This repository handles everything related to preparing a local model for the dev-agent.
- Modify codebase to check if download model is already quantized (quantization step is hardcoded in release workflow)
- Check if there is a better format for the release asset in Github, currently released in .tar.gz
- Secure download from Hugging Face Hub
- Optional quantization (GGUF / GPTQ)
- Packaging a deterministic
.tar.gzmodel bundle - Automatic publishing via GitHub Releases
| Script | Purpose |
|---|---|
download.py |
Download HF model securely |
quantize.py |
Convert or quantize to GGUF / GPTQ |
package_model.py |
Build versioned tarball for release |
Run the workflow manually from GitHub Actions:
Build & Release Model → Run workflow
Inputs:
repo_id: e.g.meta-llama/Meta-Llama-3.1-8Bmodel_version: e.g.v0.1.0