horseee/LLM-Pruner

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

1136 stars · 132 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

LLM-Pruner is a PyTorch library implementing structural pruning of large language models based on gradient information, as published at NeurIPS 2023. It compresses models like LLaMA, Llama-2/3, Vicuna, BLOOM, and Baichuan to smaller sizes with fast post-training recovery via LoRA.

Use cases

• prune a llama model to a smaller size
• compress large language models for cheaper inference
• structurally prune llm weights with gradient-based importance
• recover pruned model quality with lora fine-tuning
• shrink llama-3 or vicuna while keeping multi-task ability

When to choose

• you need task-agnostic size reduction of a supported LLM family
• you want fast pruning with minimal post-training data (50k samples)
• you are researching structural pruning methods for transformers

When to avoid

• you need quantization-only compression without structural changes
• your model architecture is not among the supported families
• you need production-grade serving rather than compression research tooling

Facets

library · maturity active

machine-learning llm-training deep-learning large-language-models machine-learning deep-learning python model-compression structural-pruning llama lora neurips-2023 pytorch gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="horseee/LLM-Pruner")