Skip to content

INSTALLATION_SUPPORT.md

Latest commit

 

History

History
74 lines (48 loc) · 2.22 KB

INSTALLATION_SUPPORT.md

File metadata and controls

74 lines (48 loc) · 2.22 KB

Installation Support

compactlogic is a Python 3.6+ package for PyTorch-based logic-gate networks. The original setup notes target PyTorch 1.9.0+ for CUDA builds. The package builds custom CUDA extensions, so the main installation requirement is a working CUDA-enabled PyTorch environment with a compatible local CUDA Toolkit.

Requirements

  • Python 3.6+
  • PyTorch (CUDA-enabled)
  • NVIDIA CUDA toolkit for building the custom CUDA extensions

⚠️ CUDA is required. Ensure your installed torch version matches your CUDA runtime/toolkit.

Tested Environments

  • PyTorch: 1.13, 2.5, 2.11
  • CUDA: 12.2, 12.8, 13.0
  • NVCC: 11.8, 13.2

Install from source

This repository is intended to be installed from source:

git clone https://github.com/eth-sri/CompactLogic.git
cd CompactLogic
pip install -e .

For the experiment scripts, install the experiment dependencies in the same Python environment:

pip install -r experiments/requirements.txt

For compiler/reporting utilities, also install:

pip install -r simulation/requirements.txt

CUDA and PyTorch compatibility

The CUDA version used by your installed PyTorch package must be compatible with the CUDA Toolkit available on the machine that builds the extensions.

You can check the locally available CUDA driver/runtime information with:

nvidia-smi

You can install PyTorch and torchvision builds for a specific CUDA version, for example:

pip install torch==2.11.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130  # CUDA 13.0

If you see an error like this while building the CUDA extension:

Failed to build difflogic

...

RuntimeError:
    The detected CUDA version (11.2) mismatches the version that was used to compile
    PyTorch (11.7). Please make sure to use the same CUDA versions.

then install a PyTorch build whose CUDA version matches your local CUDA Toolkit more closely, or use a machine/environment with a matching CUDA Toolkit. In some cases, trying a nearby older PyTorch/CUDA build can resolve mismatches even when the advertised versions appear compatible.