llama2.ts

Inference for Llama2-like Transformer models in one TypeScript file

Heavily based on the Andrej Karpathy's llama2.c.

Mostly of educational value (understand something by implementing it yourself! porting in this case but still :P)

Features

• Binary compatible (i.e. should produce exactly the same outputs as the C version given the parameters and random seed)
• Achieves around 70/25/10 tokens per second for the 15/45/110M models, respectively.
• Can run the full 7B model with 0.16 tokens per second on my laptop o_O

Includes the TinyStories 15M model.

Usage

node (via the bundled t348):

node --experimental-loader=./t348.mjs llama2.ts stories15M.bin -s 1 -t 0 -i "Once upon a time"

bun:

bun llama2.ts stories15M.bin -i "Once upon a time"

Larger TinyStories models:

wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories42M.bin
wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories110M.bin

Arguments:

• -i <string> - initial prompt
• -t <float> - temperature (0..1, 0 = deterministic argmax)
• -s <int> - random seed
• -n <int> - number of tokens to generate (0..256, default 256)
• -p <float> - p value for nucleus sampling, default 0.9

UPD: see also llama2.js by @epicure for a browser version. Glad I missed it before starting this project, otherwise I'd probably never start it :D