[Train] Support returning multiple devices in train.torch.get_device()

[amogkam]

For model parallel workloads like stable diffusion fine-tuning, we may want to use multiple GPUs per Ray Train worker. This PR implements support for train.torch.get_device() returning a List of devices for these use cases so that users do not need to manage device setting themselves.

Why are these changes needed?

Related issue number

Checks

• I've signed off every commit(by using the -s flag, i.e., git commit -s) in this PR.
• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/master/.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested :(

[krfricke]

Thanks, LGTM.

This will still require us to change the prepare_model() methods to allow to select a specific device. We can update this in a follow-up (we may just want to pass in the cuda device ID directly and otherwise default to get_device()[0]

[amogkam]

Good point, updated

[Yard1]

Isn't this technically a breaking change that would require a deprecation cycle?

[amogkam]

I think it should be fine since this is only affecting the multi-GPU case and this is still beta API, but I will update to log a warning about the change.