2.10.3-1

Add support for bfloat16.
Add ncclAvg reduction operation.
Improve performance for aggregated operations.
Improve performance for tree.
Improve network error reporting.
Add NCCL_NET parameter to force a specific network.
Add NCCL_IB_QPS_PER_CONNECTION parameter to split IB traffic onto multiple queue pairs.
Fix topology detection error in WSL2.
Fix proxy memory elements affinity (improve alltoall performance).
Fix graph search on cubemesh topologies.
Fix hang in cubemesh during NVB connections.

src/include/graph.h

@@ -13,6 +13,7 @@
 #include <stdlib.h>
 #include <ctype.h>
 #include <stdio.h>
+#include <sched.h>
 
 ncclResult_t ncclTopoCudaPath(int cudaDev, char** path);
 
@@ -33,8 +34,8 @@ ncclResult_t ncclTopoGetNetDev(struct ncclTopoSystem* system, int rank, struct n
 ncclResult_t ncclTopoCheckP2p(struct ncclTopoSystem* system, int64_t id1, int64_t id2, int* p2p, int *read, int* intermediateRank);
 ncclResult_t ncclTopoCheckGdr(struct ncclTopoSystem* topo, int64_t busId, int netDev, int read, int* useGdr);
 
-// Set CPU affinity
-ncclResult_t ncclTopoSetAffinity(struct ncclTopoSystem* system, int rank);
+// Find CPU affinity
+ncclResult_t ncclTopoGetCpuAffinity(struct ncclTopoSystem* system, int rank, cpu_set_t* affinity);
 
 #define NCCL_TOPO_CPU_ARCH_X86 1
 #define NCCL_TOPO_CPU_ARCH_POWER 2
@@ -100,6 +101,6 @@ ncclResult_t ncclTopoPostset(struct ncclComm* comm, int* firstRanks, int* treePa
 
 ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCompCap, struct ncclTopoGraph* treeGraph, struct ncclTopoGraph* ringGraph, struct ncclTopoGraph* collNetGraph);
 #include "info.h"
-ncclResult_t ncclTopoGetAlgoTime(struct ncclInfo* info, int algorithm, int protocol, float* time);
+ncclResult_t ncclTopoGetAlgoTime(struct ncclInfo* info, int algorithm, int protocol, int numPipeOps, float* time);
 
 #endif