AlphaFold 3.0.4

AlphaFold3 model parameters must be requested and downloaded by the user themselves.

Webpage

https://github.com/google-deepmind/alphafold3

Version

3.0.4

Environment

• NVIDIA Driver 610.43.02

Installation (briefly)

AlphaFold

[user@ccfep3 ~]$ curl -LsSf https://astral.sh/uv/install.sh | sh
[user@ccfep3 ~]$ export UV_PYTHON_INSTALL_DIR=/apl/alphafold/3.0.4/python-uv
[user@ccfep3 ~]$ uv python install 3.12
[user@ccfep3 ~]$ export UV_PYTHON=/apl/alphafold/3.0.4/python-uv/cpython-3.12.13-linux-x86_64-gnu/bin/python
[user@ccfep3 ~]$ export UV_PROJECT_ENVIRONMENT=/apl/alphafold/3.0.4/venv
[user@ccfep3 ~]$ uv venv $UV_PROJECT_ENVIRONMENT

[user@ccfep3 ~]$ cd /apl/alphafold/3.0.4
[user@ccfep3 3.0.4]$ git clone https://github.com/google-deepmind/alphafold3.git -b v3.0.4
[user@ccfep3 3.0.4]$ cd alphafold3/
[user@ccfep3 alphafold3]$ module load gcc-toolset/14
[user@ccfep3 alphafold3]$ uv sync --frozen --all-groups --no-editable

[user@ccfep3 alphafold3]$ . /apl/alphafold/3.0.4/venv/bin/activate
(venv) [user@ccfep3 alphafold3]$ build_data

hmmer-3.4 built for the previous version (3.0.3) is also used here (via symbolic link).

Notes

• The procedure is equivalent to that for 3.0.3.
• This version can run without GPU.
  • --flash_attention_implementation=xla --jax_backend=cpu maybe necessary.
  • During execution, many “Failed to run implementation” errors appear in the tokamax related part. But it seems the execution is completed successfully.