2.19.3-1

H800/H100 fixes and tuning.
Re-enable intra-process direct pointer buffer access when CUMEM is
enabled.

src/include/comm.h

@@ -127,7 +127,7 @@ struct ncclChannel {
   struct ncclChannelPeer** peers;
   struct ncclDevChannelPeer** devPeers;
   /* devPeer pointer array used for host side access */
-  struct ncclDevChannelPeer** devPeersHostPtr; 
+  struct ncclDevChannelPeer** devPeersHostPtr;
   struct ncclRing ring;
   int* devRingUserRanks;
   struct ncclTree tree;

src/include/graph.h

@@ -38,7 +38,6 @@ ncclResult_t ncclTopoNeedFlush(struct ncclTopoSystem* system, int64_t busId, int
 ncclResult_t ncclTopoCheckNet(struct ncclTopoSystem* system, int64_t id1, int64_t id2, int* net);
 int ncclPxnDisable(struct ncclComm* comm);
 ncclResult_t ncclTopoGetPxnRanks(struct ncclComm* comm, int** intermediateRanks, int* nranks);
-ncclResult_t ncclTopoGetLocalRank(struct ncclTopoSystem* system, int rank, int* localRank);
 
 // Find CPU affinity
 ncclResult_t ncclTopoGetCpuAffinity(struct ncclTopoSystem* system, int rank, cpu_set_t* affinity);