2.15.1-1
Add support for H100 (sm90). Make sure NCCL kernel honor user stream priorities.
Этот коммит содержится в:
+2
-2
@@ -1055,8 +1055,8 @@ void* ncclProxyService(void* _args) {
|
||||
int asyncOpCount = 0;
|
||||
while ((stop == 0 || (stop == 1 && npeers > 0)) && *comm->abortFlag == 0) {
|
||||
/* never let proxy service thread blocks in poll, or it cannot receive abortFlag. */
|
||||
if (int error = poll(pollfds, NCCL_MAX_LOCAL_RANKS+1, asyncOpCount ? 0 : 500) < 0) {
|
||||
WARN("[Proxy Service] Poll failed with error %d", error);
|
||||
if (poll(pollfds, NCCL_MAX_LOCAL_RANKS+1, asyncOpCount ? 0 : 500) < 0) {
|
||||
WARN("[Proxy Service] Poll failed: %s\n", strerror(errno));
|
||||
return NULL;
|
||||
}
|
||||
if (pollfds[NCCL_MAX_LOCAL_RANKS].revents) {
|
||||
|
||||
Ссылка в новой задаче
Block a user