• forked from wookayin/gpustat (v0.6) (The original repo' s version is higher now, but I have no time to follow up)

• The code is not quite complex: gpu_stats in main.py call print_formatted(sys.stdout, **kwargs) in core.py.
  I just modified the two files above.

• My version V.S. original version:

Another GPU trick (not relevant to gpustat, I just put the snippet here)

import os
def find_gpus(num_of_cards_needed=4):
    os.system('nvidia-smi -q -d Memory |grep -A4 GPU|grep Free >~/.tmp_free_gpus')
    # If there is no ~ in the path, return the path unchanged
    with open(os.path.expanduser ('~/.tmp_free_gpus'), 'r') as lines_txt:
        frees = lines_txt.readlines()
        idx_freeMemory_pair = [ (idx, int(x.split()[2]))
                                for idx, x in enumerate(frees) ]
    idx_freeMemory_pair.sort(reverse=True)  # 0号卡经常有人抢，让最后一张卡在下面的sort中优先
    idx_freeMemory_pair.sort(key=lambda my_tuple: my_tuple[1], reverse=True)
    usingGPUs = [str(idx_memory_pair[0]) for idx_memory_pair in
                    idx_freeMemory_pair[:num_of_cards_needed] ]
    usingGPUs = ','.join(usingGPUs)
    print('using GPUs:',end=' ')
    for pair in idx_freeMemory_pair[:num_of_cards_needed]:
        print(f'{pair[0]}号，此前空闲：{pair[1]/1024:.1f}GB')
    return usingGPUs

os.environ['CUDA_VISIBLE_DEVICES'] = find_gpus(num_of_cards_needed=1)  # must before `import torch`

\
\
\

Just less than nvidia-smi?

NOTE: This works with NVIDIA Graphics Devices only, no AMD support as of now. Contributions are welcome!

Self-Promotion: A web interface of gpustat is available (in alpha)! Check out gpustat-web.

$ gpustat

Options:

• --color : Force colored output (even when stdout is not a tty)
• --no-color : Suppress colored output
• -u, --show-user : Display username of the process owner
• -c, --show-cmd : Display the process name
• -f, --show-full-cmd : Display full command and cpu stats of running process
• -p, --show-pid : Display PID of the process
• -F, --show-fan : Display GPU fan speed
• -e, --show-codec : Display encoder and/or decoder utilization
• -P, --show-power : Display GPU power usage and/or limit (draw or draw,limit)
• -a, --show-all : Display all gpu properties above
• --watch, -i, --interval : Run in watch mode (equivalent to watch gpustat) if given. Denotes interval between updates. (#41)
• --json : JSON Output (Experimental, #10)

• To periodically watch, try gpustat --watch or gpustat -i (#41).
  • For older versions, one may use watch --color -n1.0 gpustat --color.
• Running nvidia-smi daemon (root privilege required) will make the query much faster and use less CPU (#54).
• The GPU ID (index) shown by gpustat (and nvidia-smi) is PCI BUS ID, while CUDA differently assigns the fastest GPU with the lowest ID by default. Therefore, in order to make CUDA and gpustat use same GPU index, configure the CUDA_DEVICE_ORDER environment variable to PCI_BUS_ID (before setting CUDA_VISIBLE_DEVICES for your CUDA program): export CUDA_DEVICE_ORDER=PCI_BUS_ID.

Install from PyPI:

pip install gpustat

If you don't have root privilege, please try to install on user namespace: pip install --user gpustat.

To install the latest version (master branch) via pip:

pip install git+https://github.com/wookayin/gpustat.git@master

Note that starting from v1.0, gpustat will support only Python 3.4+. For older versions (python 2.7, <3.4), you can continue using gpustat v0.x.

[0] GeForce GTX Titan X | 77'C, 96 % | 11848 / 12287 MB | python/52046(11821M)

• [0]: GPUindex (starts from 0) as PCI_BUS_ID
• GeForce GTX Titan X: GPU name
• 77'C: Temperature
• 96 %: Utilization
• 11848 / 12287 MB: GPU Memory Usage
• python/...: Running processes on GPU (and their memory usage)

See CHANGELOG.md

MIT License