Efficient Attention Mechanism for Large Language Models

Plain English Summary

This patent covers a faster way to process text in AI language models, making them handle much longer documents without the usual slowdown that comes from processing every word against every other word.

Full Description

Key Claims

1. 1

   Locality-sensitive hashing for approximate attention computation

2. 2

   Sparse attention pattern selection via learned routing

3. 3

   Memory-efficient gradient computation for long sequences

4. 4

   Hardware-optimized kernel for reduced-precision attention