-
-
Notifications
You must be signed in to change notification settings - Fork 5
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
Update feedstock to use cirun-openstack-gpu-large with Cirun #14
Update feedstock to use cirun-openstack-gpu-large with Cirun #14
Conversation
…nda-forge-pinning 2023.11.29.17.12.15
Hi! This is the friendly automated conda-forge-linting service. I just wanted to let you know that I linted all conda-recipes in your PR ( |
Please don't merge until I have adjusted the build scripts on all platforms to target all of the CUDA archs. This PR may remain open for a few weeks until I have time to do that. |
I have updated the build scripts to target all of the archs, so it's OK by me to merge this PR. However, Cirun does not succeed.
I assume @isuruf et al are still shaking out the bugs. Please let me know if I need to do something. |
CUDA 12 failing with:
|
I'm pretty sure that the errors for CUDA 11.x are because the shared library is too large when it contains byte code for all of the archs. I remember running into this error before when I first took over building this library. https://stackoverflow.com/a/47168086/4459405 We can either try adding |
I need to take a look into the roles error, sorry 🙏 |
Because the builds succeeded with fewer archs, and the fact that I know libmagma is already around 2GB in size, I accept that as evidence that the build error was due to the symbols not fitting into the default address space size of 2GB. We can either build for fewer archs or enable the larger memory space. IMO, the library is already quite large and the impact of building for more minor archs is probably not worth potential issues from increasing the address space. |
export CUDA_ARCH_LIST="${CUDA_ARCH_LIST},sm_35" | ||
export CUDAARCHS="${CUDAARCHS};35-virtual;80-virtual" | ||
export CUDA_ARCH_LIST="${CUDA_ARCH_LIST},sm_35,sm_86" | ||
export CUDAARCHS="${CUDAARCHS};35-real;86" |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Why this change? You need the virtual
arch for newer GPUs to get better performance.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
For CMAKE, 86
is an abbreviation of 86-real;86-virtual
.
With the extra build time, the list of target archs for 11.2 has expanded from
35-virual;50-real;60-real;70-real;75-real;80-real;80-virtual
to
35-real;50-real;60-real;70-real;75-real;80-real;86-real;86-virtual
35 was converted from virtual to real. 80 virtual was dropped, but 86 was added for both virtual and real.
Need to revert the Windows build scripts because Cirun is not offering Windows runners. |
Note that only builds triggered by maintainers of the feedstock (and core)
who have accepted the terms of service and privacy policy will run
on Github actions via Cirun.
at https://github.com/Quansight/open-gpu-server
Also, note that rerendering with Github actions as CI provider must be done
locally in the future for this feedstock.