Efficient Long-context Language Model Training by Core Attention Disaggregation

TL;DR

This paper introduces core attention disaggregation (CAD), a technique that enhances long-context language model training by balancing attention computation across devices, significantly improving throughput and reducing stragglers.

Contribution

The paper proposes CAD, a novel method to decouple and distribute core attention computation, enabling efficient training of long-context models on large GPU clusters.

Findings

Up to 1.35x increase in training throughput on 512 GPUs.

Elimination of data and pipeline parallel stragglers.

Near-perfect compute and memory balance achieved.

Abstract

We present core attention disaggregation (CAD), a technique that improves long-context large language model training by decoupling the core attention computation, softmax(QK^T)V, from the rest of the model and executing it on a separate pool of devices. In existing systems, core attention is colocated with other layers; at long context lengths, its quadratic compute growth compared to the near-linear growth of other components causes load imbalance and stragglers across data and pipeline parallel groups. CAD is enabled by two observations. First, core attention is stateless: it has no trainable parameters and only minimal transient data, so balancing reduces to scheduling compute-bound tasks. Second, it is composable: modern attention kernels retain high efficiency when processing fused batches of token-level shards with arbitrary lengths. CAD partitions core attention into token-level tasks and dispatches them to dedicated attention servers, which dynamically rebatch tasks to equalize compute without sacrificing kernel efficiency. We implement CAD in a system called DistCA, which uses a ping-pong execution scheme to fully overlap communication with computation and in-place execution on attention servers to reduce memory use. On 512 H200 GPUs and context lengths up to 512k tokens, DistCA improves end-to-end training throughput by up to 1.35x, eliminates data and pipeline parallel stragglers, and achieves near-perfect compute and memory balance.