Success running GPU-accelerated Stable Diffusion on Rock 5!


With the help of TVM Unity I have had a success running Stable Diffusion on the Rock 5B!

Currently generate a 512x512 image costs about 500 seconds (including model loading and GPU kernel compilation time. The actual inference time is less). The U-Net runs at 21sec per iteration.

The model is large, you probably need a >=16GB version of the board to run this.

Code: https://github.com/happyme531/RK3588-stable-diffusion-GPU (please star!!)

4 Likes

Congratulations. Is it possible to skip the first steps and proceed to the inference after they are done? And which GPU driver does it use? Is it libmali or Panfork (Panfrost)?

The model convert only need done once.

Currently Im using the stock libmali driver.