Not totally sure if we should merge this now, or wait for tf 2.14, but
figured I could put it up anyway so people could use it. With
https://github.com/tensorflow/tensorflow/pull/59825
tf-nightly can be installed using cuda pip packages. Which means we
can write a recipe for cross framework GPU support.
To install a local development version...
```shell
pip install -r requirements-cuda.txt
python pip_build.py --install
```
To install the official pip version...
```shell
pip install -r requirements-cuda.txt
pip install keras-core --no-deps
```
Note that `--no-deps` is required to avoid pulling in `tensorflow` and
`tf-nightly` at the same time.
This should work in a clean python env, as long nvidia drivers are
>=520.61.05. No conda or cuda shenanigans required!