nb::any is removed in nanobind v2.0.0

[Huizerd]

I tried to build aestream with libcaer and CUDA support, but ran into errors. It turns out that nanobind recently removed nb::any support in nb::shape, and that one should now use -1 instead. This PR fixes that here, and allows me to compile successfully with pip install git+https://github.com/Huizerd/aestream@dev/nb_any_bug --no-binary aestream -v.

Tested on Ubuntu 20.04, with gcc-10 and g++-10 selected via update-alternatives as explained here. I have the latest libcaer master from here and CUDA toolkit 12.1 from here.

I also got it built successfully in a conda environment:

name: cuda_minimal
channels:
  - pytorch
  - nvidia
dependencies:
  - python=3.11
  - pytorch-cuda=12.1
  - pytorch==2.1.2
  - torchvision==0.16.2
  - cuda=12.1
  - cuda-nvcc=12.1
  - cuda-version=12.1
  - pip
  - pip:
    - cmake
    - numpy
    - setuptools
    - wheel

Also, when I didn't have g++-10 set as default and tried to build with CUDA support, I found that setting only CXX=g++-10 wasn't enough; you also have to set the host compiler for CUDA as follows:

CXX=g++-10 NVCC_PREPEND_FLAGS='-ccbin g++-10' pip install git+https://github.com/Huizerd/aestream@dev/nb_any_bug --no-binary aestream -v

Maybe this could be added in the docs :)

[Jegp]

What a great and thorough contribution, thank you @Huizerd!

I wonder why the GPG check is failing in YUM, but I don't think you're to blame for that.

Your point of the NVCC flag would be great to have in the docs, as you suggested. Do you want to add that (https://github.com/aestream/aestream/blob/main/docs/install.md) or would you prefer I did it?

[Huizerd]

Done!

[Jegp]

Brilliant. Thank you for the addition 👍
I'll take a look at the failing builds in another branch.