CPU-only pypi and conda packages

[luweizheng]

Hi,

Is there a CPU-only pypi or conda package of ucxx. I used to install by conda install -c conda-forge -c rapidsai ucx-proc=*=cpu ucx ucx-py, and now I will move to ucxx.
I now can install by pip install ucxx-cu12 on a CPU node, and it seems that most of ucxx's functions work, but it has some warning of not having cuda devices. What's more, I am not sure if the performance of the GPU-compiled package is fast enough and whether I need to compile a CPU version.

If there is a CPU version pypi or conda package, it would be great.

[pentschev]

@luweizheng thanks for the report. One thing you don't need to do anymore is provide ucx-proc=*=cpu, since UCX 1.14 there's only one package that is supported both by CPU and CUDA, which is determined at runtime.

As for CPU packages of UCXX we still have to figure out what's possible both for conda and pip packages. In both cases the only dependency to CUDA is in libucxx (C++ implementation). @jameslamb do you have any sense of what would be necessary to provide libucxx without CUDA/RMM dependencies so it runs also on CPU-only systems both for conda and pip packages?

[jameslamb]

@jameslamb do you have any sense of what would be necessary to provide libucxx without CUDA/RMM dependencies so it runs also on CPU-only systems both for conda and pip packages?

I don't know that. I suspect we could borrow some ideas from there:

• cuml's CPU-only packages: https://github.com/rapidsai/cuml/tree/branch-24.10/conda/recipes/cuml-cpu
• the way that projects in the Legate family of packages support both GPU and CPU-only packages, e.g. https://github.com/rapidsai/legate-boost/blob/03f07bab038f718570947efeb72080c6b853d1de/conda/recipes/legate-boost/meta.yaml#L14-L27

[luweizheng]

Hoping there is a CPU-only package 🫡

[luweizheng]

Currently, the (v.0.40) conda package can run on a CPU-only computing node.