On Saturday, 21 February 2026 at 12:09:52 UTC, Sergey wrote:
On Saturday, 21 February 2026 at 10:56:39 UTC, Danny Arends
wrote:
I've finished on a port of Andrej Karpathy's microgpt.py to
the D programming language. For those unfamiliar, it is a
minimal, educational GPT implementation, great for
understanding transformer architecture.
The D port stays faithful to the original: same architecture,
same training logic. Just D instead of Python. D's operator
overloading and struct semantics make it a surprisingly
natural fit for this kind of low-level ML work.
Find it here:
https://gist.github.com/DannyArends/12704c9207797a64338a5be4f1010bcf
Nice. Thanks
The 3rd one :) Will check for some improvements in mine
https://github.com/cyrusmsk/microDpt
https://github.com/denizzzka/microgpt_dlang (WIP)
Lol, yeah guess everyone want to learn by doing... Looking at the
code, mine is quite a bit shorter. Guess using a more functional
approach with zip, map, fold, joiner et al. makes a big impact in
how compact you can get it.
I also got a basic ggml version working on Linux that uses a
ggml_tensor* for computation. But for some reason it doesn't work
properly under windows (yet).
