Update conda recipes for Enhanced Compatibility effort#893
Update conda recipes for Enhanced Compatibility effort#893rapids-bot[bot] merged 8 commits intorapidsai:branch-21.12from
conda recipes for Enhanced Compatibility effort#893Conversation
|
rerun tests |
1 similar comment
|
rerun tests |
|
Depends on #901. Keeping as a draft until that is complete. We'll probably see lots of failures until that PR is merged. |
fa82573 to
c6498fa
Compare
c6498fa to
b95ccae
Compare
`conda` seems to only build multiple variants if the variant key is explicity used in the recipe. (i.e. not via Jinja)
|
rerun tests |
|
After reviewing the CI logs, it seems that these changes are working correctly. |
|
@gpucibot merge |
| - numba >=0.49 | ||
| - numpy | ||
| - {{ pin_compatible('cudatoolkit', max_pin='x.x') }} | ||
| - {{ pin_compatible('cudatoolkit', max_pin='x', min_pin='x') }} |
There was a problem hiding this comment.
In conda-forge we decided that the min_pin should be x.x. Please see comment ( conda-forge/nvcc-feedstock#71 (comment) ) and subsequent discussion for context.
There was a problem hiding this comment.
Hmm. I'm confused how this would work? If we set min_pin to x.x, then wouldn't our pins effectively become cudatoolkit>=11.5,<12? How would this work for users that use cudatoolkit=11.{0,2,4}?
This is all assuming that we're only going to build packages with CUDA 11.5, which is what I thought the plan was, no?
After merging rapidsai#893, there are now 2 `librmm` `conda` packages being published instead of 1. Therefore the upload script needs to be updated accordingly. Skipping CI since the upload script isn't run on PRs anyway.
After merging #893, there are now 2 `librmm` `conda` packages being published instead of 1. Therefore the upload script needs to be updated accordingly. Skipping CI since the upload script isn't run on PRs anyway. Authors: - AJ Schmidt (https://github.com/ajschmidt8) Approvers: - Jordan Jacobelli (https://github.com/Ethyling) URL: #909
This PR tweaks the changes from #893 and #909 so that `rmm` produces two packages (`has_cma` and `no_cma`) instead of `librmm`. Authors: - AJ Schmidt (https://github.com/ajschmidt8) Approvers: - Jordan Jacobelli (https://github.com/Ethyling) URL: #910
This PR updates the
condarecipe build strings andcudatoolkitversion specifications as part of the Enhanced Compatibility efforts.rmmChangesThe build strings in the
condarecipe have been updated to only include the major CUDA version (i.e.librmm-21.12.00a-cuda11_gc781527_12.tar.bz2) and thecudatoolkitversion specifications will now be formatted likecudatoolkit >=x,<y.0a0(i.e.cudatoolkit >=11,<12.0a0).Moving forward, we'll build the packages with a single CUDA version (i.e.
11.5) and test them in environments with different CUDA versions (i.e.11.0,11.2,11.4, etc.).librmmChangesA
conda_build_config.yamlfile has been added to thelibrmmrecipe folder so that two variants oflibrmmare built: one with and one withoutcudaMallocAsyncsupport. A new environment variable,BUILD_FLAGS, is passed through toconda/recipes/librmm/build.shand is set according to thecudaMallocAsyncvariant value in the recipe. Finally, a build string modifier of eitherno_cmaorhas_cmais appended to the build string which is used to determine which package should be installed inci/gpu/build.sh.