search

rapidsai / jupyterlab-nvdashboard

A JupyterLab extension for displaying dashboards of GPU usage.
BSD 3-Clause "New" or "Revised" License
598 stars 78 forks source link

Update deprecated NVML API calls for NVLink query #133

Closed pentschev closed 2 years ago

pentschev commented 2 years ago

Various of the NVML API calls to query for NVLink status have been deprecated since the CUDA 11.0 release. Resetting counters is also not allowed anymore.

New results:

Screenshot 2022-06-20 at 20 07 53

Fixes #28, #114

pentschev commented 2 years ago

Do you think this is ready to merge?

IMO, yes.

More specifically, (1) are you reasonably confident that things are working,

I did test this on my workstation at home (see screenshot) and on a DGX-1, both seemed to work and I didn't notice any issues.

and (2) do you think there is any danger in dropping support for cuda<11 here?

I think not. RAPIDS dropped support for CUDA 10.x quite some time ago, right now I can't even test CUDA 10.x anymore, since all machines I have access to are on CUDA 11.x. Therefore, it seems to me like the right way forward to drop 10.x and follow more closely what RAPIDS is doing, unless there are clear use cases where we need to differ and someone can actually maintain legacy code.

pentschev commented 2 years ago

Thanks @rjzamora for reviewing and merging! 🙂