Parse nvidia-smi from python

parse_nvidia_smi.py

""" 
Parse output of nvidia-smi into a python dictionary.
This is very basic!
"""

import subprocess
import pprint

sp = subprocess.Popen(['nvidia-smi', '-q'], stdout=subprocess.PIPE, stderr=subprocess.PIPE)

out_str = sp.communicate()
out_list = out_str[0].split('\n')

out_dict = {}

for item in out_list:
    try:
        key, val = item.split(':')
        key, val = key.strip(), val.strip()
        out_dict[key] = val
    except:
        pass

pprint.pprint(out_dict)

when i run your code, I get the following error. Could you help me with this ?

TypeError Traceback (most recent call last)
in ()
5
6 out_str = sp.communicate()
----> 7 out_list = out_str[0].split('\n')
8
9 out_dict = {}

TypeError: a bytes-like object is required, not 'str'

out_str is being read as bytes not a string - you need to decode it prior to the split. Change line 12 (your line 7) to:
out_list = out_str[0].decode("utf-8").split('\n')
and it should work for you

Just in case you didn't know... http://pypi.python.org/pypi/nvidia-ml-py/

Thanks @yssybyl -- indeed the code needs that mod to run on Py3. I think @samidarko's link to nividia-ml-py is a better general solution (my gist doesn't parse everything in a useful way, e.g. it doesn't treat multiple GPUs separately).

Great effort. Just want to point out something I found.
Your code gives out the GPU utilisation of the last process only whereas on running nvidia-smi -q, it can be seen that there could be a number of different processes consuming different amounts of GPU memory.

In the image below, it can be seen that there are 3 different processes.

But your code gives out the last one, probably you need to check for nested lists or something. Again screenshot of output of your code below.

Hi, I create a repo to do this.
It parses and maps the output of nvidia-smi -q into a Python object.
You can get the attributes like:

log = NVLog()
print(log['Attached GPUs']['GPU 00000000:04:00.0']['Processes'][0]['Used GPU Memory'])

And I have re-implemented the table when running nvidia-smi.

print(log.as_table())

Please see https://github.com/petronny/nvsmi

One point need to be paid attention to.
If there are two or more processes on GPUs, the dict will overwrite the value of the key "Process ID",
also other keys.

Just in case you didn't know... http://pypi.python.org/pypi/nvidia-ml-py/

The page has no description or documentation on how to use it. Did you find any resources on how to use the pip package?

@amanuel1995 -- that package is looking a bit old now, but here's some alternative links:
https://github.com/jonsafari/nvidia-ml-py
https://github.com/nicolargo/nvidia-ml-py3
https://github.com/petronny/nvsmi
https://github.com/fbcotter/py3nvml

I haven't been keeping up to date on this front though, so there may be a better approach!