Search papers, labs, and topics across Lattice.
The paper introduces RTPurbo, a method to transform full-attention LLMs into sparse models by exploiting their intrinsic sparsity. RTPurbo retains the full KV cache only for retrieval heads and uses a lightweight token indexer for sparse attention based on query-dependent top-p selection. With only a few hundred training steps, RTPurbo achieves near-lossless accuracy and significant efficiency gains, including up to 9.36x prefill speedup at 1M context length.
Full-attention LLMs are intrinsically sparse and can be transformed into highly efficient sparse models with minimal training, sidestepping the need for expensive sparse pre-training.
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 9.36times prefill speedup at 1M context and about a 2.01times decode speedup. These results suggest that strong sparse inference can be obtained from standard full-attention training without expensive native sparse pretraining.
DAMO