User Guide

Installation

Pip install (inference only)

For inference only, install the toolkit from PyPI:

pip install git+https://github.com/b-sigpro/sbss

Pixi install (training)

If you want to train models or develop SBSS, follow these steps:

  1. (If needed) Set up Pixi by following the official installation guide.

  2. Clone the repo:

    git clone github.com:b-sigpro/sbss

  3. Adjust the MPI-related entries in pyproject.toml to match your environment:

    openmpi = { version = "4.1.*", build = "external_*" }
hdf5 = { version = "*", build = "mpi_openmpi*" }
h5py = { version = "*", build = "mpi_openmpi*" }

    • If you do not have MPI installed locally (e.g., on a laptop or a plain AWS instance), remove the build = "external_*" part so Pixi can install OpenMPI for you.

      openmpi = { version = "4.1.*" }
hdf5 = { version = "*", build = "mpi_openmpi*" }
h5py = { version = "*", build = "mpi_openmpi*" }

    • If you are on a cluster, check the available MPI version via module available or mpirun --version, and set the version of mpi (or another MPI provider) accordingly. The default MPI is openmpi, but if your environment uses another one (e.g., MPICH), replace the package and build selectors as needed.

      mpich = { version = "4.3.*" }
hdf5 = { version = "*", build = "mpi_mpich*" }
h5py = { version = "*", build = "mpi_mpich*" }

    Note

    OpenMPI is ABI-compatible within the same major version, so pick the closest version even if the exact one is not available. For example, on ABCI 3.0, HPC-X 2.20 provides OpenMPI 4.1.7, but keeping openmpi = "4.1.*" lets Pixi install 4.1.6, and it works without issues.

  4. Install dependencies with Pixi:

    pixi install