-
Notifications
You must be signed in to change notification settings - Fork 221
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
CUDA.jl v4 installs CUDA runtime despite version=local #1922
Comments
I think this is a limitation of Preferences; the package should be listed as a dependency somehow. Preferences can be layered though, so it should be possible to provide one in an upper depot. It should also be possible to set the dep in the |
Currently we only warn about this on CUDA.jl/lib/cudadrv/version.jl Line 59 in d3b1363
Should we encourage users to add "CUDA_Runtime_jll" to their environment in general? People can use |
There are two different kinds of "add to environment". If you use the Preference API it will automatically add it to I tried to give a bit off guidance on this here https://juliaparallel.org/tutorials/preferences/ |
The call to Preferences above should automatically add it to the It's true that for some users, artifacts will have been downloaded by then. Truly lazy artifacts would be a nice to have, but would need to be added to JLLWrappers (AFAIU the current architecture doesn't allow that). Hopefully you can avoid that on clusters by providing a global preference. Closing this because I don't think this a CUDA.jl bug / something we should change here, but feel free to keep the discussion alive if anything is unclear. |
Oh, sure - thanks for the help! @vchuravy regarding Preferences.jl - it would be nice being able to provide setup scripts to users for specific systems that need a local CUDA (without downloading a whole CUDA runtime first). Is there a way to make Pkg add a package ("CUDA_Runtime_jll" in this case) under It's kind of a chicken and egg problem -
|
File an issue with Preferences.jl? Should be easy enough https://github.com/JuliaPackaging/Preferences.jl/blob/e8082bc48a40dc27ed427e5ac9d17398d9949a14/src/Preferences.jl#L256-L264 (and it's not a stdlib so we can update it easily) |
Done. :-) |
This seems to happen even with Julia v1.9, but only on certain systems. For example on ORNL Ascent, with Julia v1.9.0, CUDA.jl insists on installing the CUDA runtime despite an "environments/v1.9/LocalPreferences.toml"
if "CUDA_Runtime_jll" is not part of the environment. So with
This works as it should (uses local CUDA)
But with (no explicit "CUDA_Runtime_jll") instead
CUDA.jl immediately installs it's own runtime (tested with a fresh Julia depot multiple times).
CC @williamfgc , I ran into this during the ORNL Julia HPC training session today.
The text was updated successfully, but these errors were encountered: