Possible collaboration between GONNX and gomlx?

[Swopper050]

Hi there @janpfeifer!

Awesome work on gomlx, amazing! I just discovered your package a few days ago.
Some time ago, together with a colleague of mine we built gonnx.
We were missing a package that could run our ONNX models in Go (with GRU and LSTM layers). Gonnx is by no means complete, and still missing operators and different operator sets, but I thought maybe it is interesting to look if there is a way of combining the packages, or in some way share implementation?

Our backend currently is build upon Tensor.

What would this package mostly need now? Maybe we could try combining some efforts? What do you think?

[janpfeifer]

hey Bram, I hadn't seen gonnx looks very cool!

Funny that one of my next projects is going to add Gorgonia's Tensor library as a backend for GoMLX :) While XLA is immensely faster (desirable when training or larger models), it has this huge inconvenience of having C++ libraries that needs installing, and it doesn't support Darwin (it's broken/missing pieces). So I urgently need to add a GoMLX backend using Tensor.

But one of the main reasons we (with folks from KnightsAnalytics conceived onnx-gomlx was to be able to import a ONNX model and then further fine-tune them -- GoMLX's autodiff/training support makes that relatively easy.

Hey, you be up for a sync up ? Maybe we could brainstorm where to collaborate ?

[Swopper050]

Yeah sure! I'd also contacted @wipsel (co-author of Gonnx) and he was also open for a brainstorm, I'll drop you an email.