lvsa-reproduce-paper

Reproducing LVSA paper numbers

Setup

git clone https://github.com/JiusiServe/LongVideoSparseAttention
cd LVSA

uv venv --python 3.12
source .venv/bin/activate

# Install LVSA + scoring deps
uv pip install -e ".[diffusers,hunyuan,flashinfer,dev]"
uv pip install -e vqeval/

# For VBench-Long, you need a separate venv (it pins old diffusers/transformers)
git clone https://github.com/Vchitect/VBench /path/to/VBench
python3 -m venv /path/to/vbench-venv
source /path/to/vbench-venv/bin/activate
pip install -e /path/to/VBench
deactivate

source .venv/bin/activate  # back to LVSA venv

# Model weights (downloaded separately)
# Wan 2.1 1.3B: huggingface-cli download Wan-AI/Wan2.1-T2V-1.3B-Diffusers
export MODEL_PATH=/path/to/Wan2.1-T2V-1.3B-Diffusers

SotA grid (5 prompts × 3 horizons × 4 methods)

Generate 60 videos: Dense / RIFLEx / LVSA-SDPA / LVSA-FI × 165/249/333 frames × 5 prompts.

export OUTDIR=out/sota_comparison
bash benchmarks/sota_comparison.sh

Expected wall time per cell (single A100, 50 steps, seed 16):

Total: ~9 hours on single A100, ~70 min on 8×A100 via GNU parallel.

UltraViCo is excluded — it lives in a separate repo (thu-ml/DiT-Extrapolation, branch ultra-wan) with a different CLI. See the paper appendix for the UltraViCo recipe.

Latency scaling (single-prompt frame sweep)

For the headline 3.14× claim and the latency figure on the README:

export OUTDIR=out/latency_scaling
bash benchmarks/latency_scaling.sh

Frame counts swept: 81 (1×) / 161 (2×) / 321 (4×) / 481 (6×). Methods: Dense + LVSA-FI.

Score with VQeval

bash benchmarks/score_vqeval.sh out/sota_comparison
# Writes <stem>.vqeval.json next to each mp4

VQeval scores 6 dimensions + composite. Single A100 + the bundled vqeval/ subpackage. Expect ~15 min for 60 videos.

Score with VBench-Long

VBENCH_REPO=/path/to/VBench \
VBENCH_PYTHON=/path/to/vbench-venv/bin/python \
    bash benchmarks/score_vbench.sh out/sota_comparison
# Writes <stem>.vbench.json next to each mp4

VBench-Long scores 5 dimensions: subject_consistency, temporal_flickering, motion_smoothness, background_consistency, imaging_quality.

Aggregate

python benchmarks/aggregate.py --outdir out/sota_comparison
# Writes _summary.csv (60 rows) and _summary_means.csv (12 cells)

The aggregator walks the output directory, parses tags (<model>__<backend>__<horizon>__<prompt>), loads the per-video JSONs, and emits tidy + means CSVs.

Regenerate figures

python benchmarks/generate_figures.py \
    --sota-csv     out/sota_comparison/_summary_means.csv \
    --scaling-csv  out/latency_scaling/_summary_means.csv \
    --outdir       docs/figures/

Produces 4 PNGs at 300 DPI:

• latency_scaling.png — Wan 1.3B Dense vs LVSA wall-time scaling
• crossmodel_speedup.png — speedup-vs-Dense bar chart
• hv_latency_scaling.png — HunyuanVideo wall-time scaling
• sparsity_vs_frames.png — per-query attended fraction by model

Expected results (5-prompt mean)

Wall time and speedup vs Dense

Speedup vs UltraViCo (from paper appendix data)

Quality (VQeval composite, Δ vs Dense)

Quality (VBench-Long imaging_quality, Δ vs Dense)

Tips

• Idempotency: all scripts skip cells whose .mp4 (or .vqeval.json, .vbench.json) already exists. Crash-and-resume works.
• Seed: 16 matches the paper. Change SEED=<n> to get a different RNG roll.
• Resolution: 480×832 is the default and matches the paper. Higher resolutions need LVSA_PATCHES_PER_FRAME set or VIDEO_HEIGHT/WIDTH for the vllm-omni plugin.
• 8× A100 parallel: the dev repo at scripts/paper_results/sota_job_runner.sh ships a flock-queued GNU-parallel orchestrator. The pruned recipe in benchmarks/ is single-GPU sequential.

Adapting to other models

Change the example invocation in benchmarks/sota_comparison.sh from examples/wan_generate.py to examples/hunyuan_generate.py (and the HORIZONS arrays to HunyuanVideo's range: 65/129/193/257). The aggregator and figure scripts handle any model tag.