Skip to content

Sparse attention

Sparse attention targets model-level redundancy in long video token sequences. Many spatiotemporal attention interactions contribute little to the final output, so structured sparse patterns can reduce attention cost.

The paper lists sparse attention as a model-level family. The sparse-attention agent chooses the backend, sparsified layers, sparsity pattern, and compensation policy for the deployment instance.

In Sol-Engine

Sol-Engine uses Sol-Attn for long video attention in HunyuanVideo-13B and Wan2.1-T2V-14B. LTX-2.3 uses PISA-style sparse video self-attention in selected stage-2 refinement work, combined with cache, fusion, NVFP4, and token pruning in the full optimization stack.

Implemented entries:

  • techniques/sparse_backends/sol_attn/
  • techniques/sparse_backends/sol_attn_backend.py
  • python/sglang/multimodal_gen/runtime/efficiency/transforms/sparse_attention.py
  • python/sglang/multimodal_gen/runtime/layers/attention/backends/piecewise_attn.py
  • python/sglang/multimodal_gen/runtime/layers/attention/backends/sparse_video_gen_2_attn.py
  • python/sglang/multimodal_gen/runtime/layers/attention/backends/sparse_linear_attn.py
  • python/sglang/multimodal_gen/runtime/layers/attention/backends/block_sparse_attn.py

Methods

Method Open-source status Integration Role in the design space
Sol-Attn open-source HunyuanVideo and Wan2.1 on-the-fly block thresholding with approximation correction
PISA open-source local LTX runtime piecewise sparse attention selected for the LTX-2.3 full optimization stack
SpargeAttention open-source external adapter training-free block-wise sparse attention family
Sparse VideoGen open-source external adapter spatial-temporal sparse execution with online profiling
Sparse VideoGen2 open-source local backend plus SVG adapter semantic-aware sparse-token identification and GPU layout

Practical notes

  • Sparse settings should be validated visually.
  • The value of sparse attention depends on sequence length and stage placement.
  • Report both denoise time and end-to-end time; decode and offload overhead can hide attention savings.