Skip to content

Installation

This page explains how to install INFTY for normal use, example execution, and local development.

Requirements

INFTY is a Python package built around PyTorch. The package metadata declares the following core dependencies:

  • torch
  • torchvision
  • numpy
  • matplotlib
  • scipy
  • seaborn
  • tqdm

INFTY currently requires Python >=3.8. The tested CI support matrix is Python 3.8, 3.9, 3.10, and 3.11.

Install from PyPI

python -m pip install --upgrade pip
python -m pip install infty

Verify the installation:

python -c "from infty.optim import C_Flat, ZeroFlow, UniGrad_FS; print('INFTY import succeeded')"

Install from source

git clone https://github.com/THUDM/INFTY.git
cd INFTY
python -m pip install --upgrade pip
python -m pip install .

For editable development:

python -m pip install -e .

Install PILOT and demo dependencies

The examples extra primarily supports the PILOT benchmark integration and related optional workflows. It includes packages such as easydict, Pillow, PyYAML, scikit-learn, timm, and umap-learn.

The minimal scripts under examples/infty_minimal/ use only the core INFTY installation.

From a source checkout, run this from the repository root:

python -m pip install ".[examples]"

If you installed INFTY from PyPI instead of a source checkout:

python -m pip install "infty[examples]"

For editable development with examples from a source checkout:

python -m pip install -e ".[examples]"

conda create -n infty python=3.8
conda activate infty
python -m pip install --upgrade pip
python -m pip install infty

If you need a specific CUDA build of PyTorch, install PyTorch first using the command matching your CUDA/runtime setup, then install INFTY.

Common checks

Check that infty is imported from the expected location:

python -c "import infty; print(infty.__file__)"

Check optimizer exports:

python -c "from infty.optim import C_Flat, ZeroFlow, UniGrad_FS; print(C_Flat, ZeroFlow, UniGrad_FS)"

Check plotting exports:

python -c "from infty.plot import visualize_landscape, visualize_esd, visualize_conflicts, visualize_trajectory; print('plot utilities imported')"

Troubleshooting

If installation fails, see Troubleshooting. The most common causes are incompatible PyTorch/CUDA versions, installing into the wrong Python environment, or missing PILOT/demo dependencies from the examples extra.