Merge remote-tracking branch 'nccl/master' into develop

[ROCm/rccl commit: 36343be84f]
Bu işleme şunda yer alıyor:
BertanDogancay
2025-01-23 11:48:18 -06:00
işlemeyi yapan: Corey Derochie
işleme 1b000665df
122 değiştirilmiş dosya ile 9009 ekleme ve 4702 silme
+5 -5
Dosyayı Görüntüle
@@ -39,7 +39,7 @@ static ncclResult_t socketProgressOpt(int op, struct ncclSocket* sock, void* ptr
}
}
(*offset) += bytes;
if (sock->abortFlag && __atomic_load_n(sock->abortFlag, __ATOMIC_RELAXED)) {
if (sock->abortFlag && __atomic_load_n(sock->abortFlag, __ATOMIC_ACQUIRE)) {
INFO(NCCL_NET, "socketProgressOpt: abort called");
return ncclInternalError;
}
@@ -623,12 +623,12 @@ ncclResult_t ncclSocketConnect(struct ncclSocket* sock) {
do {
NCCLCHECK(socketProgressState(sock));
} while (sock->asyncFlag == 0 &&
(sock->abortFlag == NULL || __atomic_load_n(sock->abortFlag, __ATOMIC_RELAXED) == 0) &&
(sock->abortFlag == NULL || __atomic_load_n(sock->abortFlag, __ATOMIC_ACQUIRE) == 0) &&
(sock->state == ncclSocketStateConnecting ||
sock->state == ncclSocketStateConnectPolling ||
sock->state == ncclSocketStateConnected));
if (sock->abortFlag && __atomic_load_n(sock->abortFlag, __ATOMIC_RELAXED)) return ncclInternalError;
if (sock->abortFlag && __atomic_load_n(sock->abortFlag, __ATOMIC_ACQUIRE)) return ncclInternalError;
switch (sock->state) {
case ncclSocketStateConnecting:
@@ -670,11 +670,11 @@ ncclResult_t ncclSocketAccept(struct ncclSocket* sock, struct ncclSocket* listen
do {
NCCLCHECKGOTO(socketProgressState(sock), ret, exit);
} while (sock->asyncFlag == 0 &&
(sock->abortFlag == NULL || __atomic_load_n(sock->abortFlag, __ATOMIC_RELAXED) == 0) &&
(sock->abortFlag == NULL || __atomic_load_n(sock->abortFlag, __ATOMIC_ACQUIRE) == 0) &&
(sock->state == ncclSocketStateAccepting ||
sock->state == ncclSocketStateAccepted));
if (sock->abortFlag && __atomic_load_n(sock->abortFlag, __ATOMIC_RELAXED)) return ncclInternalError;
if (sock->abortFlag && __atomic_load_n(sock->abortFlag, __ATOMIC_ACQUIRE)) return ncclInternalError;
switch (sock->state) {
case ncclSocketStateAccepting: