Skip to content
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

Windows Github actions / workflow is successfully building including Cuda 12.4 builds #229

Closed
rosslwheeler opened this issue Apr 23, 2024 · 12 comments

Comments

@rosslwheeler
Copy link
Contributor

No issue - just a heads up. The Windows builds with GitHub actions is working. Adding the tests is next and then I'll be merging those changes into the ci.yml file (@Ricardicus). The only downside is that the current Cuda installation for Windows is slow but at least it's the latest (12.4) version. Be good if we can get containers with current 12.4 Cuda toolkits for all 3 OS's - that would speed things up. Next up is updating the Makefile to include the Windows build changes. These Makefile changes will be much smaller than in prior PRs. This will all be documented too...

@Ricardicus
Copy link
Contributor

Ricardicus commented Apr 23, 2024

Nice. Perhaps add a separate job for this, there are two jobs now:

  • build-and-test-cpu, runs ubuntu that builds with and without OMP (build and test)
  • build-with-cuda, runs in an nvidia container that has cuda-tools installed (just building)

Since Windows is such a world of its own I suggest it runs as an additional job, but that is perhaps already
what you had in mind. "build-and-test-cpu-mvcc" ?

@rosslwheeler
Copy link
Contributor Author

Yep, that's the plan. Be nice if we could find some self-hosted Cuda runners that someone could let us use???

@Ricardicus
Copy link
Contributor

Yes, it would be very nice to have a container with a virtual driver installed to run here on github. I can look into that.

@Ricardicus
Copy link
Contributor

Yeah so the virtual driver requires a purchase so I don't think we can use that for this. Otherwise there are other solutions for containers but they all rely on installations on the host which I don't think we can do here on Github.

@rosslwheeler
Copy link
Contributor Author

@Ricardicus - as soon as this (#236) gets in, I'll merge the CI changes in and have you take a look.

@rosslwheeler
Copy link
Contributor Author

rosslwheeler commented Apr 24, 2024

@Ricardicus - for the windows ci build - one possible problem is the Cuda Toolkit installation speed (it's ~10 minutes) for the 12.4 Cuda SDK running from a github action. One option is to skip the CI cuda tests (and the Toolkit installation) for windows. Thoughts?

@Ricardicus
Copy link
Contributor

Windows build with CUDA support is very interesting I think, without CUDA support is not that interesting.
It seems we are not alone in this experience, there is this github action already developed that have this experience:
Jimver/cuda-toolkit#253
I looked around and found it used in another similar project, https://github.com/rustformers/llm , and looking
at the logs their build on windows takes ~30 min.
I guess this will be up to Andrej, what he prefers.

@rosslwheeler
Copy link
Contributor Author

Run it nightly? Or periodically during the day? We could catch problems at least. Might not be super timely but could work short-term.

The jimver GHA doesn't work on Windows 2022 so I ended up creating my own Cuda install GHA. The advantage of this is that we can optimize it since it lets us control the packages that are installed. I assume the best we can do is around 5 minutes though.

@rosslwheeler
Copy link
Contributor Author

@Ricardicus - can you review this PR #320? I purposely removed the Cuda build/tests until I can get them faster in CI. Let me know if this file structure is okay.

@rosslwheeler
Copy link
Contributor Author

Also, I have the build working in GitHub actions with MSBuild and the Solution/Project files. Probably won't be checking this into the main repo but have it in case we want it later.

@rosslwheeler
Copy link
Contributor Author

@Ricardicus - got the windows cuda builds down from 20+ minutes to 2 minutes. See below. Still working on finding us a self-hosted runner. #320 is ready to be merged.

image

@rosslwheeler
Copy link
Contributor Author

This has been merged - closing.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants