llama-cpp-wheels

🦙 llama-cpp-python: Multi-Architecture Wheel Builder

This repository automatically builds and publishes Python wheels for abetlen/llama-cpp-python across all major platforms and architectures using GitHub Actions and cibuildwheel.

✅ Wheels are uploaded to the GitHub Releases section of this repo — ready to download!

---

🚀 Supported Platforms

This build system supports:

OS	Architecture
Linux	x86_64, aarch64
macOS	arm64, x86_64
Windows	x86_64

---

🧪 Supported Python Versions

Wheels are built for:

• Python 3.8 → 3.13

This means all of the following tags will be available:

cp38, cp39, cp310, cp311, cp312, cp313

---

🛠️ How It Works

1. watch-upstream.yml – Scheduled Trigger (Every 12h)

• Fetches the latest release tag from the llama-cpp-python GitHub API
• Extracts the base tag (e.g., v0.3.14 from v0.3.14-cu124)

• If the release hasn’t been built yet:

  • Triggers build-wheels.yml with that version

2. build-wheels.yml – Cross-Platform Builder

• Uses cibuildwheel and docker buildx to build wheels

• Supports:

  • 🐧 Linux (QEMU for aarch64)
  • 🍎 macOS (Intel + Apple Silicon)
  • 🪟 Windows (MSVC toolchain)
• Builds wheels for Python 3.8 through 3.13
• Uploads wheels to a GitHub Release matching the tag

---

📦 Example Wheel Names

llama_cpp_python-0.3.14-cp38-cp38-manylinux_x86_64.whl
llama_cpp_python-0.3.14-cp310-cp310-manylinux_aarch64.whl
llama_cpp_python-0.3.14-cp313-cp313-macosx_11_0_arm64.whl
llama_cpp_python-0.3.14-cp311-cp311-win_amd64.whl

---

🔧 Manual Trigger

You can manually build a specific version:

1. Go to the Actions tab
2. Select build-wheels.yml
3. Click “Run workflow”
4. Enter the desired tag (e.g., v0.3.14-cu124)

---

⏰ Schedule

The watch-upstream.yml workflow runs every 12 hours via cron:

schedule:
  - cron: '0 */12 * * *'  # Every 12 hours

---

📤 Release Output

All built .whl files are published as release assets under the Releases tab for the corresponding version tag.

---

🔐 Security Notes

• GitHub-hosted runners with ephemeral environments
• Isolated builds using Docker and cibuildwheel
• Sources pulled directly from upstream: abetlen/llama-cpp-python

---

🙋 FAQ

Q: Does this build support CUDA or Metal?

A: Currently, only CPU-only builds are supported. CUDA and Metal require custom runners and are not compatible with GitHub-hosted environments.

Q: Why do wheels include cu124 in the tag?

A: That part comes from the upstream release tag. We extract the base tag (e.g., v0.3.14) for organizing releases.

Q: How do I change supported Python versions?

A: Edit the CIBW_BUILD value in build-wheels.yml. This repo is already set to build for Python 3.8–3.13.

---

📄 License

MIT © eswarthammana – Built on top of open source contributions from abetlen and pypa/cibuildwheel.

---

This site is open source. Improve this page.