Add support for H100 (sm90).
Make sure NCCL kernel honor user stream priorities.
Šī revīzija ir iekļauta:
Sylvain Jeaugey
2022-09-27 02:31:13 -07:00
vecāks 78313a6d21
revīzija da8152e57a
22 mainīti faili ar 303 papildinājumiem un 76 dzēšanām
+1 -1
Parādīt failu
@@ -27,7 +27,7 @@ ncclResult_t initChannel(struct ncclComm* comm, int channelId) {
NCCLCHECK(ncclCudaCallocAsync(&channel->devRingUserRanks, nRanks, comm->deviceStream.stream));
ncclCommPushCudaFree(comm, channel->devRingUserRanks);
NCCLCHECK(ncclStrongStreamRelease(ncclCudaGraphNull(), &comm->deviceStream));
NCCLCHECK(ncclStrongStreamRelease(ncclCudaGraphNone(), &comm->deviceStream));
for (int r=0; r < nRanks+1; ++r) {
for (int b=0; b < NCCL_MAX_CONNS; b++) {