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

enable building rccl for gfx950 (#1571)

Browse Source 
[ROCm/rccl commit: f268553ee4]
This commit is contained in:
Pedram Alizadeh
2025-02-25 16:13:48 -05:00
committed by GitHub
parent c363118589
commit acf5822a6c
24 changed files with 66 additions and 47 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/init.cc
+4 -4
View File
@@ -97,7 +97,7 @@ static uint64_t hashUniqueId(ncclUniqueId const &id) {
ncclResult_t commSetUnrollFactor(struct ncclComm* comm) {
hipDeviceProp_t devProp;
CUDACHECK(hipGetDeviceProperties(&devProp, comm->cudaDev));
if(IsArchMatch(devProp.gcnArchName, "gfx908") || (IsArchMatch(devProp.gcnArchName, "gfx94")
if(IsArchMatch(devProp.gcnArchName, "gfx908") || ((IsArchMatch(devProp.gcnArchName, "gfx94") || IsArchMatch(devProp.gcnArchName, "gfx950"))
&& devProp.multiProcessorCount > 80))
comm->unroll = NCCL_UNROLL_2;
else
@@ -1337,7 +1337,7 @@ static ncclResult_t initTransportsRank(struct ncclComm* comm, struct ncclComm* p
allGather3Data[rank].nc = std::max(allGather3Data[rank].nc, 4/ringGraph->nChannels);
if (ringGraph->nChannels > MAXCHANNELS/2)
allGather3Data[rank].nc = 1;
if (IsArchMatch(comm->topo->nodes[GPU].nodes[idx].gpu.gcn, "gfx94")) {
if (IsArchMatch(comm->topo->nodes[GPU].nodes[idx].gpu.gcn, "gfx94") || IsArchMatch(comm->topo->nodes[GPU].nodes[idx].gpu.gcn, "gfx950")) {
// Multi-node MI300A
int managed = 0;
CUDACHECK(hipDeviceGetAttribute(&managed, hipDeviceAttributeDirectManagedMemAccessFromHost, 0));
@@ -1858,7 +1858,7 @@ static ncclResult_t ncclCommInitRankFunc(struct ncclAsyncJob* job_) {
// a CUDA memory reconfig on load (c.f. NVSHMEM issue)
#ifdef USE_INDIRECT_FUNCTION_CALL
CUDACHECK(hipGetDeviceProperties(&devProp, 0));
if (ncclParamSetStackSize() == 1 && !IsArchMatch(devProp.gcnArchName,"gfx94")) {
if (ncclParamSetStackSize() == 1 && !IsArchMatch(devProp.gcnArchName,"gfx94") && !IsArchMatch(devProp.gcnArchName,"gfx950")) {
stackSize = rcclParamStackSizeOverride() ? rcclParamStackSizeOverride() : maxLocalSizeBytes;
if (stackSize == 0) {
if (IsArchMatch(devProp.gcnArchName,"gfx906"))
@@ -1924,7 +1924,7 @@ static ncclResult_t ncclCommInitRankFunc(struct ncclAsyncJob* job_) {
if (mscclEnabled() && (comm->topo->mscclEnabled || mscclForceEnabled()) && mscclppCommCompatible(comm)) {
hipDeviceProp_t devProp;
CUDACHECK(hipGetDeviceProperties(&devProp, cudaDev));
comm->mscclppCompatible = IsArchMatch(devProp.gcnArchName, "gfx94");
comm->mscclppCompatible = IsArchMatch(devProp.gcnArchName, "gfx94") || IsArchMatch(devProp.gcnArchName, "gfx950");
if (comm->mscclppCompatible) {
bool mapContainsId = (mscclpp_uniqueIdMap.count(job->commId) > 0);
auto& mscclppUniqueId = mscclpp_uniqueIdMap[job->commId];