LongVideoSparseAttention

Install LVSA and generate your first long video. Use when setting up LVSA from scratch, picking SDPA vs FlashInfer backend, configuring LVSA_REFERENCE_LATENT_FRAMES for a model, or verifying the sparse path engaged via [LVSA] log lines.

Diagnose LVSA failure modes. Use when LVSA shows no speedup vs Dense, fails to engage (silent fallback), OOMs at long sequences, the output mp4 is missing after a Docker run, quality regresses at training reference, or LVSA_SCHEDULE_* env vars don't seem to do anything.

Configure and run the LVSA vllm-omni serving plugin. Use when enabling LVSA in vllm-omni, choosing per-model env vars (LVSA_WAN_HOOK / LVSA_REFERENCE_LATENT_FRAMES), debugging silent fallbacks via [LVSA-FALLBACK] warnings, setting geometry overrides for non-default resolutions, or composing with Ulysses CP.

Add LVSA support for a new video diffusion model. Use when implementing the ModelAdapter ABC for a new DiT (single-stream like Wan, dual-stream like HunyuanVideo, or joint-attention like CogVideoX), wiring it into examples/<model>_generate.py, or adding a vllm-omni hook for it.

Reproduce LVSA paper headline numbers using the bundled benchmarks/ scripts. Use when running the SotA comparison (5 prompts × 3 horizons × 4 methods), the latency-scaling sweep, scoring videos with VQeval + VBench-Long, or regenerating the figures embedded in the README.

Tune LVSA for quality vs speed. Use when adjusting sparsity_scale, choosing window_size and n_first_frames, deciding when --rotate-keyframes pays off, composing LVSA with RIFLEx, or hitting a quality regression and needing to back off sparsity.