skills.rest
tadod12
tadod12COMMUNITY
Explore all skills by this author →
GitHub

model-pruning

Prune large language model weights using activation-aware Wanda methods.

Updated Mar 18, 2026
Search Tags:#llm#sparsity#wanda#sparsegpt#model-pruning#nm-sparsity#structured-pruning
One-click install
npx skills add https://github.com/tadod12/fraud-detection-research --skill model-pruning-tadod12
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: model-pruning
Source: https://github.com/tadod12/fraud-detection-research/tree/main/.agent/skills/19-emerging-techniques/model-pruning
Command: npx skills add https://github.com/tadod12/fraud-detection-research --skill model-pruning-tadod12

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

Model pruning reduces the computational footprint of large language models by removing redundant weights and enabling sparsity patterns that accelerate inference and reduce memory usage, often without retraining.

Core Features & Use Cases

  • Activation-aware pruning: use Wanda or similar methods to prune weights based on weight magnitudes and activation statistics.
  • Support for multiple sparsity styles: unstructured, N:M structured, and one-shot pruning with minimal accuracy loss.
  • Hardware-friendly deployment: enables faster inference on accelerators and reduces memory footprint for edge deployments.

Quick Start

Provide calibration data and run a Wanda pruning pass to prune weights with activation-aware criteria, achieving target sparsity without retraining.

Frequently Asked Questions about model-pruning

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

FAQPage Schema
How do I prune LLM weights without retraining the model?▼

To prune large language models, provide calibration data and run a Wanda pruning pass to remove weights based on magnitudes and activation statistics, achieving target sparsity without retraining.

What is N:M sparsity and how does it accelerate inference?▼

N:M structured sparsity enforces a fixed ratio of zero to non-zero weights within matrix blocks, creating hardware-friendly patterns that accelerate inference and reduce memory usage on edge deployments.

Do I need calibration data to perform activation-aware pruning?▼

Yes, activation-aware pruning requires calibration data and activation statistics to evaluate weight importance accurately and preserve model performance during sparsity management.

When should I use unstructured pruning versus structured sparsity?▼

Use unstructured pruning for maximum weight removal without pattern constraints, whereas structured N:M sparsity provides hardware-friendly deployment with better inference speedups on accelerators.

What are the limitations of one-shot pruning for large language models?▼

One-shot pruning may cause accuracy loss at high sparsity levels, requiring optional fine-tuning to manage sparsity levels and preserve performance across various models.

Related Skills

Explore More Data & Analytics Skills→
Supporter09Supporter09

model-pruning

Prune large language models with Wanda or SparseGPT-style methods.

Community
Advanced
hochoa13hochoa13

llama-cpp

Run GGUF quantized LLM inference on CPUs, Apple Silicon, and consumer GPUs.

Community
Intermediate
ihatesea69ihatesea69
3

awq-quantization

Quantize large language models to 4-bit weights using AWQ.

Community
Advanced
RheasilviaRheasilvia

obliteratus

Remove refusal behaviors from open-weight LLMs by projecting out refusal directions.

Community
Advanced