Fast Locality Sensitive Hashing for Beam Search on GPU

TL;DR

This paper introduces a GPU-optimized LSH algorithm using WTA hashing and CUDA architecture design to accelerate beam search in sequence models, achieving significant speedups without loss of accuracy.

Contribution

The paper presents a novel GPU-based LSH algorithm with architecture-aware optimizations for faster beam search in neural sequence models.

Findings

Up to 4x speedup on softmax module

2x overall speedup on GPU

No degradation in BLEU scores

Abstract

We present a GPU-based Locality Sensitive Hashing (LSH) algorithm to speed up beam search for sequence models. We utilize the winner-take-all (WTA) hash, which is based on relative ranking order of hidden dimensions and thus resilient to perturbations in numerical values. Our algorithm is designed by fully considering the underling architecture of CUDA-enabled GPUs (Algorithm/Architecture Co-design): 1) A parallel Cuckoo hash table is applied for LSH code lookup (guaranteed O(1) lookup time); 2) Candidate lists are shared across beams to maximize the parallelism; 3) Top frequent words are merged into candidate lists to improve performance. Experiments on 4 large-scale neural machine translation models demonstrate that our algorithm can achieve up to 4x speedup on softmax module, and 2x overall speedup without hurting BLEU on GPU.