Improve compatibility with CUDA-enabled pytorch on non-CUDA devices

[mainrs]

Problem Description

The pytorch CUDA-enabled libraries are more capable than the CPU-only one. They can also run on the CPU if no CUDA device is available. However, due to a race condition, the current code base calls into the CUDA driver even if one passes --no-cuda as an argument.

The issues are these lines of code:

cleanrl/cleanrl/dqn.py

Line 147 in 8cbca61

device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu")

They should first check if the flag is set and only then call torch.cuda.is_available. That way, the program runs perfectly fine in those scenarios.

Possible Solution

device = torch.device("cuda" if args.cuda and torch.cuda.is_available() else "cpu")

[pseudo-rnd-thoughts]

Could you clarify why this is a "race condition"?

it shouldn't be cuda unless both cases are true so I don't understand why ordering would matter in this case

[mainrs]

Because Python's and is lazy. If the first option is false, than it doesn't evaluate the second one. If I pass --no-cuda, Python still runs torch.cuda.is_available, even though I specified that I do not want to use CUDA.

On devices that don't have CUDA drivers (or have the stub drivers) but have the CUDA version of PyTorch installed, this throws a runtime error. However, using the CPU on such devices is valid, since the PyTorch library can still function by using the CPU.

[pseudo-rnd-thoughts]

Ok, that makes sense, I thought that pytorch would be smart enough to not raise a runtime error for this function.
I would make a PR that makes your suggested change