Regarding the server's lifecycle

[kyk666123]

If I want to perform inference with a trained model, is using trainer.dev the most convenient approach?
Also, I’ve noticed that the server seems to be released after trainer.fit or trainer.dev finishes—how can I extend its lifetime?

[ultmaster]

If I want to perform inference with a trained model, is using trainer.dev the most convenient approach?

I think verl has a test mode. In that case, you only need to run trainer.fit, but config verl in test mode.

I’ve noticed that the server seems to be released after trainer.fit or trainer.dev finishes—how can I extend its lifetime?

You can run the store server independantly, and connects to the external store. See: https://microsoft.github.io/agent-lightning/stable/tutorials/debug/#debug-the-algorithm-runner-boundary