Source Code Compilation and Build

Obtain the PaddlePaddle source code, configure CMake, and compile and check wheel.

Obtain the PaddlePaddle source code.

git clone --branch v3.3.1 --depth 1 https://github.com/PaddlePaddle/Paddle.git 
cd Paddle 
git submodule sync --recursive 
git submodule update --init --recursive

Create a build directory and configure CMake.

mkdir -p build-cp311-cuda13
cd build-cp311-cuda13 
ulimit -n 4096  
cmake .. \
  -GNinja \
  -DCMAKE_BUILD_TYPE=Release \
  -DPY_VERSION=3.11 \
  -DPYTHON_EXECUTABLE="$(command -v python3)" \
  -DWITH_GPU=ON \
  -DWITH_ARM=ON \
  -DWITH_TESTING=OFF \
  -DWITH_DISTRIBUTE=OFF \
  -DWITH_NCCL=OFF \
  -DWITH_OPENVINO=OFF \
  -DWITH_CUDNN=ON \
  -DWITH_CUDNN_FRONTEND=OFF \
  -DWITH_TP_CACHE=ON \
  -DTHIRD_PARTY_CACHE_PATH="$(pwd)/third_party_cache" \
  -DCUDA_ARCH_NAME=Ampere \
  -DCMAKE_CUDA_FLAGS="-DEIGEN_DONT_VECTORIZE -UPADDLE_WITH_SLEEF"

**Table 1** Key parameters
Parameter	Description
-DPY_VERSION=3.11	Generate a Python 3.11 ABI wheel.
-DWITH_GPU=ON	Build PaddlePaddle for the GPU version.
-DWITH_ARM=ON	Enable the Arm architecture for building.
-DWITH_TESTING=OFF	Disable the construction of test targets to reduce the construction time and dependency scale.
-DWITH_DISTRIBUTE=OFF	Disable the distributed training capability.
-DWITH_NCCL=OFF	Disable NCCL. Multi-device communication is not included in the minimum verification path.
-DWITH_OPENVINO=OFF	Disable the construction of the OpenVINO backend.
-DCUDA_ARCH_NAME=Ampere	Generate CUDA code for the A100/Ampere architecture.
-DCMAKE_CUDA_FLAGS=...	Used by the CUDA compilation unit to prevent the NVCC from parsing the Arm NEON header file in AArch64.

Perform the compilation.

export MAX_JOBS=8 
export CMAKE_BUILD_PARALLEL_LEVEL=8 
ninja -j8

Check the wheel file after the build is complete.
1
ls -lh python/dist/*.whl
The following is an example of the build product.
1
paddlepaddle_gpu-3.3.0.dev20260319-cp311-cp311-linux_aarch64.whl

Parent topic: Development Guide