Skip to content

building.md

Latest commit

 

History

History
103 lines (69 loc) · 3.5 KB

building.md

File metadata and controls

103 lines (69 loc) · 3.5 KB

Building from Source

This guide covers building AlphaFast from source for development or non-containerized deployment.

Prerequisites

Note: Building from source is only supported on Linux. macOS and Windows are not supported.

  • Python >= 3.12
  • CUDA 12.6 (toolkit and compatible driver >= 560.28.03)
  • C++20 compiler: gcc/g++ >= 10 (for pybind11 C++ extensions)
  • CMake >= 3.28
  • git
  • UV package manager

Installing UV

curl -LsSf https://astral.sh/uv/install.sh | sh

Build Steps

1. Clone the Repository

git clone https://github.com/RomeroLab/alphafast.git
cd alphafast

2. Install Dependencies

UV reads pyproject.toml and uv.lock to install all Python dependencies and compile the C++ extensions:

uv sync --frozen --all-groups

This step:

  • Creates a virtual environment
  • Installs all Python dependencies (JAX, NumPy, RDKit, etc.)
  • Compiles the C++ pybind11 extension via scikit-build-core and CMake

3. Build Chemical Components Database

The chemical components database is a set of pickled data files required before running inference:

uv run build_data

This command is provided as a script entry point defined in pyproject.toml and uses the cifpp and dssp C++ libraries built during the previous step.

4. Install MMseqs2-GPU and Foldseek

Download pre-built binaries:

# MMseqs2 with GPU support
wget https://mmseqs.com/latest/mmseqs-linux-gpu.tar.gz
tar xzf mmseqs-linux-gpu.tar.gz
cp mmseqs/bin/mmseqs ~/.local/bin/

# Foldseek with GPU support (optional, for foldseek template modes)
wget https://mmseqs.com/foldseek/foldseek-linux-gpu.tar.gz
tar xzf foldseek-linux-gpu.tar.gz
cp foldseek/bin/foldseek ~/.local/bin/

Ensure ~/.local/bin is in your PATH, or pass the binary paths explicitly with --mmseqs_binary_path and --foldseek_binary_path.

Build System Details

AlphaFast uses scikit-build-core as its build backend, which invokes CMake to compile C++ pybind11 extensions. This should be the same as AlphaFold 3's build system. The relevant files:

  • pyproject.toml: Declares build dependencies (scikit_build_core, pybind11, cmake, ninja, numpy) and project metadata.
  • CMakeLists.txt: Defines the C++ build. Uses FetchContent to download and build abseil-cpp, pybind11, pybind11_abseil, libcifpp, and dssp. Compiles all .cc files under src/alphafold3/ into a single cpp pybind11 module.
  • src/alphafold3/cpp.cc: The pybind11 entry point that exposes C++ functions to Python.

The C++20 standard is required (CMAKE_CXX_STANDARD 20).

Docker Build

To build the Docker image locally:

docker build -f docker/Dockerfile -t romerolabduke/alphafast .

The Dockerfile:

  1. Starts from nvidia/cuda:12.6.3-base-ubuntu24.04
  2. Installs system dependencies (Python 3.12, gcc, g++, git)
  3. Installs UV and creates a virtual environment
  4. Downloads MMseqs2-GPU and Foldseek binaries
  5. Installs Python dependencies via uv sync
  6. Builds the chemical components database via uv run build_data
  7. Sets XLA_FLAGS and XLA_CLIENT_MEM_FRACTION environment variables

Model Weights

Model weights (af3.bin.zst) must be obtained separately from Google DeepMind. Request access via this form. Place the file in your --weights_dir / --model_dir directory. The weights are subject to the AlphaFold 3 Model Parameters Terms of Use.