Skip to content

README.md

Latest commit

 

History

History
70 lines (59 loc) · 3.41 KB

README.md

File metadata and controls

70 lines (59 loc) · 3.41 KB
Raw

HiGHS GPU (PDLP) through OR-Tools

Notes and checks for running HiGHS' GPU PDLP through OR-Tools. These are run by hand on GPU machines; CI only covers the CPU builds (.github/workflows/ortools-*.yml). The commands are in ../or-tools-readme.md.

Files

  • check_highs_gpu.py: solves an LP with HiGHS through OR-Tools' Python API (pywraplp) with HiGHS options (default solver=pdlp). Run it through ../ortools/check_log.py --require-gpu to make sure PDLP really ran on the GPU.
  • ../ortools/highs_check/: the same check in C++, used for Bazel. It is copied into the OR-Tools tree as //highs_check by ortools/prepare.sh.
  • ../ortools/check_log.py: runs a command and checks its output: OPTIMAL status, objective, HiGHS version, and with --require-gpu the Cuda device N: <name> line that cuPDLP-C prints when it starts on a GPU (with HiGHS output on). Without that line, PDLP ran on the CPU.

Build options

  • CMake, HIGHS_GPU_LIB=ON (HiGHS option, passed straight through OR-Tools' CMake): builds the CUDA code into the HiGHS library itself (it implies CUPDLP_GPU=ON), so the GPU solver works from OR-Tools' Python wheel. Use this for the Python checks.
  • CMake, CUPDLP_GPU=ON alone: the CUDA code goes in a separate cudalin library, which OR-Tools' Python build does not pick up.
  • Bazel: --@highs//:cupdlp_gpu --@rules_cuda//cuda:archs=sm_XY. XY is the GPU's compute capability without the dot, from nvidia-smi --query-gpu=compute_cap --format=csv,noheader (8.9 -> sm_89). --//:cupdlp_gpu is only for building HiGHS on its own.

Passing HiGHS options through OR-Tools (v9.15)

  • ortools/linear_solver/solve --solver=highs --params=... cannot be used: it CHECK-fails, because HighsInterface::SetSolverSpecificParametersAsString() returns false. Run solve without --params, or use highs_check.
  • The same call from pywraplp or C++ MPSolver returns False, but the string is still stored and passed to HiGHS on Solve(), so it works. Ignore the return value.
  • Format: name=value or name value, separated by commas or newlines, e.g. solver=pdlp,presolve=off. Avoid empty lines (HiGHS' parser reads line[0] of every line), so no """\nsolver pdlp\n""" strings.
  • Python APIs:
    • pywraplp (legacy): options work as above; it cannot read MPS files directly, so check_highs_gpu.py --mps goes through model_builder and LoadModelFromProto.
    • model_builder: reads MPS files but cannot pass HiGHS options.
    • math_opt: options go through highs_pb2.HighsOptionsProto in SolveParameters(highs=...). The generic LPAlgorithm.FIRST_ORDER is not mapped to HiGHS PDLP.

Windows

  • Python 3.8+ ignores PATH when it loads extension module DLLs. check_highs_gpu.py adds %CUDA_PATH%\bin (and bin\x64 for CUDA 13) itself. Use --dll-dir for anything else, e.g. the build's Release\bin if the wheel is missing a DLL.
  • Building the Python wheel needs mypy and mypy-protobuf (pip install mypy mypy-protobuf), and stubgen.exe on PATH. For a pip install --user setup: $env:PATH += ";$env:APPDATA\Python\Python312\Scripts".
  • On Windows, HiGHS builds as a shared library whenever GPU is on.

Linux

  • Install swig, mypy and mypy-protobuf before building the Python wheel.
  • If the CUDA libraries are not in the loader path: export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH.