Install docc for PyTorch

docc plugs into PyTorch as a torch.compile backend. You keep your model as it is, add one import, and select backend="docc". In this Quickstart, you install the backend and then run an image classifier (ResNet-18) on all CPU cores and on a GPU.

Requirements

Component	Requirement
Python	3.11 – 3.14
Operating system	Linux (x86-64, arm64), macOS ≥ 14 (Apple silicon, CPU only)
PyTorch	2.13 (the tested version)
C++ compiler	Linux: clang++-21 from the official LLVM repository; macOS: Xcode Command Line Tools
NVIDIA GPUs	CUDA 12 and CuPy (cupy-cuda12x)
AMD GPUs	ROCm 7 and CuPy (cupy-rocm-7-0)

The backend generates C++ code for your model and compiles it with Clang, which is why a C++ compiler is required at runtime.

Install

1. On Linux, install Clang 21:

   wget https://apt.llvm.org/llvm.sh
   chmod +x llvm.sh
   sudo ./llvm.sh 21

2. Create a virtual environment and install PyTorch for your hardware:

   python3 -m venv .venv
   source .venv/bin/activate
   
   # CPU only
   pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu
   # NVIDIA GPUs (use cu129 for Blackwell GPUs such as RTX 50xx and B200)
   pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126 && pip install cupy-cuda12x
   # AMD GPUs
   pip install torch torchvision --index-url https://download.pytorch.org/whl/rocm7.2 && pip install cupy-rocm-7-0

3. Install the docc compiler and the PyTorch backend:

   pip install docc-compiler docc-pytorch

4. Check the installation:

   python -c "import docc.pytorch; print('docc backend is ready')"

Prepare GPU targets (optional)

docc detects the first GPU in your system and compiles for its architecture. To select a different one, set:

export DOCC_CUDA_ARCH=sm_90     # NVIDIA H100
export DOCC_ROCM_ARCH=gfx942    # AMD MI300

Next step

Continue with Run Your First Model to compile and accelerate ResNet-18.