File Limit Request: libcuml-cu12 - 650 MiB

[raydouglass]

Project URL

https://pypi.org/project/libcuml-cu12/

Does this project already exist?

• Yes

New Limit

650 MiB

Update issue title

• I have updated the title.

Which indexes

PyPI

About the project

cuML is a suite of libraries that implement machine learning algorithms and mathematical primitives functions. cuML enables data scientists, researchers, and software engineers to run traditional tabular ML tasks on GPUs without going into the details of CUDA programming. In most cases, cuML's Python API matches the API from scikit-learn.

We have worked extensively to reduce the size of libcuml-cu12 wheels.

• Originally, wheels were over 1 GiB, but now down to ~560MiB
• Utilizing static linking for libcuml.so dependencies
• Enforcing size restrictions internally when adding new features or functionality
• Future plans to reduce the size even further with newer CUDA compiler features
• Future plans to remove or devendor libcuml dependencies

Reasons for the request

cuML has built and published wheels in some form to a private index (pypi.nvidia.com) since ~October 2022. Since then, more users and third-party projects are depending on cuml / libcuml now and it would significantly improve the developer experience if the wheels were available directly from PyPI.

Two wheels (one for each x86_64 & aarch64) will be published for each release every two months.

We have requested a slight buffer on the wheel size and our future plans (detailed) above should help reduce the size in the future.

Code of Conduct

• I agree to follow the PSF Code of Conduct

[Thespi-Brain]

Hi @raydouglass,

I am afraid the size limit request is a bit too high. I suggest for ML projects and libraries to use something like NTLK. If you can reduce the size request, let me know and we will go from there, thanks.

[raydouglass]

Thanks @Thespi-Brain, I actually no longer work for the company/project, so I tagging @gforsyth who took over this effort.

[gforsyth]

Hi @Thespi-Brain!

Thanks for taking the time to review this request!
I understand that it’s a large file-size, but NLTK isn’t really a viable replacement for what cuML does. NLTK is focused more on natural language processing, and also only runs on CPU.
cuML is a permissively licensed collection of ML algorithms and mathematical primitives that other developers can use to accelerate model training and inference using NVIDIA GPUs. It's like scikit-learn, but runs on the GPU.

It supports training accelerated regression and classification models, clustering, dimensionality reduction and manifold learning, time series models, and nearest-neighbors – none of which are offered by NLTK.

As Ray mentioned above, we’re continuing to try to reduce the size of our libraries and have strict procedures in place to avoid library size increasing.

[ewdurbin]

Hi @gforsyth can you provide a link to the existing built binaries for this project If they are currently hosted somewhere else? Or an explanation of the number of expected variants and their size. That will help us understand the number of variants and their average size in order to process this request.