-
Notifications
You must be signed in to change notification settings - Fork 915
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
[FEA] CEC for CUDA 12: using nvjitlink (instead of ptxcompiler / cubinlinker) #12822
Comments
In my understanding, we have (at least) two options:
Happy to be corrected if anything above is inaccurate. |
My understanding was that this wouldn't be a build-time issue at all, but that numba would "just work" when using CUDA 12 JITted kernels across different minor CUDA versions (probably by using nvJitLink under the hood). I thought that all cudf would need to do is to make the cubinlinker and ptxcompiler requirements conditional on the CUDA version. |
The only issue I encountered when test-driving CUDA 12 pip wheel builds for cuDF (i.e. cudf-cu12, dask-cudf-cu12) was in the test step, where the install of cudf-cu12 couldn't find ptxcompiler-cu12, cubinlinker-cu12 to install from its requirements. So, if the only change for CUDA 12 is to delete those lines (perhaps using |
If simply removing the requirements makes everything work then yeah we could use that script (although it wouldn't be enough by itself, we would also need to do something for conda packages. |
Conda packages are easier because we can template the |
Yup I agree, it's not hard. Just one more thing to do in addition to what we'd do for wheels. |
cc @gmarkall |
(bump) |
Short term proposal for 23.06: #13339 |
Have tried to retitle this issue to make it clearer to others. Please feel free to revise if there is a better way to capture the intent here |
Fixes #12822 This PR provides minor version compatibility in the CUDA 12.x range through `nvjitlink` via the preliminary [nvjiitlink python binding](https://github.com/gmarkall/nvjitlink). Thus far this PR merely leverages a local installation of the library and should not be merged until `nvjitlink` is hosted on `conda-forge` and cuDF's dependencies are adjusted accordingly, likely as part of this PR. Authors: - https://github.com/brandon-b-miller - Ashwin Srinath (https://github.com/shwina) Approvers: - Bradley Dice (https://github.com/bdice) - Ashwin Srinath (https://github.com/shwina) URL: #13650
cuDF uses ptxcompiler and cubinlinker for CEC. Will this be changing for CUDA 12? We should discuss.
@bdice @gmarkall @vyasr
The text was updated successfully, but these errors were encountered: