Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps

TL;DR

This paper introduces RTPurbo, a method to transform full-attention large language models into highly sparse models by exploiting intrinsic sparsity, achieving significant efficiency gains with minimal training steps.

Contribution

The authors demonstrate that full-attention LLMs are inherently sparse and propose RTPurbo to efficiently convert them into sparse models without extensive pretraining.

Findings

RTPurbo achieves up to 9.36× prefill speedup at 1M context.

It maintains near-lossless accuracy compared to full attention models.

Substantial efficiency gains are achieved with only a few hundred training steps.

Abstract

Long-context inference in large language models is bottlenecked by the quadratic cost of full attention. Existing efficient alternatives often rely either on native sparse training or on heuristic token eviction, creating an undesirable trade-off among efficiency, training cost, and accuracy. In this work, we show that full-attention LLMs are already intrinsically sparse and can be transformed into highly sparse models with only minimal adaptation. Our approach is built on three observations: (1) only a small subset of attention heads truly requires full long-context processing; (2) long-range retrieval is governed primarily by a low-dimensional subspace, allowing relevant tokens to be retrieved efficiently with a 16-dimensional indexer; and (3) the useful token budget is strongly query-dependent, making dynamic top-$p$ selection more suitable than fixed top-$k$ sparsification. Based on these insights, we propose RTPurbo, which retains the full KV cache only for retrieval heads and introduces a lightweight token indexer for sparse attention. By exploiting the model's intrinsic sparsity, RTPurbo achieves sparsification with only a few hundred training steps. Experiments on long-context benchmarks and reasoning tasks show that RTPurbo preserves near-lossless accuracy while delivering substantial efficiency gains, including up to a 9.36$\times$ prefill speedup at 1M context and about a 2.01$\times$ decode speedup. These results suggest that strong sparse inference can be obtained from standard full-attention training without expensive native sparse pretraining.