAllReduceSupported(dataType, op)) {
INFO(NCCL_COLL,"%s: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p graphMode %d",
"mscclpp_ncclAllReduce", comm->opCount, sendBuff, recvBuff, count, dataType, op, root, comm, comm->nRanks, stream, graphMode);
INFO(NCCL_COLL,"%s: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p graphMode %d ubr %d",
"mscclpp_ncclAllReduce", comm->opCount, sendBuff, recvBuff, count, dataType, op, root, comm, comm->nRanks, stream,
graphMode, buffsRegistered);
NCCLCHECK(mscclpp_ncclAllReduce(sendBuff, recvBuff, count, dataType, op, comm->mscclpp_comm, stream));
threadLocalStatus.savedSchedulerParams.clear();
break;
}
else if (func == mscclFuncAllGather && nBytes * comm->nRanks <= comm->mscclpp_threshold) {
INFO(NCCL_COLL,"%s: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p graphMode %d" ,
"mscclpp_ncclAllGather", comm->opCount, sendBuff, recvBuff, count, dataType, op, root, comm, comm->nRanks, stream, graphMode);
INFO(NCCL_COLL,"%s: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p graphMode %d ubr %d" ,
"mscclpp_ncclAllGather", comm->opCount, sendBuff, recvBuff, count, dataType, op, root, comm, comm->nRanks, stream,
graphMode, buffsRegistered);
NCCLCHECK(mscclpp_ncclAllGather(sendBuff, recvBuff, count, dataType, comm->mscclpp_comm, stream));
threadLocalStatus.savedSchedulerParams.clear();
break;