Is there a way to specify which GPU is used for multi-gpu systems?

[lkraven]

I have a multi-gpu system and would like the engine to use the a specific gpu that would be dedicated for TTS. Is it possible to do so with environment variables or CLI options?

[remsky]

Believe you'd replace the -gpus all in the docker run or docker compose with the device ids re: https://docs.docker.com/engine/containers/resource_constraints/#gpu

[lkraven]

Answering my own question for others who might have the same issue.

Because GPU acceleration is using pytorch, CUDA environment parameters are honored.

To specify GPU, the environment can be set in the start-gpu script or elsewhere:

export CUDA_VISIBLE_DEVICES=1

Where "1" is the number of the device, or for multiple devices:

export CUDA_VISIBLE_DEVICES=1,2

See remsky's example above if you're running in a docker container.

Thank you!