RthItalia

/

Rth-lm-25b

Like 3

Hi Matthias, Thanks for the flag! You were right—the custom kernels were missing.

I've just pushed the full C++ implementation of the Fractal TCN operators (
rth_tcn_ops.cpp
) and the GGUF converter to our GitHub. You can find the patches and build instructions here: https://github.com/rthgit/ZetaGrid

Since this is a novel architecture (pure gated convolutions, no attention), your feedback would be incredibly valuable! If you manage to build it or have any thoughts on the kernels, please let me know here or open an issue on GitHub. (And if you like the project, a Star on the repo is always appreciated! ⭐)

Cheers!

Thanks. Will look into it soon(ish).