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

Re-apply unroll=1 and 112 channels for gfx950 (#1706)

Browse Source 
* Reapply "[SRC] Enable unroll=1 for gfx950 (#1602)" (#1667)
This reverts commit a6972c0d09.

* Reapply "[GRAPH] Increase default nChannels to 112 for gfx950 (#1596)" (#1620)
This reverts commit 1a2eca1756.

[ROCm/rccl commit: 12517a957e]
This commit is contained in:
Nilesh M Negi
2025-05-28 14:58:10 -05:00
committed by GitHub
parent 22120c6303
commit 19ed482121
9 changed files with 123 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/init.cc
+7 -3
View File
@@ -101,8 +101,9 @@ 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, "gfx942") || IsArchMatch(devProp.gcnArchName, "gfx950"))
&& devProp.multiProcessorCount > 80))
if(IsArchMatch(devProp.gcnArchName, "gfx950"))
comm->unroll = NCCL_UNROLL_1;
else if(IsArchMatch(devProp.gcnArchName, "gfx908") || ((IsArchMatch(devProp.gcnArchName, "gfx942") && devProp.multiProcessorCount > 80)))
comm->unroll = NCCL_UNROLL_2;
else
comm->unroll = NCCL_UNROLL_4;
@@ -1348,7 +1349,7 @@ 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, "gfx942") || IsArchMatch(comm->topo->nodes[GPU].nodes[idx].gpu.gcn, "gfx950")) {
if (IsArchMatch(comm->topo->nodes[GPU].nodes[idx].gpu.gcn, "gfx942")) {
// Multi-node MI300A
int managed = 0;
CUDACHECK(hipDeviceGetAttribute(&managed, hipDeviceAttributeDirectManagedMemAccessFromHost, 0));
@@ -1365,6 +1366,9 @@ static ncclResult_t initTransportsRank(struct ncclComm* comm, struct ncclComm* p
allGather3Data[rank].nc = 4;
}
}
if (IsArchMatch(comm->topo->nodes[GPU].nodes[idx].gpu.gcn, "gfx950")) {
allGather3Data[rank].nc = 4;
}
allGather3Data[rank].pivotA2AEnabled = comm->topo->pivotA2AEnabled && rcclParamPivotAlltoallEnable();
comm->topo->ll128Enabled = comm->topo->ll128Enabled || rcclParamLL128ForceEnable();