I have an application where I would like to train a model using my GPU, while simultaneously running inference with another model on the CPU. Is there any way to control which device is used on a per-model basis?
@Fredrik-M for now that would have to be a separate process due to the way the nd4j binaries are loaded. We’re working on modularization now where you can pick which data types and ops you want to include. We could look at improving the multi backend support as well.
We would likely do this through our workspaces abstraction which allows you to set a device. COuld ou file an issue with more of what you’re looking for? Thanks!