Skip to content

SpargeAttention

SpargeAttention is a training-free sparse attention family cited by the paper. It predicts near-zero attention entries with a block-wise similarity metric and skips low-value work.

Role in Sol-Engine

SpargeAttention is part of the sparse-attention design space. It is a candidate when the attention map has clear block-level structure and the deployment can tolerate approximate attention entries.

Search knobs

  • block size.
  • similarity threshold.
  • dense fallback layers.
  • per-stage sparsity budget.

Validation

Compare against dense attention at the target resolution and duration; sparsity that works at shorter context can fail at longer temporal length.

References