SLA2: Sparse-Linear Attention with Learnable Routing and QAT

TL;DR

SLA2 enhances sparse-linear attention in diffusion models by introducing a learnable routing, a more accurate attention formulation, and quantization, achieving high sparsity and significant speedup without quality loss.

Contribution

SLA2 proposes a learnable routing and a faithful sparse-linear attention formulation, improving efficiency and accuracy in diffusion model attention mechanisms.

Findings

Achieves 97% attention sparsity in video diffusion models.

Provides 18.6x attention speedup while maintaining quality.

Introduces quantization-aware fine-tuning for low-bit attention.

Abstract

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.