Installation Guide

Quick Install

Requirements: Python 3.12 or higher

BLISS is pip installable:

pip install bliss-toolkit

Developer Installation

1. Install uv:

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

2. Install the fftw library (used by GalSim):

   sudo apt-get install libfftw3-dev  # Ubuntu/Debian
   brew install fftw                   # macOS
   

3. Install git-lfs if you haven’t already:

   git-lfs install
   

4. Clone the repo and install dependencies:

   git clone git@github.com:prob-ml/bliss.git
   cd bliss
   uv sync --all-extras
   

5. Activate the virtual environment:

   source .venv/bin/activate
   

6. Verify installation by running tests:

   pytest
   pytest --gpu
   

7. (Optional) Install pre-commit hooks for development:

   pre-commit install
   

GPU Requirements

BLISS requires a CUDA-capable GPU for training. Check your CUDA version with:

nvidia-smi

BLISS uses PyTorch with CUDA 12.8 support, which works on both older GPUs (RTX 2080 Ti) and newer architectures.

Core Dependencies

BLISS requires the following core packages:

• PyTorch (>=2.7.0): Deep learning framework

• PyTorch Lightning (>=2.3.3): Training infrastructure

• Hydra (>=1.0.4): Configuration management

• GalSim (>=2.4.10): Galaxy image simulation

• Astropy (>=6.1.1): Astronomical utilities

• NumPy, SciPy, Matplotlib: Scientific computing

All dependencies are automatically installed via uv sync --all-extras.

Troubleshooting

GalSim installation fails:

Ensure FFTW is installed before running uv sync:

sudo apt-get install libfftw3-dev  # Ubuntu/Debian
brew install fftw                   # macOS

CUDA not found:

Verify your CUDA installation:

nvcc --version
nvidia-smi

Memory errors during training:

Reduce batch size in configuration:

bliss mode=train cached_simulator.batch_size=8