2.9.8-1

Fix memory leaks.
Fix crash in bootstrap error case.
Fix Collnet clean-up issue.
Make PCI switch vendor/device optional for XML injection.
Add support for nvidia-peermem module.


[ROCm/rccl commit: ca8485b0d0]

projects/rccl/src/group.cc

@@ -164,6 +164,7 @@ ncclResult_t ncclGroupEnd() {
   for (int i=0; i<ncclGroupIndex; i++) doneArray[i] = 1;
   ncclResult_t ret = ncclGroupError;
   int usingCudaGraphAll = -1;
+  cudaGraph_t* graphs = NULL;
   if (ret != ncclSuccess) goto group_cleanup;
 
   /* Launch async ncclCommInitRank */
@@ -307,7 +308,6 @@ sched_delta:
    */
 
   // Check whether we are in cuda graph mode
-  cudaGraph_t* graphs;
   NCCLCHECK(ncclCalloc(&graphs, ncclGroupIndex));
   for (int i=0; i<ncclGroupIndex; i++) {
     struct ncclAsyncArgs* args = ncclGroupArgs+i;
@@ -407,5 +407,6 @@ end:
   ncclGroupError = ncclSuccess;
   ncclGroupIndex = 0;
   CUDACHECK(cudaSetDevice(savedDev)); // do other clean-ups first before calling cudaSetDevice, because this call can fail too
+  if (graphs) free(graphs);
   return ret;
 }