Polar Sparsity: High Throughput Batched LLM Inferencing with Scalable Contextual Sparsity

Susav Shrestha, Bradley Settlemyer, Nikoli Dryden, A. L. Narasimha Reddy

Keywords: Large Language Models, Contextual Sparsity, Activation Sparsity, Head Sparsity, Select Head Attention, Batched Inference, Efficient Transformers, Machine Learning Systems

TL;DR: Polar Sparsity scales contextual sparsity to large batches by exploiting stable attention head sparsity and using efficient GPU kernels, achieving up to 2.2× speedups with minimal accuracy loss.

Abstract: Accelerating large language model (LLM) inference is critical for real-world deployments requiring high throughput and low latency. Contextual sparsity, where each token dynamically activates only a small subset of the model parameters, shows promise but does not scale to large batch sizes due to union of active neurons quickly approaching dense computation. We introduce Polar Sparsity, highlighting a key shift in sparsity importance from MLP to Attention layers as we scale batch size and sequence length. While MLP layers become more compute-efficient under batching, their sparsity vanishes. In contrast, attention becomes increasingly more expensive at scale, while their head sparsity remains stable and batch-invariant. We develop Selective Head Attention with hardware-efficient, sparsity-aware GPU kernels, delivering up to \(2.2\times\) end-to-end speedups for models like OPT, LLaMA-2 \& 3, Qwen, Mistral across various batch sizes and sequence lengths without compromising accuracy. To our knowledge, this is the first work to demonstrate that contextual sparsity can scale effectively to large batch sizes, delivering substantial inference acceleration with minimal changes, making Polar Sparsity practical for large-scale, high-throughput LLM deployment systems.

Primary Area: Deep learning (e.g., architectures, generative models, optimization for deep networks, foundation models, LLMs)

Submission Number: 9859