With a nice latent space vis ;)

It has many limitations:

- no dynamic shapes
- hardcoded weights

But it should be possible to greatly improve this in the future.

Where `resnet18_stablehlo.mlir` is a model exported with:

Here's a small example:

The motivation for it comes from working on inegrating
Apple embedding atlas into ragnar: github.com/tidyverse/ra...

We ended up using the python package, but we could have used a #rstats / JS only solution if we had some nice mosaic integration.

Yep! We have some work in github.com/r-xla
It's a long term project though. Theoretically it's already possible to create a graph in jax , export it to stablehlo and ecute it in R with no python dep.

Happy to help making it work on MPS. It should just work if you create a tensor on the MPS device, using eg:

torch_randn(10,10, device="mps")

This is only happens on Linux when using the CPU. But many workflows involve transforming data on the CPU before executing the model on the GPU, so this fix should improve performance of most torch programs running on Linux!

But LibTorch is really optimized to work with Intel MKL, and indeed statically links to a version of MKL for faster math kernels, but we were not using it leaving a lot of performance on the table.

The problem was that previous versions of torch would use whatever BLAS library R is configured to use (which in most system will be a single threaded BLAS, and if you had some fun configuring R, it could be OpenBLAS.

Nice! Wrapping the bucket with a CDN, such as CloudFront, should also make downloads faster, due to regional caching and other optimizations that CDN's do. It also allows you to decouple the storage location from how things are distributed, so it gives you freedom to change storage structures later.

I know nothing about nix, but how do folks download from the cache later? In general using a CloudFront distribution instead of directly accessing the bucket will reduce a lot the egress costs.

After long investigation, I still don't understand what is really happening and what's wrong, but... The fix turned out being quite simple. Just had to add the RTLD_DEEPBIND flag when opening the dynamic library 🤷‍♂️

github.com/mlverse/torc...