Moving from pynvml to nvidia-ml-py

[jakirkham]

Currently this is making use of pynvml in a few places:

deployment/source/guides/mig.md

Line 33 in 3b6a991

Alternatively, MIG instance names can be obtained programatically using [NVML](https://developer.nvidia.com/nvidia-management-library-nvml) or [PyNVML](https://github.com/gpuopenanalytics/pynvml). Please refer to the [NVML API](https://docs.nvidia.com/deploy/nvml-api/) to write appropriate utilities for that purpose.

deployment/source/cloud/azure/azure-vm-multi.md

Lines 57 to 60 in 3b6a991

	import pynvml
	pynvml.nvmlInit()
	return pynvml.nvmlDeviceGetName(pynvml.nvmlDeviceGetHandleByIndex(0))

However we would like to move to nvidia-ml-py in the future. Raising this issue to track this work

[hcho3]

@jakirkham Is there a Conda package for nvidia-ml-py ?

[jakirkham]

@hcho3 sorry for missing your question 🤦‍♂️

Yeah here's the feedstock in conda-forge

[jacobtomlinson]

The APIs of pynvml and nvidia-ml-py should be identical(import pynvml; ...) with the exception of pynvml.smi so this should be a quick change.

To close this issue we need to:

• Check if we install pynvml anywhere and replace it will nvidia-ml-py
• Check if we link to pynvml anywhere and replace it with a link to nvidia-ml-py
• Check if we use pynvml.smi anywhere and remove it as that submodule is going away
• Check any import pynvml; ... code and ensure they still work as expected with nvidia-ml-py installed instead

[jakirkham]

As part of RAPIDS 25.02, we have upgraded to pynvml version 12, which requires/uses nvidia-ml-py under-the-hood

It would be good to recheck the deployment docs and examples work with it