Selecting action of N agents inside a single GPU with torch.distributed.rpc

[wayi (Yi Wang)]

Is this the same question as this post?

RPC does not seem to help in forward time

Using RPC only decreases “backward time”. Hello I am using distributed.rpc changing the example in examples/main.py at 01539f9eada34aef67ae7b3d674f30634c35f468 · pytorch/examples · GitHub I have N agents and 1 simulation. I would like to select actions in parallel, if possible in different GPU’s. However, when I execute my code the only thing that seems to reduce is the update part. Below there is the code for selecting the actions which does not seem to be reduced at all, any ideas? from to…