VainF/Torch-Pruning

Health v2 · maintenance only

Adoption not part of the score

3359 stars · 387 forks observed · 2026-10-02

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

Torch-Pruning is a PyTorch framework for structural neural network pruning based on the DepGraph algorithm from CVPR 2023. It automatically groups coupled parameters across arbitrary architectures (CNNs, Transformers, LLMs, etc.) and physically removes them to compress and accelerate models.

Use cases

• prune a pytorch model to make it faster and smaller
• compress large language models with structural pruning
• prune huggingface transformers or timm models
• remove channels from a resnet or vision transformer
• speed up inference by slimming yolov8 or sam models
• compare pruning criteria like magnitude or taylor on my network

When to choose

• you need architecture-agnostic structural pruning that physically removes parameters rather than masking them
• you want to prune modern architectures like ViTs, LLMs, diffusion models, or detection models
• you need a research-grade toolkit with many ready-made pruning examples

When to avoid

• you only need unstructured sparsity or weight masking, which torch.nn.utils.prune already covers
• you need post-training quantization or distillation rather than pruning
• you work outside PyTorch, e.g. TensorFlow or JAX

Facets

library · maturity active

machine-learning deep-learning developer-tools deep-learning machine-learning computer-vision large-language-models python cross-platform pruning model-compression pytorch depgraph structural-pruning model-acceleration transformers llm gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="VainF/Torch-Pruning")