Today we skip paper lists and do one deep dive.

  • Paper: SLA2: Sparse-Linear Attention with Learnable Routing and QAT
  • arXiv: https://arxiv.org/abs/2602.12675
  • Recency: within 7 days
  • Attention: 47 upvotes on HF Daily Papers

Why this paper matters

  • SLA2 improves sparse-linear attention in diffusion models by introducing a learnable router, direct attention formulation, and quantization-aware fine-tuning for enhanced efficiency and quality.

Problem framing

Sparse-Linear Attention (SLA) combines sparse and linear attention to accelerate diffusion models and has shown strong performance in video generation. However, (i) SLA relies on a heuristic split that assigns computations to the sparse or linear branch based on attention-weight magnitude, which can be suboptimal. Additionally, (ii) after formally analyzing the attention error in SLA, we identify a mismatch between SLA and a direct decomposition into sparse and linear attention. We propose SLA2, which introduces (I) a learnable router that dynamically selects whether each attention computation should use sparse or linear attention, (II) a more faithful and direct sparse-linear attention formulation that uses a learnable ratio to combine the sparse and linear attention branches, and (III) a sparse + low-bit attention design, where low-bit attention is introduced via quantization-aware fine-tuning to reduce quantization error. Experiments show that on video diffusion models, SLA2 can achieve 97% attention sparsity and deliver an 18.6x attention speedup while preserving generation quality.

Practical implementation checklist

  • Reproduce minimal path first.
  • Lock environment + seeds.
  • Run one-variable ablations.
  • Track quality/latency/cost together.
  • Document failure modes explicitly.