LLKLL/DecoupAttn-plans

DecoupAttn Plans

The offline plans and calibration tables of DecoupAttn: Decoupled Sparse Attention for Training-Free Acceleration of Video Diffusion. Code: github.com/YJLi-new/DecoupAttn.

DecoupAttn compiles one plan per model offline. Distortion Budget gives every unit u = (t, l, h) one option (Dense, Static_θ with θ ∈ {0.05, 0.10, 0.15, 0.20, 0.30, 0.45}, or CompOnly) and every call one estimator (Full or Centroid) at density 0.20, from the block ranking D_u, the distortion ε_u and the sensitivity J_u. This dataset holds the plans of the paper and the ε_u and J_u files they were solved from.

Contents

Key	Model	Block shape	Reassign	ε_u clips	J_u	PSNR of this plan	PSNR in Table 1
cp1	Cosmos-Predict1	2×4×16	yes	4	here	– (plan lost)	27.77
cp25	Cosmos-Predict2.5	4×4×8	yes	4	GitHub data/sensitivity/	24.70	24.70
wan21	Wan2.1-14B (480p)	4×4×8	yes	4	GitHub data/sensitivity/	24.70	24.70
wan22	Wan2.2-A14B	1×16×8	yes	4	here	18.43	18.43
hy	HunyuanVideo	1×8×16	yes	1	here	27.25	27.47
nano	Cosmos3-Nano (I2W)	4×4×8	yes	4	here	21.95	21.56
t25	Cosmos-Transfer2.5 (edge)	2×4×16	yes	4	GitHub data/sensitivity/	33.81	33.81
dr	DiffusionRenderer	2×4×16	once	4	GitHub data/sensitivity/	50.87	50.87
alaya	AlayaRenderer	4×4×8	yes	4	GitHub data/sensitivity/	28.15	28.15
unirelight	UniRelight	1×16×8	yes	4	here	42.75	42.75

PSNR (dB) is measured against the same-seed dense output on the test prompts or clips of each model, as in Table 1. Block shapes are frames × rows × cols of one key block. "Reassign" is the Reassign Cache rule (δ_stale > 1.0 dB); for DiffusionRenderer the clusters are built once.

HunyuanVideo and Cosmos3-Nano. The plan files behind Table 1 were lost. The plans here are re-runs of the same calibration with the same constants; their PSNR differs from Table 1 by −0.22 dB (HunyuanVideo, 8 test prompts) and +0.39 dB (Cosmos3-Nano, 32 test prompts).

Cosmos-Predict1. The plan file was lost. Its ε_u and J_u files are included; the block ranking D_u has to be measured again before the plan can be solved (see below).

Layout

<key>/
├── plan.npz                        the plan, read by decoupattn.plan.load_plan
├── plan.json                       summary of the run that compiled it: block shape, reassignment, option histogram,
│                                   executed density, sensitivity source, solver
├── distortion/<clip>_s1000.npz     distortion ε_u on the calibration clips (seed 1000)
└── sensitivity/J_s1000.jsonl       group-injection runs behind J_u (models whose J_u is not in the GitHub repository)

Each folder has the layout of a DecoupAttn run directory, so it can be passed as --run-dir. plan.json keeps the keys of the runs behind the paper, which differ from the summary that decoupattn calibrate writes. The block ranking tables D_u are not included (up to 53 GB per model).

Usage

pip install -U huggingface_hub
hf download LLKLL/DecoupAttn-plans --repo-type dataset --local-dir plans

Generate the test videos of a model with its plan (installation and checkpoints: see the GitHub README):

python -m decoupattn generate --model cp25 --run-dir runs/cp25 --plan plans/cp25/plan.npz --split test
python -m decoupattn generate --model cp25 --run-dir runs/cp25 --dense --split test
python -m decoupattn evaluate --model cp25 --run-dir runs/cp25 --split test

To solve a plan again, copy the model folder into a new run directory and run decoupattn calibrate there. The GPU stages skip clips whose outputs exist, so the distortion and sensitivity files here are reused and only the block shape, the reassignment probe and D_u are measured. The Cosmos-Predict1 sensitivity file names its groups r0_t0, … and its floor dense2; set prefix = "r" and floor = "dense2" in [planner.sensitivity] of configs/models/cp1.toml to read it.

Citation

@misc{li2026decoupattn,
  title  = {{DecoupAttn}: Decoupled Sparse Attention for Training-Free Acceleration of Video Diffusion},
  author = {Li, Yanjin and Qiu, Bingbing and Huang, Qiang and Qian, Rui and Wang, Jian and Zhang, Mingda},
  year   = {2026}
}

License

Apache License 2.0. The plans were compiled from the outputs of the models listed above, which keep their own licenses and terms of use; the calibration prompts of the video models are from MovieGen Video Bench (CC BY-NC 4.0), and only their line numbers are stored in the GitHub repository.