Command-line interface¶

The cupdlpx executable solves LPs in .mps or .mps.gz format and writes solution vectors and a solve summary to the output directory. Build it from source with the C libraries.

Build requirements¶

See hardware requirements for supported GPUs and required CUDA or ROCm versions.

Installation¶

Clone the repository:

git clone https://github.com/MIT-Lu-Lab/cuPDLPx.git
cd cuPDLPx

CUDA¶

Configure CMake and build:

cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --clean-first --parallel

The build selects the sparse matrix–vector backend from the cuSPARSE version:

To target a specific CUDA architecture, pass -DCMAKE_CUDA_ARCHITECTURES=<architecture> when configuring.

ROCm¶

Enable HIP and provide the architecture of the target AMD GPU:

cmake -B build \
  -DUSE_HIP=ON \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_HIP_ARCHITECTURES=gfx90a \
  -DCMAKE_PREFIX_PATH=/opt/rocm
cmake --build build --clean-first --parallel

Common architecture values include gfx90a for MI200, gfx1100 for RDNA3, and gfx1201 for RDNA4. Replace the value with the architecture supported by the installed ROCm toolchain.

On HIP builds, cuBLAS, cuSPARSE, and CUB calls are mapped to hipBLAS, hipSPARSE, and hipCUB.

Build outputs¶

Both configurations create:

• build/cupdlpx, the command-line solver;
• a static core library; and
• a shared cupdlpx library.

Usage¶

cupdlpx [OPTIONS] <mps_file> <output_directory>

For a source build, the executable is normally ./build/cupdlpx:

mkdir -p results
./build/cupdlpx problem.mps.gz results

The output directory must already exist.

Solver options¶

Pass solver options before the input and output paths. See Parameters for defaults and allowed values, or run ./build/cupdlpx --help.

./build/cupdlpx \
  --time_limit 600 \
  --eps_opt 1e-6 \
  --eps_feas 1e-6 \
  --opt_norm linf \
  problem.mps.gz results

Output files¶

For problem.mps.gz, cuPDLPx creates:

results/
├── problem_summary.txt
├── problem_primal_solution.txt
└── problem_dual_solution.txt

The primal and dual files contain one value per line. The summary records the termination reason, model dimensions, objective values, residuals, iteration count, and phase timings. Dual slacks are available through the in-memory Python, Julia, and C interfaces rather than as a separate CLI output file.

Advanced build options¶

The native build exposes the following CMake options:

For example, build only the shared library:

cmake -B build \
  -DCUPDLPX_BUILD_STATIC_LIB=OFF \
  -DCUPDLPX_BUILD_CLI=OFF
cmake --build build --parallel

See GPU backends for architecture defaults and platform requirements.