Hugging Face link with 25+ models :)<p>At Pruna.AI we built a tool that automatically implements and combines optimizations and ML compression algorithms for your model/hardware to make them more efficient at inference. We decided to run it on the most popular Stable Diffusion models on an NVIDIA A100 and publish the models and gains on Hugging Facen, so you can all leverage them for your AI image generation projects.<p>The instructions and comparison metrics are on each model README, with about 2x speedup for A100. The models are compatible with most other NVIDIA GPUs. The first couple of generations are slower so warm up the model to get the full speedup.<p>Soon we'll update these models with even more gains and publish others for LLMs, Computer Vision etc<p>According to each use-case our tool can do a mix of methods including quantization, compilation, pruning, other architecture-specific algorithms etc. We're always adding new algorithms to the mix and making the implementation frictionless by providing easy config options and improving compatibility with new setups.<p>If you have custom requirements or finetuned models you'll need to leverage our tool yourself to make your models more efficient, so feel free to contact us for that.
In the meantime we're happy if we can make efficient ML more accessible to everyone, starting with these stable diffusion models on Hugging Face. Tell us if there's other ones you want us to publish on Hugging Face :)