Enable cuFile on ARM #405
Enable cuFile on ARM #405
Conversation
jakirkham
added
improvement
jakirkham
force-pushed
the
enable_arm_cufile
branch
from
71b7f2d
to
1c417d7
Compare
| @@ -4,7 +4,7 @@ files: | |||
| output: conda | |||
| matrix: | |||
| cuda: ["11.8", "12.5"] | |||
| arch: [x86_64] | |||
| arch: [aarch64, x86_64] | |||
There was a problem hiding this comment.
I would prefer to revert this. We don't supply ARM environment files for most RAPIDS repos. If you'd like to add ARM conda envs for all of RAPIDS, I would do that in a separate PR and track the effort in rapidsai/build-planning for all repos.
| arch: [aarch64, x86_64] | |
| arch: [x86_64] |
There was a problem hiding this comment.
Given there are other users asking about ARM specifically with KvikIO (and running into issues without more explicit guidance), the environments provide a value add for them. For example: #393
Edit: That said, if there is a good way to drop CUDA 11.8 ARM, think that is worth doing as that is probably more confusing than helpful (especially given there is no cuFile for CUDA 11.8 ARM)
There was a problem hiding this comment.
I suppose it's fine to include this. Maybe this should be done for all of RAPIDS, with a build-planning issue for tracking.
We don't have a good way to exclude (CUDA 11 + ARM), unless you create a separate file key entry for (CUDA 12 + ARM) That's not worth doing imo.
As we see increased ARM usage, it'll be somewhat helpful to have these environment files readily accessible. I have had to explain how to generate the ARM environments with rapids-dependency-file-generator before and it's a little complicated for new DFG users.
There was a problem hiding this comment.
Agree this is a good idea
Have raised issue: rapidsai/build-planning#87
Please add anything additional to that
|
Thanks all! 🙏 Anything else needed here? |
|
/merge |
Fixes #339
Builds CUDA 12.2+ ARM packages with cuFile and adds cuFile as a dependency. Continues to build CUDA 11 ARM packages without cuFile.
As cuFile already has the needed logic to
dlopencuFile at runtime (with fallbacks for when it is missing), there shouldn't be any additional handling needed for legacy CUDA versions with ARM support.