Why do Python NVML discovery helpers use nvmlInit rather than nvmlInit_v2?

I am trying to understand an NVML initialization choice in PyTorch after a GPU fault on a shared multi-GPU server. Technical summary In this source revision, _raw_device_count_nvml() and _raw_device_uuid_nvml() call nvmlInit() through ctypes.CDLL. NVIDIA’s documentation describes legacy initialization as initializing all GPUs, whereas nvmlInit_v2() defers initialization until device-handle acquisition. The C header aliases nvmlInit to nvmlInit_v2, but my understanding is that this macro does n...