[Train] Fix Deepspeed device ranks check in Lightning 2.0.5

[woshiyyya]

Why are these changes needed?

LightningTrainer deepspeed ci test failed due to lightning upgrade from 2.0.4 to 2.0.5, which introduced a check on device ranks in DeepSpeedStrategy Link. This PR aims to address the incompatibility and fix the test.

Related issue number

Fix #37374

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 added any new APIs to the API Reference. For example, if I added a
    method in Tune, I've added it in doc/source/tune/api/ under the
    corresponding .rst file.
• 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 :(

[matthewdeng]

Would something like this make sense? Not sure if the devices indices are always starting from 0.

Suggested change

	self.parallel_devices = list(range(torch.cuda.device_count()))
	devices = train.torch.get_device()
	if not isinstance(devices, list):
	devices = [devices]
	self.parallel_devices = [d.index for d in devices if d.index is not None]

[woshiyyya]

The parallel_devices here need to be set to all the cuda device ids for current node. train.torch.get_device() only return the one for the current worker.

For example, if we have two workers each with 1 GPU

• worker_0: CUDA_VISIBLE_DEVICES=2,3, train.torch.get_device()=torch.device(cuda:0)
• worker_1: CUDA_VISIBLE_DEVICES=2,3, train.torch.get_device()=torch.device(cuda:1)

parallel_devices here should be [torch.device(cuda:0), torch.device(cuda:1)]

Lightning added a special check for deepspeed, which requires the indices of parallel_devices equals to list(range(len(parallel_devices)))