Binary distribution for cuQuantum support

[doichanj]

Summary

Add binary distribution for cuQuantum support

Details and comments

Added binary build workflow in .github/workflow/deploy.yml for cuQuantum support.

User will be able to install by pip install qiskit-aer-cuQuantum

[doichanj]

https://github.com/doichanj/qiskit-aer/runs/7084200066?check_suite_focus=true

[doichanj]

I think large file size issue is a common problem for CUDA 11, but cuQuantum required CUDA 11 or above.

https://discuss.python.org/t/what-to-do-about-gpus-and-the-built-distributions-that-support-them/7125

[mtreinish]

Since we're hitting issues with pypi file size, maybe we can use: https://github.com/softprops/action-gh-release to add the wheels as a release artifacts on github for the time being.

[leofang]

@doichanj we're preparing detailed instruction for cuQuantum downstream to avoid hitting the file size limit, please stay tuned 🙂

Generally, if it's only ~300 MB it's possible to request size limit bump by submitting a ticket to https://github.com/pypa/pypi-support/issues, but we do not consider it a nice experience for users and developers. We'd rather advocate for the following changes:

• use dynamic instead of static linking
  • cuQuantum already have wheels such as custatevec-cu11 and cutensornet-cu11 living on PyPI
  • NVIDIA CUDA wheels are also up (nvidia-*-cu11 for CUDA 11 and nvidia-*-cu12 for CUDA 12, ex: nvidia-cublas-cu11)
• use auditwheel repair --exclude (https://github.com/pypa/auditwheel/pull/368) to whitelist certain shared libraries
  • it's possible to not do repair at all, if you already build the wheel in a manylinux container and meet certain conditions

However there are some caveats in terms of the build system, we'll try to explain them in the upcoming instruction. This is how our cuquantum-python-cuXX wheels can be kept under very comfortable size.

[leofang]

@doichanj we're preparing detailed instruction for cuQuantum downstream to avoid hitting the file size limit, please stay tuned 🙂

FYI: https://github.com/NVIDIA/cuQuantum/tree/main/extra/demo_build_with_wheels, let us know if it helps! 🙂

[hhorii]

no longer necessary because NVIDIA provides appliance of qiskit-aer-gpu with cuQuantum.

[leofang]

Hi @hhorii, would it be possible to keep this PR opened? Even with our cuquantum Appliance container, we can't cover all use cases, and according to our understanding (@tlubowe knows far better than I do 🙂) the majority of users still want a simple pip-install based solution, without container. If you don't have bandwidth for this PR, maybe @tlubowe can find someone on our side to help?