README.md

metadata

title: Committed API
emoji: 🤖
colorFrom: indigo
colorTo: gray
sdk: docker
app_port: 7860
pinned: false

Committed — Inference API

Headless FastAPI inference service for Committed: a Qwen3-1.7B QLoRA fine-tune that generates Conventional Commits messages from code diffs, served on CPU via llama.cpp with GBNF grammar-constrained decoding.

This Space exposes the HTTP contract; it has no UI. For the interactive demo, see the Gradio Space.

Endpoints

• POST /generate — body {"diff": "<unified diff>"} → {"message": "<commit message>"}
• GET /health — {"status": "ok"} once the model is loaded
• GET /docs — interactive Swagger UI (FastAPI default)

Model

Serves the fine-tuned GGUF marzoukbaig14/committed-gguf / committed-finetuned-Q4_K_M.gguf (the serving artifact of record, ADR 0048). Overridable via the COMMITTED_MODEL_REPO / COMMITTED_MODEL_FILE Space Variables.

Notes

The container builds by installing the committed package from GitHub (the single source of truth). The model is pulled from the Hub at first startup and cached, so the first boot is slower than subsequent ones. The Space may sleep after a period of inactivity; the first request after a sleep incurs a cold start.