Custom Installation

Build the backend from source if you want the latest main branch, a debug build, or to work on docc itself. The backend (pytorch/) is pure Python and builds on the docc Python package (python/), which contains the compiler. Both are installed from the same checkout so that their versions match.

Requirements

The requirements of the Python Custom Installation apply: Clang 21, CMake, Ninja, and the system libraries.

tip

Skip the setup with the Docker images: daisytuner/docc-build-env-llvm21-ubuntu-24.04 contains all build dependencies, and the run-env images add CUDA and ROCm. PyTorch itself is installed from the requirements files below.

Build

git clone https://github.com/daisytuner/docc.git
cd docc
git submodule update --init --recursive

python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip

1. Install PyTorch and the build dependencies. Choose the requirements file for your hardware:

   File	Installs
   pytorch/requirements.txt	PyTorch for CPUs
   pytorch/requirements-cuda.txt	PyTorch for CUDA 12.6 and CuPy
   pytorch/requirements-rocm.txt	PyTorch for ROCm 7.2 and CuPy
   pytorch/requirements-base.txt	Only the build and test dependencies; bring your own PyTorch

   pip install -r pytorch/requirements.txt

   The CUDA 12.6 build of PyTorch does not support Blackwell GPUs (RTX 50xx, B200). For those, install requirements-base.txt, cupy-cuda12x, and PyTorch from the cu129 index instead.

2. Build and install the compiler:

   pip install --no-build-isolation -e python/

3. Install the backend:

   pip install --no-build-isolation -e pytorch/

Check your build with the ResNet-18 example from the Quickstart:

python pytorch/examples/resnet18.py

Changes to the Python sources of both packages take effect immediately. After changing C++ sources, run step 2 again.

Build options

The compiler build accepts the same CMake options as the Python package, for example the C++ compiler used for generated code. See Python Custom Installation.

Debug builds

Build the compiler with debug information and keep the CMake build directory, so that rebuilds after C++ changes are incremental:

pip install --no-build-isolation -e python/ \
    -C cmake.build-type=Debug \
    -C build-dir=build-python

Debug Options explains how to run the tests and inspect what docc does with your model.