Well, Mathematica is lispy, and there is:

https://www.wolfram.com/data-science-platform/

I suppose Hylang is quite slow (and possibly has leaky abstractions). Ideally the main C/C++ libraries like libtorch should be wrapped in Common Lisp and not Python.

You have a few options for this: The TH package wraps Blas, etc., and has MINST, GAN, etc. examples.

My preference is using py4cl to use pre-trained TensorFlow models in Common Lisp.

I wrote a short book on Hy that has some deep learning examples, but I don't use Hy that often any longer. I am looking forward to the 1.0 changes.