minimal transformer inference in c. ~400 lines total.

make
./download.sh
./tinyllm models/stories15M.bin "Once upon a time"

./tinyllm <model.bin> [prompt] [-t temp] [-p topp] [-n steps] [-s seed] [-z tokenizer]

runs llama-style models (tinyllamas, llama2.c format). implements:

• rmsnorm
• rotary position embeddings (rope)
• grouped query attention (gqa)
• swiglu ffn
• kv cache
• top-p sampling
• bpe tokenizer

no dependencies except libc and libm.