Fix crash when only a subset of GPUs are visible within a container.

Fixes #326.


[ROCm/rccl commit: f36540f55a]

projects/rccl/src/graph/topo.cc

@@ -520,6 +520,7 @@ ncclResult_t ncclTopoGetSystem(struct ncclComm* comm, struct ncclTopoSystem** sy
       NCCLCHECK(int64ToBusId(comm->peerInfo[r].busId, busId));
       struct ncclXmlNode* node;
       NCCLCHECK(ncclTopoFillGpu(xml, busId, &node));
+      if (node == NULL) continue;
       NCCLCHECK(xmlSetAttrInt(node, "rank", r));
       NCCLCHECK(xmlInitAttrInt(node, "gdr", comm->peerInfo[r].gdrSupport));
     }

projects/rccl/src/graph/xml.cc

@@ -569,7 +569,7 @@ ncclResult_t ncclTopoGetXmlFromGpu(struct ncclXmlNode* pciNode, nvmlDevice_t nvm
     NCCLCHECK(xmlSetAttrInt(gpuNode, "dev", dev));
   }
   NCCLCHECK(xmlGetAttrInt(gpuNode, "dev", &dev));
-  if (dev == -1) return ncclSuccess;
+  if (dev == -1) { *gpuNodeRet = NULL; return ncclSuccess; }
 
   NCCLCHECK(xmlGetAttrIndex(gpuNode, "sm", &index));
   if (index == -1) {