fix(api/libnvml): fix process info support for NVIDIA R535 driver (CUDA 12.2+)

[XuehaiPan]

Issue Type

• Bug fix

Description

The start with the NVIDIA R510 driver, the new version 3 APIs have been added for nvmlDeviceGet{Compute,Graphics,MPSCompute}RunningProcesses. But the version 3 functions still use the version 2 type struct as the function argument type:

class c_nvmlProcessInfo_v2_t(pynvml._PrintableStructure):
    _fields_ = [
        ('pid', ctypes.c_uint),
        ('usedGpuMemory', ctypes.c_ulonglong),
        ('gpuInstanceId', ctypes.c_uint),
        ('computeInstanceId', ctypes.c_uint),
    ]
    _fmt_ = {
        'usedGpuMemory': '%d B',
    }

Recently, the NVIDIA R535 driver came out. The version 3 APIs starts to use the new version 3 type struct without a version bump. This results in invalid memory access and produces the wrong results.

class c_nvmlProcessInfo_v3_t(pynvml._PrintableStructure):
    _fields_ = [
        ('pid', ctypes.c_uint),
        ('usedGpuMemory', ctypes.c_ulonglong),
        ('gpuInstanceId', ctypes.c_uint),
        ('computeInstanceId', ctypes.c_uint),
        ('usedGpuCcProtectedMemory', ctypes.c_ulonglong),
    ]
    _fmt_ = {
        'usedGpuMemory': '%d B',
        'usedGpuCcProtectedMemory': '%d B',
    }

The two type structs have different sizes:

>>> ctypes.sizeof(libnvml.c_nvmlProcessInfo_v2_t)
24
>>> ctypes.sizeof(libnvml.c_nvmlProcessInfo_v3_t)
32

This PR adds a helper function that determines the API version and type struct version of nvmlDeviceGet{Compute,Graphics,MPSCompute}RunningProcesses on the first API call.

Motivation and Context

Fixes #75
Fixes #76