Source Code Compilation and Build

Build Paddle Inference from the PaddlePaddle source code tree. After the build is complete, configure the GPU-oriented inference path and complete the compilation.

Obtain the PaddlePaddle 3.3.1 source code.

Build Paddle Inference from the PaddlePaddle source code tree.

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

If OpenVINO is not enabled, you can skip the third_party/openvino submodule to avoid downloading unused dependencies.

git config -f .gitmodules --get-regexp path | awk '{print $2}' \
  | grep -v '^third_party/openvino$' \
  | xargs -r git submodule update --init --recursive --

Configure CMake.

The following configuration is for the GPU inference path and generates CUDA code based on the single Ampere architecture.

mkdir -p build-inference-cuda13 
cd build-inference-cuda13

cmake .. \
  -GNinja \
  -DCMAKE_BUILD_TYPE=Release \
  -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_PYTHON=OFF \
  -DON_INFER=ON \
  -DWITH_INFERENCE_API_TEST=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 parameter description
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

You are advised to use 8 concurrent jobs to balance the build speed and system load.

After the build is complete, the product directories are build-inference-cuda13/paddle_inference_install_dir/ and build-inference-cuda13/paddle_inference_c_install_dir/.

Example of the C++ API product:

paddle_inference_install_dir/paddle/include/paddle_inference_api.h 
paddle_inference_install_dir/paddle/lib/libpaddle_inference.so 
paddle_inference_install_dir/paddle/lib/libpaddle_inference.a 
paddle_inference_install_dir/version.txt

Example of the C API product:

paddle_inference_c_install_dir/paddle/include/pd_inference_api.h 
paddle_inference_c_install_dir/paddle/lib/libpaddle_inference_c.so 
paddle_inference_c_install_dir/paddle/lib/libpaddle_inference_c.a

Parent topic: Developer Guide