Lock pynvml version below 12.0.0 for branch-0.9

[renweizhukov]

On JupyterLab v3.6.8, used conda to install the stable version:

conda install -c rapidsai -c conda-forge jupyterlab-nvdashboard

After launching jupyter lab, I got the following error:

AttributeError: module 'pynvml' has no attribute 'nvml'

Found that pynvml v12.0.0 was installed while requirements.txt sets pynvml>=11.0.0. To fix this, we either (1) set pynvml version as

pynvml>=11.0.0,<12.0.0

in requirements.txt or (2) set it in conda command as below:

conda install -c rapidsai -c conda-forge pynvml=11.5.3 jupyterlab-nvdashboard

[jakirkham]

Sorry to hear you ran into this issue

In the latest jupyterlab-nvdashboard release, 0.12.0, we pinned pynvml to 11 ( #210 )

However earlier versions of jupyterlab-nvdashboard do not have that pin

Without seeing more details from the install above, my guess is that conda is picking an older version of jupyterlab-nvdashboard that did not have the version constraint and installing it with pynvml

My current recommendation would be to use jupyterlab-nvdashboard>=0.12.0 when installing. That will ensure conda picks the latest version of jupyterlab-nvdashboard, which has the pynvml constraint. If that is not an option, your workaround also seems reasonable

We are also exploring setting up repodata patching for the rapidsai channel ( conda/infrastructure#1094 ), which would allow us to effectively apply this pynvml constraint to older versions of the jupyterlab-nvdashboard packages. However that is not setup yet

The next release of jupyterlab-nvdashboard version 0.13.0 will be compatible with pynvml ( #215 ). That will be released as part of the RAPIDS 25.02 release timeline

[renweizhukov]

Thank you for the quick response!

Probably because my JupyterLab was 3 < 4, conda picked up jupyterlab-nvdashboard v0.9.0 instead of >=0.12.0.

Also thank you for the recommendation! Will try it when my workaround stops working :)