e-GPU for running pre-trained PyTorch models from a macbook

Hi everyone,
I am seeking for advice before investing in an e-GPU.

Doing PhD experiments in deep learning for sound synthesis and audio effects.
I would like to make some demonstrations, for instance when presenting in conference, of my models running real-time.

So the setup is not for training models, but for running them in real-time from my macbook. For this an average GPU and 4GB memory should be enough.

I have seen the Nvidia Jetson Nano and consider it later for doing standalone instruments.

But for now, I would like to ask about e-GPUs that are Mac OS compatible, with cuda and pytorch support.
Are there some rather straight-forward solutions nowadays ? That are well compatible and allow for easy deployment from training models on servers to using them in live conditions from a laptop ?
Regarding budget, we could go up to 500€ for the enclosure and 500€ for the card.

If needed, few more details about the implementations. I already experimented with exporting models with LibTorch to C++ and embedding them as PureData externals to make realtime instruments and plugins. I want to experiment more with this kind of deployment and would thus need an e-GPU to avoid latency.

Thanks in advance and happy learning everyone !