Install OpenCUDA-Lab 0.1 (exportable preview)

This is not a frozen production ABI (SOVERSION is still 0).

The PyPI package is a source distribution. pip install opencudalab compiles libopencuda inside pip on that machine. The third person does not run CMake by hand. A Mac wheel is never the CUDA artifact.

pip (what a third person runs)

Needs: Python 3.9+, CMake, a C/C++ compiler (Visual Studio Build Tools on Windows), and nvcc on PATH for GPU backends.

pip install opencudalab
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124
python -c "from opencudalab import Runtime, install_pytorch_allocator; print('ok')"

ocl_pytorch_alloc is always compiled into the shared library. After pip install, import opencudalab loads the DLL from site-packages.

If CUDA is on PATH (nvcc), the GPU backends are built automatically. Force the mock backend: set OCL_FORCE_MOCK=1 before pip install.

Until the project is published to PyPI, the same command works from a clone:

pip install .

USB tarball (no GitHub / no PyPI yet)

./scripts/pack_v1.sh

Copy dist/opencudalab-0.1.0-usb.tar.gz. Do not copy a Mac .dylib.

Windows + NVIDIA + CUDA (friend laptop)

Needs: NVIDIA GPU, CUDA Toolkit, Visual Studio 2022 Build Tools (C++ workload), Python 3.10+, CMake on PATH (winget install Kitware.CMake).

In x64 Native Tools Command Prompt or a terminal where nvcc works:

tar -xf opencudalab-0.1.0.tar.gz
cd opencudalab-0.1.0

python -m venv .venv
.venv\Scripts\activate
python -m pip install -U pip

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124

pip install opencudalab

Then the with-vs-without comparison:

python -m opencudalab --compare --batches 8

Or without pip, CMake + the example:

cmake -S . -B build -DOCL_ENABLE_CUDA=ON
cmake --build build --config Release
python examples/python/resnet_cifar10.py --compare --lib build\Release\opencuda.dll

If native Windows build fights you, WSL2 Ubuntu + CUDA on WSL is the same Linux path as the paper's GPU box:

cmake -S . -B build -DOCL_ENABLE_CUDA=ON && cmake --build build -j
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124
python examples/python/resnet_cifar10.py --compare --lib build/libopencuda.so

Look at torch_peak_alloc_MiB and ocl_n_oom_actual in the table.

pip (Mac / Linux)

From this repo, or from PyPI after publish:

pip install opencudalab
python -c "import opencudalab; print(opencudalab.library_path())"
python -m opencudalab --cpu --compare --dataset synthetic

On this Mac that installs the mock backend (no NVIDIA GPU).

To publish the source distribution (maintainers):

./scripts/publish_pypi.sh           # build + print upload commands
./scripts/publish_pypi.sh --test    # TestPyPI (needs token)
./scripts/publish_pypi.sh --upload  # PyPI (needs token)

Homebrew (macOS, mock backend only)

brew install cmake
./scripts/pack_v1.sh
brew install --build-from-source ./packaging/homebrew/opencudalab.rb

Python API (0.1.3+)

opencudalab.Runtime / Pipeline wrap the C ABI. ctypes layouts are not part of notebooks. Mock-backend checks (no GPU):

cmake -S . -B build -DOCL_ENABLE_CUDA=OFF
cmake --build build -j
PYTHONPATH=python pytest tests/python/test_api.py

C / pkg-config

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j
cmake --install build --prefix /usr/local
pkg-config --libs --cflags opencuda

Java (JNA, same DLL/.so/.dylib)

cd bindings/java
mvn -q package
# Windows: put opencuda.dll on PATH or java.library.path
mvn -q exec:java