Hello,
I’m trying to do something a bit weird: I’m trying to distribute functions in a graphmodule graph over a cluster (e.g. through some compile analysis passes + replacing the node with logic that serializes inputs, runs the function on a worker, and returns/deserializes the result). This means I need a way to serialize the torch OpOverload object to send to the worker node on the cluster.
Is there a way to cleanly do this? I’m hoping in particular that there’s some insights from torch’s RPC stack, but I don’t have a deep enough understanding of how it is built to know otoh.
Greatly appreciate your help!