py_smi

Convenient access to pynvml (the library behind nvidia-smi)

Installation

Install latest from pypi:

$ pip install python-smi

Links:

How to use

Here’s a quick demo of all the methods available:

from py_smi import NVML

nv = NVML()
nv.driver_version, nv.cuda_version

('535.183.06', '12.2')

All methods have a single parameter, which is the index of the GPU to get information about.

nv.info(0)

_Info(name='NVIDIA RTX A6000', serial='1322123048138', uuid='GPU-61e56e6f-2a64-c0f4-b26c-ab3ead0eed5b', persistence_mode=1, bus_id='00000000:01:00.0', display_active=0, performance_state=8, fan_speed=30, temperature=32, compute_mode=0)

[nv.mem(i) for i in range(3)]

[_Memory(free=2193.25, total=49140.0, used=46946.75),
 _Memory(free=48672.4375, total=49140.0, used=467.5625),
 _Memory(free=48672.4375, total=49140.0, used=467.5625)]

The index defaults to 0.

nv.utilization()

_Utilization(gpu=0, memory=0, enc=0, dec=0)

nv.power()

_Power(usage=17.22, limit=300.0)

nv.clocks()

_Clocks(graphics=0, sm=0, mem=405)

nv.pcie_throughput()

_PCIeThroughput(rx=0.0, tx=0.0)

nv.processes()

[_ProcessInfo(pid=201084, name='/home/jhoward/miniconda3/bin/python3.12', memory=46476.0)]

nv.dmon()

_DMon(pwr=17.039, gtemp=32, sm=0, mem=0, enc=0, dec=0, mclk=405, pclk=0)

Contributing

I’ve added the obvious pieces based on how I use nvidia-smi, but I’m sure there’s missing useful features, so PRs are welcome! Note that this is an nbdev project so the source notebooks must be changed, rather than editing .py or .md files directly.