No CUDA when combining with_resources and with_parameters

[ju2ez]

What happened + What you expected to happen

When passing multiple objects to tune.with_parameters CUDA is not recognized anymore in the trainable process.

What I did:

The adding of a parameters next to the trainable in the with_parameters function leads to tune not detecting the GPU anymore.

Any thoughts?

Versions / Dependencies

Ray Version 2.1.0

Reproduction script

def trainable(cfg, test):
    print(torch.cuda.is_available())

tune.Tuner(tune.with_parameters(trainable=tune.with_resources(trainable, {"gpu:1"}), 
test={"1": 2} ))
tuner.fit()

Issue Severity

Medium: It is a significant difficulty but I can work around it.

[justinvyu]

This has been resolved by #29740 - this is an issue with nesting these wrappers, and a quick workaround for now is to switch the order: tune.with_resources(tune.with_parameters(...), ...)