net_ib: return ncclSuccess if read roceTypePath failed and errno is E…

…INVAL. #890
NVIDIA · Sep 5, 2024 · dc445aa · dc445aa · gcongiu · Jan 13, 2025
1 parent 178b6b7
commit dc445aa
Showing 1 changed file with 4 additions and 0 deletions.
diff --git a/src/transport/net_ib.cc b/src/transport/net_ib.cc
@@ -248,6 +248,10 @@ static ncclResult_t ncclIbRoceGetVersionNum(const char* deviceName, int portNum,
   close(fd);
 
   if (ret == -1) {
+    WARN("NET/IB: read failed: %s, roceTypePath: %s", strerror(errno), roceTypePath);
+    if (errno == EINVAL) {
+      return ncclSuccess;
+    }
     return ncclSystemError;
   }