Polar Sparsity Images

Polar Sparsity Images

This directory contains images and figures for the Polar Sparsity research page.

Current Placeholders

The following placeholder images need to be replaced with actual figures from the paper:

  1. placeholder.svg - Generic placeholder for all figures

Figures to Add

From the NeurIPS 2025 paper, extract and add the following figures:

  1. Figure 1a - Transformer decode latency breakdown
    • Shows how attention layers dominate as batch size increases
    • Bar chart comparing MLP vs Attention latency across batch sizes
  2. Figure 6b - OPT-66B Decode Throughput
    • Comparison of Dense, Deja Vu, and Polar Sparsity
    • Bar chart across different batch sizes
  3. Additional figures to consider:
    • Sparsity pattern visualizations
    • Accuracy vs speedup tradeoff graphs
    • Kernel performance breakdowns
    • Model architecture diagrams

Image Format Guidelines

  • Use PNG or SVG format for best quality
  • Recommended resolution: At least 1200px wide for main figures
  • Optimize images for web (compress without losing quality)
  • Use descriptive filenames (e.g., latency-breakdown-batch-size.png)

How to Replace Placeholders

  1. Export figures from the paper (high resolution)
  2. Save them in this directory with descriptive names
  3. Update the image src in /Users/susavshrestha/kernels/susavlsh10.github.io/_pages/polar-sparsity.md
  4. Update alt text and captions as needed