PrunaAI/pruna

Combines multiple compression techniques—quantization, pruning, distillation, caching, and compilation—through a composable `smash()` API that chains algorithms for synergistic optimization. Supports diverse model architectures including LLMs, diffusion models, vision transformers, and speech models across PyTorch and other frameworks. Includes built-in evaluation tools to benchmark speed, memory, and quality trade-offs across different optimization configurations.

1,142 stars. Actively maintained with 21 commits in the last 30 days.