Merge remote-tracking branch 'nccl/master' into develop

2020-07-01 16:51:25 -07:00
parent d3548cc474 1952325569
commit da3b197d6c
5 changed files with 15 additions and 8 deletions
@@ -687,9 +687,14 @@ ncclResult_t ncclTopoGetXmlFromGpu(struct ncclXmlNode* pciNode, nvmlDevice_t nvm
    if (index == -1) {
      const char* busId;
      NCCLCHECK(xmlGetAttr(sub, "target", &busId));
-      char* path;
-      NCCLCHECK(getPciPath(busId, &path));
-      NCCLCHECK(ncclTopoSetAttrFromSys(sub, path, "class", "tclass"));
+      if (strcmp(busId, "fffffff:ffff:ff") == 0) {
+        // Remote NVLink device is not visible inside this VM. Assume NVSwitch.
+        NCCLCHECK(xmlSetAttr(sub, "tclass", "0x068000"));
+      } else {
+        char* path;
+        NCCLCHECK(getPciPath(busId, &path));
+        NCCLCHECK(ncclTopoSetAttrFromSys(sub, path, "class", "tclass"));
+      }
    }
  }
  *gpuNodeRet = gpuNode;