Restart CUDA 12 migration with CUDA arch builds #4799
Conversation
As all CUDA Linux builds are now on equal footing, fold the CUDA 12 arch migration into the CUDA 12 migration and restart it. That way the latest migrator can be retrieved.
|
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 ( |
|
Isn't this still dependent on having had the arch migration on that feedstock first? Or have all CUDA feedstocks passed that migration already? |
|
During the core meeting on 2023 July 26, we discussed how CUDA arch libraries make use of a newer GLIBC. In particular some libraries use symbols from GLIBC versions newer than 2.17. However GLIBC 2.28 support in conda-forge is still under discussion ( conda-forge/conda-forge.github.io#1941 ) That said, one of the points brought up in the meeting is that from a conda-forge build perspective, mainly we are interested to know the GLIBC usage of static libraries that are linked in builds. With CUDA, this is Have since done this and reported it in the last conda-forge meeting. Only saw GLIBC 2.17 symbols in use. Am including the results below for completeness: $ conda create -n cuda cuda-cudart-dev -y
Channels:
- conda-forge
Platform: linux-aarch64
Collecting package metadata (repodata.json): done
Solving environment: done
## Package Plan ##
environment location: /opt/conda/envs/cuda
added / updated specs:
- cuda-cudart-dev
The following packages will be downloaded:
package | build
---------------------------|-----------------
_openmp_mutex-4.5 | 2_gnu 23 KB conda-forge
cuda-cccl_linux-aarch64-12.0.90| h579c4fd_1 1.1 MB conda-forge
cuda-cudart-12.0.107 | hac28a21_6 21 KB conda-forge
cuda-cudart-dev-12.0.107 | hac28a21_6 22 KB conda-forge
cuda-cudart-dev_linux-aarch64-12.0.107| hac28a21_6 317 KB conda-forge
cuda-cudart-static-12.0.107| hac28a21_6 22 KB conda-forge
cuda-cudart-static_linux-aarch64-12.0.107| hac28a21_6 641 KB conda-forge
cuda-cudart_linux-aarch64-12.0.107| hac28a21_6 180 KB conda-forge
cuda-version-12.0 | hffde075_2 20 KB conda-forge
------------------------------------------------------------
Total: 2.3 MB
The following NEW packages will be INSTALLED:
_openmp_mutex conda-forge/linux-aarch64::_openmp_mutex-4.5-2_gnu
cuda-cccl_linux-a~ conda-forge/noarch::cuda-cccl_linux-aarch64-12.0.90-h579c4fd_1
cuda-cudart conda-forge/linux-aarch64::cuda-cudart-12.0.107-hac28a21_6
cuda-cudart-dev conda-forge/linux-aarch64::cuda-cudart-dev-12.0.107-hac28a21_6
cuda-cudart-dev_l~ conda-forge/noarch::cuda-cudart-dev_linux-aarch64-12.0.107-hac28a21_6
cuda-cudart-static conda-forge/linux-aarch64::cuda-cudart-static-12.0.107-hac28a21_6
cuda-cudart-stati~ conda-forge/noarch::cuda-cudart-static_linux-aarch64-12.0.107-hac28a21_6
cuda-cudart_linux~ conda-forge/noarch::cuda-cudart_linux-aarch64-12.0.107-hac28a21_6
cuda-version conda-forge/noarch::cuda-version-12.0-hffde075_2
libgcc-ng conda-forge/linux-aarch64::libgcc-ng-13.1.0-h2b4548d_0
libgomp conda-forge/linux-aarch64::libgomp-13.1.0-h2b4548d_0
libstdcxx-ng conda-forge/linux-aarch64::libstdcxx-ng-13.1.0-h452befe_0
Downloading and Extracting Packages
Preparing transaction: done
Verifying transaction: done
Executing transaction: done
#
# To activate this environment, use
#
# $ conda activate cuda
#
# To deactivate an active environment, use
#
# $ conda deactivate
$ strings /opt/conda/envs/cuda/targets/sbsa-linux/lib/libcudart_static.a | grep GLIBC
GLIBC_2.17
GLIBC_2.17 |
If the arch migration has not occurred on a feedstock (either via the bot or a maintainer manually enabling those builds), then the CUDA 12 migrator will not add those jobs If the the arch migration has run (or is run later), then the CUDA 12 arch builds will be added IOW they are orthogonal Edit: Here's an example with a feedstock where arch builds are disabled, the CUDA 12 migration is run, and then arch builds are enabled ( conda-forge/nccl-feedstock#93 ). Note CUDA 12 arch builds are not added until arch builds are enabled |
|
@conda-forge/core, would be interested to hear your thoughts on this |
| migration_number: | ||
| 1 | ||
| 2 |
There was a problem hiding this comment.
Need to double check this doesn't drop the migrator in migrated feedstocks
There was a problem hiding this comment.
IIUC this code path in conda-smithy continues to use the migrator already in a feedstock when a new migration_number occurs. IOW it will wait until the migrator bot add the new migration file (or a user adds it manually)
Removal doesn't occur as long as migrator_ts is the same, which is the case here
There was a problem hiding this comment.
Also IIUC the migrators used by conda-smithy can be overridden by passing --exclusive-config-file during rerender. So tried this with the ucx-split-feedstock using latest changes and found no change. So it looks like this is not an issue
Edit: Should add can see this migration file being used based on log lines that look like this pointing to the local copy of the migrator
INFO:conda_smithy.configure_feedstock:Applying migrations: /Users/jkirkham/Developer/conda-forge/feedstocks/feedstocks/ucx-split/.ci_support/migrations/cuda_112_ppc64le_aarch64.yaml,/Users/jkirkham/Developer/conda-forge/feedstocks/feedstocks/ucx-split/.ci_support/migrations/cuda120.yaml
There was a problem hiding this comment.
Also have done a few re-renderings myself and with the bot after this was merged and everything seems to be working correctly
Since `cudnn` is no longer in `zip_keys` with `cuda_compiler_version` and the version of `cudnn` used by CUDA 12 is the same as CUDA 11, go ahead and drop `cudnn` from the CUDA 12 migrator. It is not actually needed or used here.
|
Based on discussion in the conda-forge meeting today, it sounds like we are ok going ahead with this. Only item was to check on the behavior of Given this going to go ahead and get this started. Happy to follow up on anything else as needed. Thanks all! 🙏 |
|
Posted an update about this in the conda-forge CUDA 12 bringup issue: conda-forge/conda-forge.github.io#1963 (comment) |
As all CUDA Linux builds are now on equal footing, fold the CUDA 12 arch migration into the CUDA 12 migration and restart it. That way the latest migrator can be retrieved.
Checklist
0(if the version changed)conda-smithy(Use the phrase@conda-forge-admin, please rerenderin a comment in this PR for automated rerendering)