Problem

When using and importing torch and SAHI inside a multiprocessing subprocess in a multi-GPU environment, sometimes
only a single GPU is visible thus torch.cuda.device_count() returns 1. In that case I was having problems loading the model to the correct device. SAHI performs an extra verification: it checks that the requested CUDA device index is < total CUDA devices.
In a subprocess that only sees one GPU (even if the system has multiple GPUs), torch.cuda.device_count() returns 1.
If the code tries to load on "cuda:1" or higher, the verification fails, and SAHI falls back to cuda:0 (the default).

In line 88 of utils/torch_utils.py:

arg = "cuda:" + (device if int(device) < torch.cuda.device_count() else "0")

I changed it to remove that extra-verification that prevented loading the model to the right GPU:

arg = f"cuda:{device}" if device else "cuda:0"

Key change:

• Skip the global check against torch.cuda.device_count().

• Trust the device string/int provided by the user ("cuda:0", 0, "cuda:1", etc.).

• This makes SAHI compatible with the standard multi-GPU pattern where each worker process is pinned to a GPU using CUDA_VISIBLE_DEVICES.