[Feature]: Allow specifying CUDA device in configure_cuda

[bo-kwok]

Description

Hi,

I’m currently using the setup_torch_device and configure_cuda methods. Right now, configure_cuda always sets:

self.torch_device = torch.device("cuda")

which defaults to the first available GPU (cuda:0). On multi-GPU systems, the first GPU might already be in use, and it would be useful to allow selecting a specific GPU.

Proposed:
Add an optional argument to configure_cuda, e.g., device_index, so that the device can be explicitly set:

def configure_cuda(self, ort_providers, device_index=0):
    self.torch_device = torch.device(f"cuda:{device_index}")

This would allow both PyTorch and ONNX Runtime to use the specified GPU, instead of always defaulting to cuda:0.

Thank you~

[bo-kwok]

Additionally, I noticed that one could try something like this:

_separator = Separator(output_dir=base_output)
_separator.torch_device = torch.device('device:2')

However, torch_device does not have a public setter method, so I’m not sure if assigning it directly is safe, or whether it might affect other parts of the code that rely on torch_device.