Skip to content

PISA

PISA is the sparse attention method selected for LTX-2.3. It uses piecewise sparse attention: important blocks are computed exactly, while less important blocks can be approximated or skipped depending on configuration.

Sol-Engine placement

Sol-Engine uses PISA in selected LTX-2.3 stage-2 video self-attention blocks together with cache, token pruning, NVFP4, and kernel fusion.

Tunable knobs

  • sparsity.
  • block size.
  • route mode.
  • dense fallback layers.
  • approximate-remainder policy.
  • stage and layer placement.

Validation

Validate temporal coherence and fine detail. Sparse attention errors can be subtle in single frames but visible over motion.

Wan / LingBot usage

  • LingBot refiner. LingBot applies PISA to the 1080p refiner (density 0.10), where the spatiotemporal sequence is longest.
  • Single-GPU Wan-14B. PISA runs on single-GPU Wan-14B through the dispatch_attention_fn entry (density 0.10), firing on real self-attention of hooked layers, with attention routed via DIFFUSERS_ATTN_BACKEND=_native_cudnn. The multi-GPU context-parallel path drives PISA through its _native_attention_forward_op hook.

References