Check for _is_cuda() in compute_num_jobs

[bnellnm]

Check _is_cuda() before trying to determine number of nvcc threads in compute_num_jobs.

This will avoid an error where CUDA_HOME is not defined on machines without CUDA/nvcc installed, e.g. AMD.

[hliuca]

the _is_cuda isn't correct... it return True on HIP... the old implementation should be resotred.

[hliuca]

OLD:
def _is_cuda() -> bool:
return (torch.version.cuda is not None) and not _is_neuron()

New
def _is_cuda() -> bool:
return torch.version.cuda is not None

New is wrong.

[simon-mo]

@hliuca is the suggested fix adding back and not _is_neuron()? I'm confused how this is related to AMD

[hliuca]

false alarm... please ignore. thank you.

[jamestwhedbee]

I have run into this too, do we expect this to merge before v0.3.4 is released?

[hliuca]

As this blocks ROCm/HIP, this PR should be merged as soon as possible. Thank you :-)

[hliuca]

_is_cuda() has been restored in #3537 :-)

def _is_cuda() -> bool:

• return torch.version.cuda is not None

• return torch.version.cuda is not None and not _is_neuron()