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

Rail optimization for rings (#1140)

Browse Source 
- Modifies the ring creation algorithm to be friendlier to rail-optimized topologies (should not affect classic fabric topologies)

[ROCm/rccl commit: 4cb62f999a]
This commit is contained in:
gilbertlee-amd
2024-04-15 12:03:57 -06:00
committed by GitHub
parent 8ddb74e3b1
commit 422a7ffcbb
11 changed files with 296 additions and 69 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/graph/search.cc
+2 -1
View File
@@ -946,7 +946,7 @@ ncclResult_t ncclTopoCompute(ncclTopoSystem* system, struct ncclTopoGraph* graph
NCCLCHECK(parseGraphLight(strTrees, system, graph, NULL));
system->treeDefined=true;
} else {
NCCLCHECK(parseGraph(str, system, graph, NULL, NULL));
NCCLCHECK(parseGraph(str, system, graph, NULL, NULL, false));
int arch, vendor, model;
NCCLCHECK(ncclTopoCpuType(system, &arch, &vendor, &model));
if (graph->nChannels && arch == NCCL_TOPO_CPU_ARCH_X86 && vendor == NCCL_TOPO_CPU_VENDOR_AMD && model == NCCL_TOPO_CPU_TYPE_ROME) {
@@ -959,6 +959,7 @@ ncclResult_t ncclTopoCompute(ncclTopoSystem* system, struct ncclTopoGraph* graph
if (graph->nChannels) return ncclSuccess;
// try to match Rome 4P2H
NCCLCHECK(parseRome4P2H(system, graph));
if (graph->nChannels) return ncclSuccess;
// try to match 1H16P
NCCLCHECK(parse1H16P(system, graph));