I’m trying to use tuned-example DDPPO.
This configuration uses “num_gpus_per_worekr” = 1, so I think it would use gpu. However, it doesn’t uses gpu when I check it with nvidia-smi. Is there any way to make ddppo use the gpu?
My experiment setups are followings.
It should. The whole point of DDPPO is to compute gradients on the worker nodes. Can you file an issue against RLlib here?
with setup and steps to reproduce, etc. we will take a look. thanks.