lvsa

Install LVSA and generate your first long video. Picks SDPA vs FlashInfer, sets the right LVSA_REFERENCE_LATENT_FRAMES per model, verifies the sparse path engaged.

Tune LVSA's sparsity_scale, window_size, n_first_frames, and reference_latent_frames for a target model and quality/speed trade-off.

Diagnose LVSA failure modes: silent dense fallback, OOM at long sequences, missing output mp4 in Docker, quality regression at training reference.

Add LVSA support for a new video diffusion model by implementing the ModelAdapter ABC and wiring it into examples/.

Configure and run the LVSA vllm-omni serving plugin: env vars, geometry overrides, hooks for Wan vs HunyuanVideo, multi-GPU Ulysses.

Reproduce the headline paper numbers (SotA grid, latency scaling) using the bundled benchmarks/ scripts, VQeval, and VBench-Long.