SnesGPT, micro-GPT ported to ASM on the Super Nintendo
This stuffs a full decoder-only transformer (RMSNorm, ReLU^2 MLP, multi-head attention with KV cache) into a SNES ROM and makes the 3.58 MHz 65816 actually perform matrix math by using the PPU hardware multiplier and Q8.8 fixed-point math. The repo documents the pipeline end-to-end — training/export, ROM lookup tables, and an assembler build — so it reads like a curated engineering demo rather than a sketchy toy.
