Microgpt in pure C hits 10M tps on Apple m5

The most atomic way to train and inference a GPT in pure, dependency-free C.

A character-level transformer with forward pass, backprop, Adam and sampling, in one C file with nothing beyond libc. It trains on ~32k names in a couple of seconds and generates new ones.

make run

Or run it directly, on any corpus with one item per line:

./microgpt data/names.txt

Builds on macOS, Linux and Windows (MSYS2), on ARM64 with NEON and x86-64 with AVX2. The Makefile picks the flags for the host.

step 5000 / 20000 | loss 2.6036  (avg 2.2940)
step 10000 / 20000 | loss 1.9639  (avg 2.2564)
step 15000 / 20000 | loss 2.7007  (avg 2.2151)
step 20000 / 20000 | loss 2.3463  (avg 2.2201)

inference
sample  1: kayley
sample  2: maria
sample  3: arana
sample  4: shayan
sample  5: jayden
sample  6: saria
sample  7: kaylen
sample  8: amari
sample  9: alina
sample 10: mailyn
  c fp32+NEON       10168430 tok/sec

The model has 4192 parameters and generalises rather than memorises. Trained on 20000 of the 32033 names, it scores 2.2054 nats per character on those and 2.2039 on the 12033 it never saw, beating an interpolated trigram that has nearly five times as many parameters.

Training and inference use separate forward passes. gpt_forward

stores activations for backprop; gpt_forward_infer

is a specialised single-token path whose logits match it to within fp32 rounding. docs/PERFORMANCE.md covers how that path works and what limits it.