Add support for CUDA 12.0, drop Kepler (sm_35).
Support for H100 features.
Make socket code more robust and protected. Solves #555.
Improve performance on large CUDA graphs, reducing dependencies.
Reduce inter-socket bandwidth on AMD CPUs to favor better paths.
Various fixes to ncclCommAbort.
Make service thread polling resistant to EINTR.
Compile with profiling API by default.
Extend NVTX instrumentation with call arguments.
Cette révision appartient à :
Sylvain Jeaugey
2022-11-29 04:27:46 -08:00
Parent 614b49f0de
révision 28189e2df8
46 fichiers modifiés avec 3325 ajouts et 1037 suppressions
+9 -3
Voir le fichier
@@ -10,10 +10,16 @@
#include "nccl.h"
#include "comm.h"
struct ncclBootstrapHandle {
uint64_t magic;
union ncclSocketAddress addr;
};
static_assert(sizeof(struct ncclBootstrapHandle) <= sizeof(ncclUniqueId), "Bootstrap handle is too large to fit inside NCCL unique ID");
ncclResult_t bootstrapNetInit();
ncclResult_t bootstrapCreateRoot(ncclUniqueId* commId, bool idFromEnv);
ncclResult_t bootstrapGetUniqueId(ncclUniqueId* out);
ncclResult_t bootstrapInit(ncclUniqueId* id, struct ncclComm* comm);
ncclResult_t bootstrapCreateRoot(struct ncclBootstrapHandle* handle, bool idFromEnv);
ncclResult_t bootstrapGetUniqueId(struct ncclBootstrapHandle* handle);
ncclResult_t bootstrapInit(struct ncclBootstrapHandle* handle, struct ncclComm* comm);
ncclResult_t bootstrapAllGather(void* commState, void* allData, int size);
ncclResult_t bootstrapSend(void* commState, int peer, int tag, void* data, int size);
ncclResult_t bootstrapRecv(void* commState, int peer, int tag, void* data, int size);