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
This commit is contained in:
Mustafa Abduljabbar
2025-09-03 08:54:13 -04:00
committed by GitHub
parent 361d596229
commit 7ccc6f268f
5 changed files with 75 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/enqueue.cc
+2
View File
@@ -2103,6 +2103,8 @@ static ncclResult_t topoGetAlgoInfo(
if (info->algorithm == NCCL_ALGO_TREE) nt = NCCL_MAX_NTHREADS; // Tree now uses all threads always.
if (info->algorithm == NCCL_ALGO_PAT) nt = NCCL_MAX_NTHREADS;
info->nWarps = nt/comm->WarpSize;
rcclOverrideAlgorithm(ncclAlgoStr, table, info);
rcclOverrideProtocol(ncclProtoStr, table, info);
return ncclSuccess;
}