Is it possible to run the PPO training script on an other GPU than 0?

[sschertler]

I'd like to run multiple instances in parallel. Using export CUDA_VISIBLE_DEVICES=1 in the second session leads to an EOF error caused by the multithreading.connection library. Thanks in Advance!

[mx781]

I'm unsure if we've tested this, but have you tried simply passing --train_gpu 1 to the training script? You might also want to change --inference_gpus, depending on what exactly you're doing.