Learning When to Attend: Conditional Memory Access for Long-Context LLMs

TL;DR

L2A introduces a conditional attention mechanism for long-context language models, enabling efficient long-range memory access, reducing computational costs, and maintaining performance with extended context lengths.

Contribution

The paper proposes L2A, a novel layer that selectively enables global attention, significantly extending context length and improving efficiency in long-context language models.

Findings

L2A extends context length from 32K to 128K tokens.

L2A reduces global attention computation by 80%.

L2A achieves up to 2x training throughput improvements.

Abstract

Language models struggle to generalize beyond pretraining context lengths, limiting long-horizon reasoning and retrieval. Continued pretraining on long-context data can help but is expensive due to the quadratic scaling of Attention. We observe that most tokens do not require (Global) Attention over the entire sequence and can rely on local context. Based on this, we propose L2A (Learning To Attend), a layer that enables conditional (token-wise) long-range memory access by deciding when to invoke global attention. We evaluate L2A on Qwen 2.5 and Qwen 3 models, extending their effective context length from 32K to 128K tokens. L2A matches the performance of standard long-context training to within 3% while skipping Global Attention for $\sim$80% of tokens, outperforming prior baselines. We also design custom Triton kernels to efficiently implement this token-wise conditional Attention on GPUs, achieving up to $\sim$2x improvements in training throughput and time-to-first-token over FlashAttention. Moreover, L2A enables post-training pruning of highly sparse Global Attention layers, reducing KV cache memory by up to 50% with negligible performance loss.