llama-tokens

Babashka scripts for experimenting with ::M completions. Intended to be used with llama-server from llama.cpp.

Dependencies

• babashka (bb command)
• golang (go command)
• llama-server from llama.cpp

Setup

Go to Hugging Face and find a gguf model you have enough "compute" to run, and run it something like this:

./build/bin/llama-server -m ~/models/Qwen3-14B-UD-Q5_K_XL.gguf -c 28000 --port 8080 --host 127.0.0.1

Copy the example config and point it at your server:

cp resources/config.edn.example resources/config.edn

Scripts

generate_chat

Uses the qwen3 chat template to produce an example

go run go/generate_chat.go

tokenize.bb

Tokenizes a string and prints the token IDs.

./tokenize.bb "Hello, world!"
# tokens: [9906 11 1917 0]

complete.bb

Runs a completion against the server.

./complete.bb "Once upon a time"

Options:

Flag	Description
-n N / --max-tokens N	Maximum tokens to generate (default: 30)
-c / --char-tokens	Tokenize the prompt character-by-character before sending
-h / --help	Show help