Your code1 doesn’t make any sense to me. You are creating a graph without any nodes. If you wish to argue that your empty subgraph node is a contraindication, I don’t agree with you. If you prefer, I will state that you are creating an empty subgraph (no nodes in it) and then attempt to instantiate it. That doesn’t make any sense, and I don’t know of any suggestions in the CUDA docs that indicate that is OK. As soon as you add an actual node (code2) that issue seems to be resolved.
For the valgrind issue on code3, my suggestion would be to retest on the latest CUDA version, and if the issue still presents, then file a bug. It’s possible that there may be some API misuse in your example, but after a quick look I didn’t spot it.