Thanks Jose, Stuart and Niklas for all the helpful and informative suggestions. I do have a preference for CPU-based training with giant batch sizes so KANN looks like a great streamlined library to try.

I will certainly take a look at Kompute also, looks interesting and I have a spare GPU for experimentation. It's encouraging to hear about OpenBSD Vulkan support.

I have made good progress compiling PyTorch with CPU support also, so will keep you updated on my progress.

Frank

On Wednesday, July 17th, 2024 at 10:59 PM, Stuart Henderson <stu@spacehopper.org> wrote:

> On 2024/07/17 09:50, Frank Baster wrote:
> 
> > Is there any interest in machine learning ports for PyTorch / JAX / TensorFlow? I had a quick
> > look at mlpack but couldn't find support for recurrent neural networks.
> > 
> > PyTorch looks like the most do-able.
> 
> 
> OpenBSD is probably not a great target for those, the GPU-based code
> seems to just be for Linux/Windows (and I think needs binary-only
> drivers).
> 
> KANN would probably be a better fit to OpenBSD. It's CPU-only everywhere
> so we're not at such a disadvantage compared to other OS, and much
> lighter weight. There's no standalone port but fwiw it is used as part
> of rspamd via a vendored copy - indeed KANN upstream recommends that
> it's used that way.