Fix ProcessRaisedException

✅ Solution

ProcessRaisedException in accelerate often arises from CUDA initialization failures within subprocesses launched by `notebook_launcher` or similar utilities, particularly in environments like Kaggle notebooks. Address this by explicitly setting the CUDA device visible to the subprocesses using `torch.cuda.set_device(device_index)` or `os.environ["CUDA_VISIBLE_DEVICES"]=<device_index>` at the beginning of the launched function, ensuring each process uses an available and correctly initialized CUDA device or disabling CUDA if not needed (`torch.device("cpu")`). This helps isolate and manage CUDA context per process, preventing initialization conflicts.