Update feedstock to use cirun-openstack-cpu-large with Cirun

[conda-forge-admin]

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.

• Maintainers have accepted the terms of service and privacy policy
  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.

Closes #19

[conda-forge-webservices]

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 (recipe) and found it was in an excellent condition.

[jaimergp]

@RaulPPelaez, push an empty commit here to test the Cirun triggers. Thanks!

[RaulPPelaez]

Seems to be running. Should I merge #19 here or the other way around?

[jaimergp]

Whatever you prefer, this PR is here for your convenience. Pick one and iterate there :)

[RaulPPelaez]

Also, note that rerendering with Github actions as CI provider must be done
locally in the future for this feedstock.

@jaimergp could you clarify what this means? I cannot longer call
@conda-forge-admin, please rerender
in this feedstock?

Also I cannot manage to make the CI run in #19 so I will move the commits here...

[github-actions]

Hi! This is the friendly automated conda-forge-webservice.

I tried to rerender for you, but it looks like there was nothing to do.

This message was generated by GitHub actions workflow run https://github.com/conda-forge/cutlass-feedstock/actions/runs/8018330283.

[github-actions]

Hi! This is the friendly automated conda-forge-webservice.

I tried to rerender for you, but it looks like there was nothing to do.

This message was generated by GitHub actions workflow run https://github.com/conda-forge/cutlass-feedstock/actions/runs/8018331935.

[h-vetinari]

Running into:

[ 72%] Linking CXX shared library lib/libcutlass.so
$BUILD_PREFIX/x86_64-conda-linux-gnu/bin/ld: cannot find -lcuda_driver: No such file or directory

I'm not sure how linking the driver would have worked on azure (though we also haven't been able to build anything >=v3.2), but by itself, this looks like we might have to use the gpu runner after all?

@jaimergp @jakirkham

[RaulPPelaez]

I believe it should be enough to add cuda-driver-dev as a dependency. Lets try

[jaimergp]

@jaimergp could you clarify what this means? I cannot longer call

Correct. GH prevents workflows from modifying other workflows in a PR for security reasons. Since cirun requires GHA workflows and rerendering happens via GHA jobs under the hood... we hit that limitation. We might come up with a workaround, but for now this is what we have. You'll need to rerender locally with conda-smithy rerender.

[RaulPPelaez]

Adding cuda driver was not enough. I can see it installed, but the same error arises.

[h-vetinari]

I believe it should be enough to add cuda-driver-dev as a dependency. Lets try

Neither cuda-driver-dev nor cuda-driver-dev_linux-64 contain libcuda_driver.so; IIUC, the actual driver is the only piece of the puzzle that we aren't able to distribute. CC @jakirkham @leofang @adibbley

If so, then I think we do need cirun-openstack-gpu-large after all (unless there's an option for cutlass not to directly link to the driver of course).

[leofang]

To the best of my knowledge, I am unaware of such thing as libcuda_driver.so. Doesn't seem to come from NVIDIA?

[leofang]

Is it coming from the CMake FindCUDAToolkit module?
https://cmake.org/cmake/help/latest/module/FindCUDAToolkit.html#cuda-driver-library

[leofang]

Haven't looked into how cutlass's CMake works, but this preprocessing log seems fishy

-- CUDART: Not Found
-- CUDA Driver: Not Found
-- NVRTC: Not Found

[leofang]

This seems to be correct to me, if you need the header and to link to libcuda.so.

[h-vetinari]

Thanks for your input @leofang!

To the best of my knowledge, I am unaware of such thing as libcuda_driver.so. Doesn't seem to come from NVIDIA?

Well, cutlass if from NVIDIA, so in some way yes. Looking at the source upstream now, I see that cutlass does a sort of CMake-wrapper around the driver library.

It looks like we need to teach the search just above to look in the right place for finding libcuda.so, then the rest should work.

[RaulPPelaez]

Actually I am with @leofang here, wth is libcuda_driver.so? I cannot find references to it online.

Maybe setting CUDA_TOOLKIT_ROOT_DIR to CONDA_PREFIX?

[h-vetinari]

Actually I am with @leofang here, wth is libcuda_driver.so? I cannot find references to it online.

Click the first link in my comment

[leofang]

Again, NVIDIA does not ship libcuda_driver.so. If this is from CUTLASS CMake, you need to ask CUTLASS folks about how to patch/fix it.

[leofang]

Thanks very much for all the efforts and sorry for the struggling. Now that we restored cutlass as a header-only package (#22) and also released 3.3.0 (#24), this PR is no longer needed. Let me proceed and close it.