Fix DMABUF support (#1218)

* Fix DMABUF support

* Reduce log output by moving dmabuf allocation details to TRACE

* Enable peer memory GDR support if ib_umem_get_peer is in kernel
このコミットが含まれているのは:
Wenkai Du
2024-06-25 08:00:15 -07:00
committed by GitHub
コミット 5d7078e383
3個のファイルの変更33行の追加3行の削除
+5 -1
ファイルの表示
@@ -722,7 +722,11 @@ static ncclResult_t fillInfo(struct ncclComm* comm, struct ncclPeerInfo* info, u
#endif
CUDACHECK(hipFree(ptr));
info->hasFineGrain = true;
NCCLCHECK(ncclGpuGdrSupport(comm, &info->gdrSupport));
// GPU supports GDR if DMABUF is supported
if (dmaBufSupported(comm) == ncclSuccess)
info->gdrSupport = 1;
else
NCCLCHECK(ncclGpuGdrSupport(comm, &info->gdrSupport));
}
else {
info->hasFineGrain = false;