scalene
scalene copied to clipboard
plasma-umass
Could this package work with newer versions of nvidia-ml-py?
I see from looking at previous issues that there was an issue with certain versions of nvidia-ml-py causing a problem because of a use of a print statement instead of a function. However, has anyone checked to see if newer versions of nvidia-ml-py still have this problem?
The reason I ask is that it looks like the nvida-ml-py package is being maintained more regularly than pynvml at this point (https://pypi.org/project/nvidia-ml-py/#history vs https://pypi.org/project/pynvml/#history), and I'm trying to set up an environment primarily using conda, which explicitly blocks installation of both pynvml and nvidia-ml-py (whereas I believe pip would just have the later install clobber the prior install).
Thanks for letting us know; it had for long time seemed like nvidia-ml-py was in effect abandonware. IIRC the API is the same, but I have no idea if they have fixed the previous issues.
I just tested this and it appears to work fine. Can you verify on your end? (Just uninstall pynvml)
Another impact of depending on pynvml instead of nvidia-ml-py is that the latest versions of scalene cant be installed together with ray because ray depends on gpustat that depends on nvidia-ml-py that cant be installed together with pynvml
For example, creating a simple environment using conda create -n test_env python=3.9.18 ray-default=2.8.1 scalene=1.5.31.1 fails.
Actually, at the moment, pynvml has not been updated in a while, but nvidia-ml-py is getting regular updates.
@emeryberger Any thoughts about reverting this package to use nvidia-ml-py at this point? As a couple of us have mentioned, it seems to be getting updated much more regularly than pynvml as of now.
As of the latest release (1.5.42), Scalene is back to using nvidia-ml-py.
@emeryberger I just noticed that both nvidia-ml-py and nvidia-ml-py3 are specified as dependencies now. Why is that? From PyPi, it looks like nvidia-ml-py3 only had a single release 7 years ago. It seems that only nvidia-ml-py should be a dependency.
@emeryberger Sorry to bother you, but could you release 1.5.42.3 (ideally to both PyPi and conda-forge) without the nvidia-ml-py3 dependency? Thanks.