[train] support memory per worker

[matthewdeng]

Why are these changes needed?

This enables scheduling Ray Train Workers by specifying "memory" in ScalingConfig.resources_per_worker. This additional logic is necessary because Ray Actors require the special memory kwarg rather than a "memory" entry in the resources dictionary.

As part of this change, I am also changing the interface of BackendExecutor and WorkerGroup to simply take a dictionary for all resources. This matches the interface of ScalingConfig and PlacementGroup, and centralizes all logic to convert to Ray Actor/Task kwargs in WorkerGroup.

Example:

from ray.train import ScalingConfig
from ray.train.torch import TorchTrainer


scaling_config = ScalingConfig(num_workers=2, resources_per_worker={"memory": 10_000})


def train_func(): 
    ...


trainer = TorchTrainer(train_func, scaling_config=scaling_config)
trainer.fit()

Before:

ValueError: The resources dictionary must not contain the key 'memory' or 'object_store_memory'

After:

(TorchTrainer pid=67598) Started distributed worker processes: 
(TorchTrainer pid=67598) - (ip=127.0.0.1, pid=67604) world_rank=0, local_rank=0, node_rank=0
(TorchTrainer pid=67598) - (ip=127.0.0.1, pid=67605) world_rank=1, local_rank=1, node_rank=0
(RayTrainWorker pid=67604) Setting up process group for: env:// [rank=0, world_size=2]

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 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 :(

[justinvyu]

Thanks! See the comment about also needing to update ScalingConfig.as_placement_group_factory, and also a request for the unit test.

[woshiyyya]

Good to Go!

Also Thanks for cleaning up the redundant input arguments.