[CUDA] drop CUDA 10 support, start supporting CUDA 12 (fixes #5789)#6099
[CUDA] drop CUDA 10 support, start supporting CUDA 12 (fixes #5789)#6099
Conversation
| --inspect \ | ||
| --ignore 'compiled-objects-have-debug-symbols,distro-too-large-compressed' \ | ||
| --max-allowed-size-uncompressed '60M' \ | ||
| --max-allowed-size-uncompressed '70M' \ |
There was a problem hiding this comment.
With generating code for 2 more CUDA architectures, the linux CUDA wheels are now around 63 MB uncompressed.
==================== running pydistcheck ====================
checking '/LightGBM/dist/lightgbm-4.1.0.99-py3-none-manylinux_2_27_x86_64.whl'
----- package inspection summary -----
file size
* compressed size: 40.6M
* uncompressed size: 63.3M
* compression space saving: 35.9%
contents
* directories: 0
* files: 17 (1 compiled)
size by extension
* .so - 64423.5K (99.3%)
* .py - 404.7K (0.6%)
* no-extension - 21.5K (0.0%)
* .txt - 0.0K (0.0%)
* .typed - 0.0K (0.0%)
Build link: https://github.com/microsoft/LightGBM/actions/runs/6186545556/job/16794443363?pr=6099
We could start dropping older architectures at some point if we think that's a problem, but I don't think that should even be considered until we start possibly distributing precompiled CUDA wheels via a repository like PyPI.
Until then, I think building for more architectures to support a wider range of GPUs is desirable, and a few more MB on disk for the wheel isn't a big concern.
|
I'd specifically like a review from @shiyu1994 before we merge this one, to be sure I'm not missing something. I don't know a lot about CUDA 😅 |
|
@jameslamb Thanks for working on this! This is very helpful. We definitely need to support CUDA 12. I'll review this in the next few days. |
|
Sure no problem, take your time! |
|
thanks @shiyu1994 ! |
|
This pull request has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this. |
Fixes #5789.
Notes for Reviewers
CUDA 12.2.2 is the latest release of the CUDA toolkit (https://developer.nvidia.com/cuda-toolkit-archive), but the latest official Docker image is 12.2.0 (https://catalog.ngc.nvidia.com/orgs/nvidia/containers/cuda), so I made the new job here
12.2.0.I think it's important the
method: sourceCI job be tested against the latest version of CUDA, since as of now installing from source is the only way users can work with a CUDA-enabled version of the Python package.