From the repository root:
pip install .Import works without a CUDA toolkit:
from tailwarp import TailWarpClient
client = TailWarpClient()
result = client.compute_cvar([-0.1, -0.05, 0.0, 0.02], alpha=0.95)
print(result.value, result.method, result.device)# Documents CUDA-native wheel path (extension built separately via CMake)
pip install ".[cuda]"
# Dashboard dependencies
pip install ".[dashboard]"Build host-only pybind11 module (no GPU required for compute_cvar):
cmake -B build -DBUILD_PYTHON=ON -DBUILD_TESTS=OFF
cmake --build build --target _native
# Copy or install _native*.so into python/tailwarp/ for local devWhen _native is missing, all public methods use the NumPy CPU fallback.