We get a segfault when we destroy the current ncclComm:

#0 0x00007fdba8a61838 in transportDestroyProxy (comm=comm@entry=0x5604b2781850) from /opt/anaconda3/envs/dlipy2/lib/python2.7/site-packages/torch/lib/../../../../libnccl.so.2
#1 0x00007fdba8a5d82b in commDestroy (comm=0x5604b2781850) from /opt/anaconda3/envs/dlipy2/lib/python2.7/site-packages/torch/lib/../../../../libnccl.so.2
#2 ncclCommDestroy (comm=0x5604b2781850) at /tmp/tmpxft_0000756e_00000000-5_init.compute_70.cudafe1.cpp:1158

in transportDestroyProxy:

  while (proxyState->pools != NULL) {
    struct ncclProxyPool *next = proxyState->pools->next;
    free(proxyState->pools);
    proxyState->pools = next;
}

attempting to dereference:

(gdb) p * & comm->proxyState->pools->next
Cannot access memory at address 0x21f0

ENV
    Lenovo x3650 M5 (Intel x64_64)
        48core / 377GB
    Ubuntu 18.04.2 LTS
        uname -r: 4.15.0-45-generic
    CUDA: 10.1
    NCCL: 2.4.2-1
        Note: the same code works fine with NCCL 2.3.5-5

(We built a version of the NCCL libraries with ENABLE_TRACE, however, we do not see any additional output beyond that enabled by "NCCL_DEBUG=INFO NCCL_DEBUG_SUBSYS=ALL)