UPDATE: THIS REPO WILL BE ARCHIVED> THE UPDATED REPO CAN BE FOUND AT https://github.com/RegularJoe-CEO/Geodesic-Attention-Engine-GAE-

Waller Operator v1.0 - Memory-Efficient Attention for LLM Inference

The Waller Operator eliminates O(N²) memory bottlenecks in transformer attention, achieving O(N log N) memory and constant ~14ms latency across 4K-2.6M token sequences on NVIDIA H100.

🚀 Quick Start

Download waller_eval_x86 and run:

chmod +x waller_eval_x86
./waller_eval_x86 --help

Adjust CLI parameters for custom sequence lengths, batch sizes, head dimensions, and more:

./waller_eval_x86 --seq_len 131072 --batch_size 1 --num_heads 32 --head_dim 128

Requirements:

• NVIDIA H100 GPU (sm_90) or A100 (sm_80)
• CUDA 12.8
• x86_64 Linux

📊 Benchmark Results

See full results at: waller-eval/RESULTS.md

Key Performance:

• Memory: O(N log N) vs O(N²) standard attention
• Latency: ~14ms constant across 4K-2.6M tokens
• Equivalence: Mathematically identical to standard softmax attention
• No retraining required

🔬 Technical Details

• Built: Feb 2026
• Platform: x86_64 Linux
• CUDA Compute: sm_80 (A100), sm_90 (H100)
• Size: 0.97 MB

📜 Patent & Licensing

Patent Pending: US Provisional Application filed Feb 2026
Title: Memory-Efficient Attention Computation System and Method

Licensing inquiries: e@ewaller.com

🏢 Waller Systems

Breakthrough IP in AI inference, energy efficiency, and materials science.

Contact: e@ewaller.com
Web: luxiedge.com