2.20.3-1

Add support for alternating rings, allow for cross-nic rings without
cross-rail communication.
Add support for user buffer registration for network send/recv.
Optimize aggregated operations to better utilize all channels.
Add flattening for BCM PCI gen5 switches.
Add support for inter-node NVLink communication
Add support for port fusion in NET/IB.
Add support for ReduceScatter and AllGather using Collnet.
Update net API to v8.
Fix hang during A2A connection.

src/misc/shmutils.cc

@@ -169,7 +169,7 @@ ncclResult_t ncclShmemAllgather(struct ncclComm *comm, struct ncclShmemCollBuff
   int curRound = shmem->round;
   size_t mycnt;
 
-  if (comm == NULL || shmem == NULL || sendbuff == NULL || recvbuff == NULL) {
+  if (comm == NULL || shmem == NULL || sendbuff == NULL || recvbuff == NULL || shmem->maxTypeSize < typeSize) {
     ret = ncclInvalidArgument;
     goto exit;
   }
@@ -184,7 +184,7 @@ ncclResult_t ncclShmemAllgather(struct ncclComm *comm, struct ncclShmemCollBuff
     uint64_t t0 = clockNano();
     while(__atomic_load_n(shmem->cnt[curRound], __ATOMIC_ACQUIRE) != comm->localRanks + 1) {
       if (clockNano() - t0 >= 5 * 1000) sched_yield();
-      if (*comm->abortFlag == 1) {
+      if (__atomic_load_n(comm->abortFlag, __ATOMIC_RELAXED) == 1) {
         ret = ncclInternalError;
         goto exit;
       }