Shaoyi Zheng

Wenbo Lu, Shaoyi Zheng, Yuxuan Xia et al.

Diffusion models excel in high-fidelity image generation but face scalability limits due to transformers' quadratic attention complexity. Plug-and-play token reduction methods like ToMeSD and ToFu reduce FLOPs by merging redundant tokens in generated images but rely on GPU-inefficient operations (e.g., sorting, scattered writes), introducing overheads that negate theoretical speedups when paired with optimized attention implementations (e.g., FlashAttention). To bridge this gap, we propose Token Merge with Attention (ToMA), an off-the-shelf method that redesigns token reduction for GPU-aligned efficiency, with three key contributions: 1) a reformulation of token merge as a submodular optimization problem to select diverse tokens; 2) merge/unmerge as an attention-like linear transformation via GPU-friendly matrix operations; and 3) exploiting latent locality and sequential redundancy (pattern reuse) to minimize overhead. ToMA reduces SDXL/Flux generation latency by 24%/23%, respectively (with DINO $Δ< 0.07$), outperforming prior methods. This work bridges the gap between theoretical and practical efficiency for transformers in diffusion.

Xin Teng, Canyu Zhang, Shaoyi Zheng et al.

Retrieval-augmented generation (RAG) for long-context question answering is bottlenecked by inference-time prefilling over large retrieved contexts. A common strategy is to precompute key-value (KV) caches for individual documents and selectively recompute a small subset of tokens to restore global causal dependencies, but existing methods rely on heuristics or representation discrepancies without modeling whether selected tokens can effectively influence generation. We cast selective KV recomputation as an information flow problem and show that a simple attention-norm signal from the query reliably identifies tokens that are both semantically relevant and structurally positioned to propagate information, when computed under an inference-consistent RoPE geometry. We therefore reconstruct global positional assignments for retrieved chunks and introduce an information-flow-guided chunk reordering strategy. Experiments on LLM and VLM benchmarks demonstrate consistent gains over prior methods under comparable efficiency budgets.

Shaoyi Zheng, Canyu Zhang, Tianyi Zhou et al.

In-context learning (ICL) enables efficient few-shot learning in large language models (LLMs) without training, but suffers from the quadratic input complexity of transformers, limiting the maximum number of exemplars. While various efficient ICL approaches partition the context into blocks to process (e.g., ensembling, compression, cross-attention), they often ignore the information redundancy or under-representation caused by different partition strategies, leading to suboptimal performance. To tackle this problem, we propose Sub-CP, a block-aware context selection framework that leverages submodular objectives to control block diversity. Sub-CP supports a flexible spectrum of selection strategies, allowing each block to range from globally diverse to locally coherent. This allows fine-grained control over semantic structure while enabling precomputation. Extensive experiments across diverse tasks on multiple datasets show that Sub-CP consistently improves performance across model scales.

Shaoyi Zheng, Wenbo Lu, Yuxuan Xia et al.

Designing sparse attention for diffusion transformers requires reconciling two-dimensional spatial locality with GPU efficiency, a trade-off that current methods struggle to achieve. Existing approaches enforce two-dimensional spatial locality but often incur uncoalesced memory access. We present HilbertA, a 2D-aware and GPU-efficient sparse attention mechanism. HilbertA reorders image tokens along Hilbert curves to achieve a contiguous memory layout while preserving spatial neighborhoods, and employs a sliding schedule across layers to enable long-range information propagation without repeated or uncoalesced memory access. To further enhance cross-tile communication and positional awareness, HilbertA introduces a small central shared region. Implemented in Triton, HilbertA delivers comparable image quality with significant acceleration over prior methods on Flux.1-dev, demonstrating the feasibility of hardware-aligned two-dimensional sparse attention for high-resolution image generation. HilbertA delivers attention speedups of $2.3\times$ when generating $1024\times 1024$ images, and up to $4.17\times$ at $2048\times 2048$, while achieving image quality comparable to or surpassing baselines.