Zheng Tian

Jiayi Yuan, Cameron Shinn, Kai Xu et al.

The growing demand for long-context inference capabilities in Large Language Models (LLMs) has intensified the computational and memory bottlenecks inherent to the standard attention mechanism. To address this challenge, we introduce BLASST, a drop-in sparse attention method that dynamically prunes the attention matrix without any pre-computation or proxy scores. Our method uses a fixed threshold and existing information from online softmax to identify negligible attention scores, skipping softmax computation, Value block loading, and the subsequent matrix multiplication. This fits seamlessly into existing FlashAttention kernel designs with negligible latency overhead. The approach is applicable to both prefill and decode stages across all attention variants (MHA, GQA, MQA, and MLA), providing a unified solution for accelerating long-context inference. We develop an automated calibration procedure that reveals a simple inverse relationship between optimal threshold and context length, enabling robust deployment across diverse scenarios. Maintaining high accuracy, we demonstrate a 1.62x speedup for prefill at 74.7% sparsity and a 1.48x speedup for decode at 73.2% sparsity on modern GPUs. Furthermore, we explore sparsity-aware training as a natural extension, showing that models can be trained to be inherently more robust to sparse attention patterns, pushing the accuracy-sparsity frontier even further.

Mingcheng Chen, Haoran Zhao, Yuxiang Zhao et al.

Data-driven black-box model-based optimization (MBO) problems arise in a great number of practical application scenarios, where the goal is to find a design over the whole space maximizing a black-box target function based on a static offline dataset. In this work, we consider a more general but challenging MBO setting, named constrained MBO (CoMBO), where only part of the design space can be optimized while the rest is constrained by the environment. A new challenge arising from CoMBO is that most observed designs that satisfy the constraints are mediocre in evaluation. Therefore, we focus on optimizing these mediocre designs in the offline dataset while maintaining the given constraints rather than further boosting the best observed design in the traditional MBO setting. We propose retrieval-enhanced offline model-based optimization (ROMO), a new derivable forward approach that retrieves the offline dataset and aggregates relevant samples to provide a trusted prediction, and use it for gradient-based optimization. ROMO is simple to implement and outperforms state-of-the-art approaches in the CoMBO setting. Empirically, we conduct experiments on a synthetic Hartmann (3D) function dataset, an industrial CIO dataset, and a suite of modified tasks in the Design-Bench benchmark. Results show that ROMO performs well in a wide range of constrained optimization tasks.