Event Tensor: A Unified Abstraction for Compiling Dynamic Megakernel

TL;DR

Event Tensor introduces a unified compiler abstraction that effectively handles dynamic shapes and data-dependent computations in GPU megakernels, improving LLM inference performance.

Contribution

It proposes Event Tensor, a novel abstraction enabling dynamic megakernels to better support real workloads with shape and data-dependent dynamism.

Findings

ETC achieves state-of-the-art LLM serving latency.

ETC significantly reduces system warmup overhead.

Abstract

Modern GPU workloads, especially large language model (LLM) inference, suffer from kernel launch overheads and coarse synchronization that limit inter-kernel parallelism. Recent megakernel techniques fuse multiple operators into a single persistent kernel to eliminate launch gaps and expose inter-kernel parallelism, but struggle to handle dynamic shapes and data-dependent computation in real workloads. We present Event Tensor, a unified compiler abstraction for dynamic megakernels. Event Tensor encodes dependencies between tiled tasks, and enables first-class support for both shape and data-dependent dynamism. Built atop this abstraction, our Event Tensor Compiler (ETC) applies static and dynamic scheduling transformations to generate high-performance persistent kernels. Evaluations show that ETC achieves state-of-the-art LLM serving latency while significantly reducing system warmup overhead.