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

Revert "Enable Ll128 on gfx90a (#823)" (#829)

Browse Source 
This reverts commit 420f8af6a0.

Also increase number of parallel jobs for linking
This commit is contained in:
Wenkai Du
2023-07-27 20:25:18 -07:00
committed by GitHub
parent c424979c14
commit 3db371c9a5
3 changed files with 8 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graph/tuning.cc
+4
View File
@@ -451,9 +451,13 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
int pEnable = protoEnable[p];
if (pEnable == 2 && p == NCCL_PROTO_LL128) {
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
#if defined(ENABLE_LL128)
// Enable LL128 by default only on gfx90a with available tuning table
pEnable = (graphs[a]->typeInter <= PATH_PXB) && graphs[a]->typeIntra <= PATH_NVL &&
(comm->topo->nodes[GPU].nodes[0].gpu.gcn == 910 && comm->topo->ll128Enabled) ? 1 : 0;
#else
pEnable = 0;
#endif
#else
// Enable LL128 by default only on Volta/Ampere/Hopper+NVLink. Other cases are not tested and may cause silent data corruption.
pEnable = 1;