VainF/Torch-Pruning

Leverages DepGraph, a dependency-aware algorithm that automatically identifies and groups coupled parameters across layers to enable true structural pruning—removing entire channels/filters rather than masking individual weights. Supports diverse architectures including LLMs (Llama, Qwen, DeepSeek), vision transformers, diffusion models, and detection frameworks, with examples from Huggingface, Timm, and Torchvision. Provides both high-level pruning APIs for common strategies (global, isomorphic, sparse training) and low-level functions for custom pruning workflows on PyTorch 1.x and 2.x.

3,267 stars and 21,337 monthly downloads. Used by 1 other package. No commits in the last 6 months. Available on PyPI.