如何在代码中打印GPU使用信息

依赖CUDA nvcc

watch -n 1 nvidia-smi

pip install gpustat 

gpustat -cp -i

使用Ctrl+C可以退出。

!pip install nvidia-ml-py3

import nvidia_smi
nvidia_smi.nvmlInit()
deviceCount = nvidia_smi.nvmlDeviceGetCount()
for i in range(deviceCount):
    handle = nvidia_smi.nvmlDeviceGetHandleByIndex(i)
    util = nvidia_smi.nvmlDeviceGetUtilizationRates(handle)
    mem = nvidia_smi.nvmlDeviceGetMemoryInfo(handle)
    print(f"|Device {i}| Mem Free: {mem.free/1024**2:5.2f}MB / {mem.total/1024**2:5.2f}MB | gpu-util: {util.gpu:3.1%} | gpu-mem: {util.memory:3.1%} |")

用户可以将GPU信息显示操作看作一个装饰器，在模型训练过程中就可以实时的显示GPU状态信息。

def gputil_decorator(func):
    def wrapper(*args, **kwargs):
        import nvidia_smi
        import prettytable as pt

        try:
            table = pt.PrettyTable(['Devices','Mem Free','GPU-util','GPU-mem'])
            nvidia_smi.nvmlInit()
            deviceCount = nvidia_smi.nvmlDeviceGetCount()
            for i in range(deviceCount):
                handle = nvidia_smi.nvmlDeviceGetHandleByIndex(i)
                res = nvidia_smi.nvmlDeviceGetUtilizationRates(handle)
                mem = nvidia_smi.nvmlDeviceGetMemoryInfo(handle)
                table.add_row([i, f"{mem.free/1024**2:5.2f}MB/{mem.total/1024**2:5.2f}MB", f"{res.gpu:3.1%}", f"{res.memory:3.1%}"])

        except nvidia_smi.NVMLError as error:
            print(error)

        print(table)
        return func(*args, **kwargs)
    return wrapper

nvidia-ml-py3可以直接查询nvml c-lib库，而无需通过nvidia-smi。因此，这个模块比nvidia-smi周围的包装器快得多。

from pynvml import *
nvmlInit()
handle = nvmlDeviceGetHandleByIndex(0)
info = nvmlDeviceGetMemoryInfo(handle)
print("Total memory:", info.total)
print("Free memory:", info.free)
print("Used memory:", info.used)

!pip install gputil

import GPUtil as GPU
GPU.showUtilization()

import GPUtil as GPU
GPUs = GPU.getGPUs()
for gpu in GPUs:
    print("GPU RAM Free: {0:.0f}MB | Used: {1:.0f}MB | Util {2:3.0f}% | Total {3:.0f}MB".format(gpu.memoryFree, gpu.memoryUsed, gpu.memoryUtil*100, gpu.memoryTotal))

注：用户在使用pytorch/tensorflow等深度学习框架时也可以使用框架自带的api进行查询。