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

Cette révision appartient à :
BertanDogancay
2024-01-30 14:43:43 -08:00
révision 9ff53eeeae
11 fichiers modifiés avec 183 ajouts et 187 suppressions
+1 -3
Voir le fichier
@@ -305,7 +305,7 @@ struct ncclComm {
// Flag to ask NCCL kernels to abort
volatile uint32_t *abortFlag;
volatile uint32_t *childAbortFlag;
volatile uint32_t *abortFlagRefCount;
uint32_t *abortFlagRefCount;
// Flags for enable P2P NET
uint32_t p2pNet;
@@ -353,8 +353,6 @@ struct ncclComm {
int nvlsRegSupport;
/* sharable NVLS resource. */
struct ncclNvlsSharedRes* nvlsResources;
struct ncclShmemCollBuff nvlsShmem;
void *nvlsShmemHandle;
ssize_t channelSize; // User requested work size (bytes) for channel partitions
+2 -5
Voir le fichier
@@ -206,7 +206,6 @@ struct ncclProxyRpcResponseHeader {
};
struct ncclProxyState {
int internalRefCount;
int refCount;
int tpRank;
int tpnRanks;
@@ -221,11 +220,10 @@ struct ncclProxyState {
ncclNet_t* ncclNet;
ncclCollNet_t* ncclCollNet;
volatile uint32_t* abortFlag;
volatile uint32_t* abortFlagRefCount;
// Service thread
pthread_t thread;
struct ncclSocket* listenSock;
volatile int stop;
int stop;
CUcontext cudaCtx;
ncclResult_t asyncResult;
@@ -306,8 +304,7 @@ ncclResult_t ncclProxyClientGetFdBlocking(struct ncclComm* comm, struct ncclProx
ncclResult_t ncclProxyStop(struct ncclComm* comm);
ncclResult_t ncclProxyShmUnlink(struct ncclComm* comm);
ncclResult_t ncclProxyDestroy(struct ncclProxyState *proxyState);
ncclResult_t ncclProxyTryDetach(struct ncclProxyState *proxyState);
ncclResult_t ncclProxyDestroy(struct ncclComm* comm);
ncclResult_t mscclSaveProxy(struct ncclComm* comm, struct ncclChannel* channel, int type, int peer, struct ncclProxyOp* op, int connIndex);
#endif
+2
Voir le fichier
@@ -68,6 +68,8 @@ struct ncclNvlsSharedRes {
char shareableHandle[NVLS_HANDLE_SIZE];
size_t ucGran;
int nChannels;
struct ncclShmemCollBuff nvlsShmem;
void *nvlsShmemHandle;
};
#endif /* CUDART_VERSION >= 12010 */