6e068002f5
Teach nvbench_compare to parse GPU timing summaries into structured values and prefer the robust median/IQR summaries when both compared measurements provide them. Fall back to the existing mean/stdev summaries when robust summaries are not available. Classify comparisons with the larger available relative noise estimate instead of the smaller one, keep unavailable noise distinct from encoded infinite noise, and report improvements separately from regressions. Keep the process exit code as success for completed comparisons; regression counts are reported in the summary instead of being used as the process status. Make plotting tolerate unavailable noise by leaving gaps in confidence bands, sort plotted series by the plotted axis, and avoid reusing pyplot state across plot calls. Add focused Python tests for robust-summary preference, unavailable-noise classification, non-finite timing centers, plot-along handling when the selected axis is absent, and the exit-code contract.
CUDA Kernel Benchmarking Package
This package provides a Python API to the CUDA Kernel Benchmarking
Library NVBench.
Installation
Install from PyPI:
python -m pip install cuda-bench
Use an optional dependency if you want pip to install a compatible
cuda-bindings package as well:
python -m pip install "cuda-bench[cu12]" # Install cuda-bindings 12.x
python -m pip install "cuda-bench[cu13]" # Install cuda-bindings 13.x
The published Linux wheel is compatible with both CUDA 12.x and CUDA 13.x
Python environments. It contains two native extensions: one built with a CUDA
12.x Toolkit and installed under cuda.bench.cu12, and one built with a CUDA
13.x Toolkit and installed under cuda.bench.cu13. At runtime, cuda-bench
queries the installed cuda.bindings package to determine the CUDA major
version and loads the matching native extension.
The cu12 and cu13 extras do not select different cuda-bench wheels. They
only select the compatible cuda-bindings dependency family. If your
environment already provides an appropriate cuda-bindings 12.x or 13.x
package, installing plain cuda-bench is sufficient.
A local CUDA Toolkit is not required when installing a published wheel, but the
NVIDIA driver must support the CUDA runtime used by the installed
cuda.bindings package. Use the same CUDA major version for other CUDA Python
binary packages in the environment, for example cupy-cuda12x with
cuda-bench[cu12] or cupy-cuda13x with cuda-bench[cu13].
Building from source
Ensure recent version of CMake
Since nvbench requires CMake >=3.30.4, either install a recent CMake or
create a conda environment with CMake and Ninja:
conda create -n build_env --yes cmake ninja
conda activate build_env
Ensure CUDA compiler
Building cuda-bench from source requires a CUDA Toolkit with nvcc. Ensure
that the appropriate environment variables are set. For example, on Linux,
assuming the CUDA Toolkit is installed system-wide:
export CUDACXX=/usr/local/cuda/bin/nvcc
export CUDAARCHS=all-major
Unlike the published wheel, a local source build only builds the native extension for the CUDA Toolkit found by CMake. The CUDA major version selected in the install command below must match that Toolkit.
Build Python project
Now switch to the Python package directory and install cuda-bench from source:
cd nvbench/python
python -m pip install ".[cu12]" # If CUDACXX points to a CUDA 12.x toolkit
python -m pip install ".[cu13]" # If CUDACXX points to a CUDA 13.x toolkit
Editable installs (python -m pip install -e .) are currently not supported.
They do not install the versioned CUDA extension layout used by cuda-bench.
Re-run the non-editable install command after making source changes.
Verify that package works
python test/run_1.py
Run examples
# Example benchmarking numba.cuda kernel
python examples/throughput.py
# Example benchmarking kernels authored using cuda.core
python examples/axes.py
# Example benchmarking algorithms from cuda.cccl.parallel
python examples/cccl_parallel_segmented_reduce.py
# Example benchmarking CuPy function
python examples/cupy_extract.py