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

Add 8P6L multi-node models (#239)

Browse Source 
[ROCm/rccl commit: d5f90e19b5]
This commit is contained in:
Wenkai Du
2020-07-21 14:10:36 -07:00
committed by GitHub
parent 56d8c7c893
commit f604fc774e
15 changed files with 830 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/graph/connect.cc
+3 -3
View File
@@ -9,6 +9,7 @@
#include "graph.h"
#include "trees.h"
#include "rings.h"
#include "topo.h"
/******************************************************************/
/********************* Internode connection ***********************/
@@ -288,9 +289,8 @@ ncclResult_t ncclTopoPostset(struct ncclComm* comm, int* firstRanks, struct nccl
memcpy(ringPrev+nChannels*nranks, ringPrev, nChannels*nranks*sizeof(int));
memcpy(ringNext+nChannels*nranks, ringNext, nChannels*nranks*sizeof(int));
char *str = NULL;
NCCLCHECK(parseChordalRing(comm->topo, &str));
int end = std::min((int)ncclMaxNchannels(), (str ? nChannels*3 : ncclMinNchannels()));
int end = std::min((int)ncclMaxNchannels(),
((comm->topo->nodes[NET].count == 0 && comm->topo->type == RCCL_TOPO_CR8G) ? nChannels*3 : ncclMinNchannels()));
// Duplication should be complete now
nChannels = comm->nChannels = std::min(MAXCHANNELS,nChannels*2);