pruna

Why we included this project

Pruna suits teams whose transformer, diffusion, or speech models are close to shipping but struggle with latency, memory, or GPU bills. It packs quantization, pruning, distillation, caching, and hardware compilation behind a single smash() call, so you can stack several optimizations on one model instead of gluing together separate libraries. The API stays small: load a pretrained model as usual, pass a config listing the techniques you want, and get back an optimized version that still behaves like the original. An evaluation interface reports what the compressed model actually lost, which matters before you trust it in production.