llama.cpp Windows Builder (Community Scripts)

Automated, reproducible build scripts for Windows 10/11 that get you from a clean machine to a working llama.cpp toolchain with minimal friction.
These scripts can install prerequisites (Git, CMake, Visual Studio + C++ workload, CUDA) and then configure + compile llama.cpp using CMake (Ninja, Release).

License: GPLv3
Upstream project: ggml-org/llama.cpp
Clone URL: https://github.com/ggml-org/llama.cpp.git (You don't need to clone the repo yourself, the installer does this automatically for you)

---

✨ What’s in this repo?

• llama_install_prerequisites_and_build.bat
  End-to-end “happy path” for a fresh Windows install.
  Installs prerequisites if missing (Winget, Git, CMake, Visual Studio Community 2022 + Desktop C++ workload, CUDA Toolkit), then clones and builds llama.cpp.

• llama_build_only.bat
  Fast path when your machine is already set up.
  Clones the latest llama.cpp from ggml-org, configures CMake, and builds it.

• install_cuda.bat (helper)
  Called automatically by the main script. You normally do not run this yourself.
  Detects CUDA via nvcc; if missing, installs a default CUDA version (currently 12.8).

• Tested with RTX 3060, RTX 3090, RTX 5090 on Windows 10/11.

---

🚀 Quick Start

1. Run (as Administrator)

   llama_install_prerequisites_and_build.bat

   • The script checks/installs prerequisites with winget.
   • If Visual Studio or CUDA are missing, it installs them (you may be prompted and/or a reboot may be required).
   • After prerequisites are in place, it configures CMake (Ninja, Release) and builds llama.cpp.
   • Build flags (CUDA architectures) are automatically adapted based on the detected GPU(s).

2. Already set up? If you know Git, CMake, Visual Studio C++ workload, and (optionally) CUDA are installed:

   llama_build_only.bat

Note: At the end of a run, the script prints the path to a detailed log file (look for a line such as [LOG] Details: "...").

---

🧰 What gets installed (when needed)

• Winget (via App Installer on Windows 10/11)
• Git
• CMake
• Visual Studio Community 2022 + Desktop development with C++ workload
• CUDA Toolkit (default 12.8; only installed if nvcc is not found)

The main script attempts to install only what’s missing. On Windows 10, ensure App Installer is up to date (for winget).

---

🏗️ Build Configuration

The scripts configure and build llama.cpp using CMake + Ninja in Release mode. Typical flags include enabling the CUDA backend. A representative configuration (subject to upstream changes and your GPU) looks like:

cmake -S . -B build -G Ninja ^
  -DGGML_CUDA=ON ^
  -DGGML_CUDA_GRAPHS=ON ^
  -DGGML_CUDA_FA_ALL_QUANTS=ON ^
  -DLLAMA_CURL=ON ^
  -DCMAKE_CUDA_ARCHITECTURES=86;120 ^
  -DCMAKE_BUILD_TYPE=Release

Output binaries (depending on the current llama.cpp layout) typically land in:

build\bin\

Expect tools like llama-cli.exe, llama-server.exe, quantize.exe, etc. (exact names may change upstream).

---

🔁 Updating / Rebuilding

• To rebuild with the latest upstream changes, simply re-run llama_build_only.bat.

• If you need a clean configure:

  • Delete the build\ directory and re-run the script.

• If you want to force a fresh clone:

  • Delete the llama.cpp\ folder and re-run.

---

📦 What this repo does not do

• It does not download model files. You’ll need compatible GGUF models placed where your runtime expects them.
• It does not configure runtime servers or service wrappers beyond building the official tools.

---

Known limitations

• Currently Windows-only (Linux planned later).
• Winget IDs may change upstream.

---

✅ Requirements

• Windows 10 (21H2+) or Windows 11
• Administrator privileges (to install prerequisites)
• Internet access
• Disk space: Visual Studio + CUDA can require 10–20 GB+. The llama.cpp build itself is small by comparison.

---

🧯 Troubleshooting

Winget says “No package found matching input criteria”

• Update sources: winget source update
• Ensure App Installer is current (Microsoft Store) and winget --version works in a new terminal.

Ninja/CMake not found

• Re-run the main script (it installs them if missing).

• Or install manually:

  winget install Kitware.CMake
  winget install Ninja-build.Ninja

Visual Studio / C++ workload missing

• Re-run the main script; it uses vswhere + winget to install VS Community 2022 with the required workload.

CUDA / nvcc not found

• The main script calls install_cuda.bat automatically if nvcc is missing.
• If you installed CUDA during the session, open a new terminal (or reboot) so PATH changes take effect.

Build fails after installing toolchains

• Close and reopen your terminal (PATH updates).
• Delete build\ and re-configure.
• Ensure GPU drivers are up to date (NVIDIA).
• Check the printed log file path for exact errors.

---

🔐 Safety & Notes

• These scripts make system-level changes (installing developer tools and CUDA). Review the code and run at your own discretion and risk.
• A reboot may be required for Visual Studio/CUDA. If prompted, accept the reboot and re-run (the main script prepares an automatic resume via RunOnce where possible).

---

🤝 Contributing

PRs and issues are welcome! Please:

• Keep scripts idempotent and readable.
• Prefer winget IDs and quiet/non-interactive installs where reasonable.
• Test on Windows 10 and Windows 11.

---

🙏 Acknowledgements

• llama.cpp by the GGML organization and open-source contributors (originally started by Georgi Gerganov).
• Thanks to everyone building tools that make local AI accessible.

---

📜 License

This project is licensed under GPLv3. The upstream llama.cpp project is licensed separately; refer to its repository for terms.

---

📄 Legal / Disclaimer

• This project is an independent community initiative.

• It is not affiliated with, endorsed by, or officially associated with the original llama.cpp project or its maintainers.

• All references to llama.cpp are purely descriptive and for compatibility purposes only.

• This repository aims to simplify usage for local developers, while respecting the original work and licensing.

Created and maintained with 💙 by py-sandy