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

Force enable proto and/or algo after model selection (#1799)

Browse Source 
* Force enable proto or algo

* Remove inc nccl_common.h

* Move logic and add error checks

* Fix topo_expl compatibility

* Allow algo/proto overrides

* Remove extra function decl

* Clarify warning message

* Move algo/proto overrides into separate functions

* Update CHANGELOG.md

[ROCm/rccl commit: 7ccc6f268f]
This commit is contained in:
Mustafa Abduljabbar
2025-09-03 08:54:13 -04:00
committed by GitHub
parent 1999f2eba8
commit 1a7ab8dfc8
5 changed files with 75 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/include/rccl_common.h
+3
View File
@@ -81,6 +81,9 @@ inline size_t rcclGetSizePerRank(ncclFunc_t const& func, size_t const& nBytes, i
// For AR, this is the send/recv size per rank
return (func == ncclFuncReduceScatter || func == ncclFuncAllGather || func == ncclFuncBroadcast) ? nBytes / nRanks : nBytes;
}
ncclResult_t rcclGetAlgoProtoIndex(const char *envStr, const char* algoProtoString[], int nEntries, int& result);
ncclResult_t rcclOverrideProtocol(const char* ncclProtoStr[], float table[][NCCL_NUM_PROTOCOLS], struct ncclTaskColl* info);
ncclResult_t rcclOverrideAlgorithm(const char* ncclAlgoStr[], float table[][NCCL_NUM_PROTOCOLS], struct ncclTaskColl* info);
void rcclUpdateCollectiveProtocol(struct ncclComm* comm, size_t const& nBytes, struct ncclTaskColl* info);
void rcclUpdateThreadThreshold(struct ncclComm* comm, size_t const& nBytes, struct ncclTaskColl* info, int& threadThreshold);
void rcclSetPipelining(struct ncclComm* comm, size_t const& nBytes, struct ncclTaskColl* info);