cuQuantum Benchmark Suite (cuquantum-benchmarks) v0.3.0 is released!

[leofang]

Hello cuQuantum users, we are pleased to announce the release of NVIDIA cuQuantum Performance Benchmark Suite v0.3.0 🚀

New Features & Enhancements

• Add more API-level benchmarks. See cuquantum-benchmarks api --help and e.g. cuquantum-benchmarks api --benchmark tensor_decompose --help for more details.
  • apply_generalized_permutation_matrix
  • cusv_sampler
  • tensor_decompose
• Improvements to the cutn backend:
  • Support --nhypersamples CLI option to control the number of hypersamples.
  • The contraction optimizer outcome is printed if the benchmark is run with the verbose mode (-v).
• All benchmarks are now covered by the NVTX ranges for ease of profiling.
  • This requires the nvtx Python package, installable via pip install nvtx or conda install -c conda-forge nvtx.
• The help messages for CLI options are now clearer.
  • As the benchmark suite grows, some CLI options are dynamically generated/recognized (based on the top-level inputs, such as "api"/"circuit"). Be sure to do the --help query for the selected benchmark and/or frontend-backend.
• Internal code refactoring and cleanup.

Bugs fixed

• Fix a bug that the subdirectories in the cache dir might not be created correctly.
• Fix a bug in our Qiskit circuit generator.
• Various fixes for the cusvaer backend:
  • Fix an unexpected device index is set.
  • Fix unnecessary synchronization.

Installation Guide

Same as in the previous release, it can be easily installed by cloning this repository as follows:

$ git clone https://github.com/NVIDIA/cuQuantum.git
$ cd cuQuantum/benchmarks/
$ pip install .
$ cuquantum-benchmarks --help

This benchmark suite works either without or with our cuQuantum Appliance container. For further information please see the benchmark project README.

Happy benchmarking!