model-pruning

Reduces LLM size via Wanda or SparseGPT pruning workflows.

Updated Feb 15, 2026

Search Tags:#sparsity #inference acceleration #llm compression #model pruning #wanda #sparsegpt #n:m pruning

Ready to run

npx skills add https://github.com/ricable/mcai --skill model-pruning-ricable

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: model-pruning
Source: https://github.com/ricable/mcai/tree/main/.agents/skills/model-pruning
Command: npx skills add https://github.com/ricable/mcai --skill model-pruning-ricable

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill helps compress large language models without retraining so they use less memory, run faster, and stay accurate enough for practical deployment.

Core Features & Use Cases

One-shot pruning workflows based on Wanda-style weight and activation scoring.
Higher-quality layer-wise compression using SparseGPT reconstruction methods.
Hardware-friendly N:M sparsity for accelerator-ready inference speedups.
Use it to reduce model size, lower serving costs, or prepare a model for constrained hardware.

Quick Start

Ask the assistant to prune your target language model to a chosen sparsity level using Wanda or SparseGPT and save the compressed checkpoint.

Frequently Asked Questions about model-pruning

High-intent search queries and answers about installing and using this skill.

FAQPage Schema

How do I compress a large language model to reduce inference cost without retraining?▼

Compress large language models without retraining using one-shot pruning workflows based on Wanda or SparseGPT. This reduces model size and inference cost while maintaining accuracy through layer-wise masking and weight-activation importance scoring.

What is the best way to apply N:M sparsity to LLMs for hardware-friendly acceleration?▼

Apply N:M sparsity to LLMs for hardware-friendly acceleration using structured pruning workflows. This creates accelerator-ready sparse models that deliver inference speedups on compatible hardware while satisfying calibration data handling requirements.

How does Wanda pruning compare to SparseGPT reconstruction for LLM compression?▼

Wanda pruning uses weight and activation scoring for one-shot compression, while SparseGPT offers higher-quality layer-wise compression through reconstruction. Both reduce model size without retraining, but SparseGPT typically preserves more accuracy at higher sparsity levels.

Can I prune a large language model for constrained hardware without losing too much accuracy?▼

Prune large language models for constrained hardware with minimal accuracy loss using calibration-based pruning. Methods like SparseGPT apply layer-wise masking and importance scoring to preserve model quality while achieving target sparsity levels.

Do I need calibration data to perform one-shot pruning on an LLM checkpoint?▼

Calibration data handling is required for one-shot pruning workflows. The process uses calibration samples to score weight and activation importance, enabling accurate layer-wise masking when applying Wanda or SparseGPT compression to the target model checkpoint.