I am trying to run an example simulation on a system with AMD GPUs.
This is the config in my pyproject.toml file
options.num-supernodes = 4
options.backend.client-resources.num-gpus = 0.25
options.backend.client-resources.num-cpus = 4
But the GPU is never used by the program. What is missing ? Is there any additional configuration I need to do. On the system all the 8 GPUs are detected when i run a sample pytorch script.
Hello @prasadwrites , thanks for creating the post. What ML framework are you using? You probably need to send the model+dataloaders to the device that was exposed to the ClientApp after increasing num-gpus?.
For example, take a look at how we do it in this pytorch example. Because we want the code to automatically use the GPU if there’s one, we typically add the line
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
to all our examples.
I hope the issue you are facing is just that! (that you forgot to do send the model to cuda).
Let us know if you got it working