2.9.6-1

Add support for CUDA graphs.
Fuse BCM Gen4 switches to avoid suboptimal performance on some platforms. Issue #439.
Fix bootstrap issue caused by connection reordering.
Fix CPU locking block.
Improve CollNet algorithm.
Improve performance on DGX A100 for communicators with only one GPU per node.

src/collectives/device/prims_ll.h

@@ -1,5 +1,5 @@
 /*************************************************************************
- * Copyright (c) 2016-2020, NVIDIA CORPORATION. All rights reserved.
+ * Copyright (c) 2016-2021, NVIDIA CORPORATION. All rights reserved.
  *
  * See LICENSE.txt for license information
  ************************************************************************/
@@ -203,8 +203,9 @@ class ncclLLPrimitives {
     // Make sure step is updated before we read it.
     barrier();
 
-    for (int i=0; i<NRECV && recvPeers[i] >= 0; i++) loadRecvConn(&channel->devPeers[recvPeers[i]].recv.conn, i);
-    for (int i=0; i<NSEND && sendPeers[i] >= 0; i++) loadSendConn(&channel->devPeers[sendPeers[i]].send.conn, i);
+    // If we are going to support oneshot collNet + LL, then we would need to add connector index here
+    for (int i=0; i<NRECV && recvPeers[i] >= 0; i++) loadRecvConn(&channel->devPeers[recvPeers[i]].recv->conn, i);
+    for (int i=0; i<NSEND && sendPeers[i] >= 0; i++) loadSendConn(&channel->devPeers[sendPeers[i]].send->conn, i);
     loadRecvSync();
     loadSendSync();
   }