foldl2022

Gemma-4 E2B/E4B models reuses K-V cache from other layers, which do things in a "transposed" way: not reuse Q/K/V matrices within a single layer, but reuse across different layers.

So, show me the model weights, please.

A nice side-effect of "variables at the top": you keep your functions short.

Lazarus is still my favorite when developing desktop apps.

Language is not a problem. Pascal is just C/C++ in another favor. LCL/VCL is a wonderful library, everything just works like what I am expecting.

Tan looks like Pat. Amazing.

Just open-wights it. We need this. :)

Off the topic: I am also in favor of minimalist.

In the case of AlphaGeometry, I made AlphaGeometryRE to get rid of tensorflow/jax/flax, and 100+ Python packages.

Author here. It took quite a lot of time and efforts to implement this old model. The situation is worsened by `meliad` framework. Glad to see that `meliad` is not active now.

Is one ESP32 enough?

Glad that Nim had solved this problem with a macro `capture`.

It's up to the developers to capture loop variables or not.