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

removed gfx940 and gfx941 (#1606)

Browse Source 
* removed gfx940 and gfx941

* removed gfx940 and gfx941

* Update "gfx94" to "gfx942" in init.cc

* Updated remaining "gfx94" updates to "gfx942"

* Update filenames and variables from gfx940 to gfx942

---------

Co-authored-by: akolliasAMD <akollias@amd.com>
This commit is contained in:
corey-derochie-amd
2025-03-20 09:34:53 -06:00
committed by GitHub
parent 6403afff4a
commit 6505639cf4
30 changed files with 74 additions and 79 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graph/connect.cc
+2 -2
View File
@@ -286,7 +286,7 @@ static ncclResult_t setTreeDown(struct ncclTree* tree, int* indexes, int d) {
static ncclResult_t connectTrees(struct ncclComm* comm, int* treeToParent, int* treeToChild0, int* treeToChild1, int* treePatterns) {
const int channelLimit = (IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx94") || IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx950")) ? 2*CHANNEL_LIMIT : CHANNEL_LIMIT;
const int channelLimit = (IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx942") || IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx950")) ? 2*CHANNEL_LIMIT : CHANNEL_LIMIT;
const int nChannels = (comm->nChannels > channelLimit) ? comm->nChannels / 2 : comm->nChannels;
const int nNodes = comm->nNodes, node = comm->node;
@@ -804,7 +804,7 @@ ncclResult_t ncclTopoPostset(struct ncclComm* comm, int* firstRanks, int* treePa
}
// Only use full MAXCHANNELS for gfx94x and gfx950
int maxChannels = (IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx94") || IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx950")) ?
int maxChannels = (IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx942") || IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx950")) ?
((comm->topo->nodes[GPU].nodes[0].gpu.cu == 80 || comm->topo->nodes[GPU].nodes[0].gpu.cu == 20 || comm->topo->nodes[GPU].nodes[0].gpu.cu == 38)
? comm->topo->nodes[GPU].nodes[0].gpu.cu : MAXCHANNELS) : 2*CHANNEL_LIMIT;