Skip to content

Conversation

@CarloLucibello
Copy link
Member

@CarloLucibello CarloLucibello commented Dec 5, 2025

Fix #2625 while maintaining retrocompatibility, that is

using CUDA, Flux

is all you need for CUDA support, provided cuDNN is available in your environment.

This solution is still not orthodox, so I'm not sure it won't have bad interactions with other package extensions.

The other possibilities are:

  • Recommended to use using CUDA, cuDNN, Flux
  • Introduce an auxiliary package (FluxCUDA) requiring CUDA and cuDNN, so that we can do using FluxCUDA, Flux.

None of these 2 options would allow old code to keep working on julia 1.12 as this PR instead does.

See also JuliaLang/julia#60307

@CarloLucibello
Copy link
Member Author

@potaslab would you like to give this a try?

@CarloLucibello CarloLucibello merged commit 9bade5d into master Dec 7, 2025
7 of 11 checks passed
@mcabbott
Copy link
Member

mcabbott commented Dec 7, 2025

Thanks Carlo!

Sorry I meant to actually try, but if this try block works then it's very close in spirit to what we had, so seems like the right thing to do.

@mcabbott mcabbott deleted the cl/cudaext2 branch December 7, 2025 18:11
@CarloLucibello
Copy link
Member Author

I will do some more testing before tagging, but so far it seems identical to previous behavior

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Julia 1.12: Deadlock detected in loading Flux ext

3 participants