Bug#982679: RFP: keops -- Kernel Operations on the GPU, with autodiff, without memory overflows
Package: wnpp
Severity: wishlist
* Package name : keops
Version : 1.4.2
* URL : https://github.com/getkeops/keops
* License : MIT
Programming Lang: C++, Cuda
Description : Kernel Operations on the GPU, with autodiff, without memory overflows
The KeOps library lets you compute generic reductions of very large
arrays whose entries are given by a mathematical formula. It combines a
tiled reduction scheme with an automatic differentiation engine, and can
be used through Matlab, Python (NumPy or PyTorch) or R backends. It is
perfectly suited to the computation of Kernel matrix-vector products and
the associated gradients, even when the full kernel matrix does not fit
into the GPU memory.
---
I am particularly interested in the python interface (PyKeOps).
While this package requires cuda so it can work on the GPU, it also has
a useful CPU mode that is usable in 100% free software. I don't know how
easy/hard it would be to split between main, contrib (and non-free?)
though.
The python GPU mode requires some suitable versions of pytorch (new
versions of pytorch sometimes break it), and a C++ compiler compatible
with cuda (which usually means 1 or 2 versions older than current in
debian), so it is helpful if packaging ensures a working combination of
packages.
I use pykeops directly, but it is also a dependency for other useful
packages like geomloss (http://www.kernel-operations.io/geomloss/).
I am not planning on packaging it myself, sorry.
Reply to: