pytorch - 💡(How to fix) Fix Accelerator-independent API to determine if a specific backend is available? [1 comments, 2 participants]

🚀 The feature, motivation and pitch

The most common pattern I've seen to determine whether you have, say, cuda available, is to do something like this:

if torch.cuda.is_available():
    do_some_cuda_stuff_bro()

This lets you write code that runs equivalently on CUDA and CPU code with customizations for the CUDA portion. If you have an out-of-tree backend, though, the backend itself will not be available, so if you want to write something equivalent like:

if torch.my_backend.is_available():
    configure_my_backend()

You'll get an AttributeError rather than graceful degradation, so instead you have to do something like this:

if hasattr(torch, "my_backend") and torch.my_backend.is_available():
    configure_my_backend()

Which is a bit unwieldy. I suspect that many people writing backend-independent code will want a single easy check for "is the backend present" and "is the accelerator available", so it would be nice if there were an official, standard way to do this check.

A possible interface for this could be a generic torch.is_available like this:

def is_available(backend: str) -> bool:
  module = getattr(torch, backend, None)
  return module is not None and module.is_available()

Alternatives

No response

Additional context

No response

cc @bdhirsh @NmomoN @mengpenghui @fwenguang @cdzhan @1274085042 @PHLens @albanD @guangyey @EikanWang