Wave Forcing

Towards Speed-of-Light Streaming Video Generation

Wave Forcing is a block-causal streaming video diffusion method, paired with Wave Parallelism and the WaveRT runtime. It preserves a mixed-noise frontier and fresh cleaner-to-noisier context while removing the reverse dependencies that prevent efficient pipelining. Across GPUs, full-model replicas process different chunk-stage tasks and publish layer-wise K/V states in one direction. Communication is prefetched and hidden under the longer attention computation on the critical rank.

The inference/runtime code in this repository is now open source.

Release status

Asset	Status
Inference code and WaveRT	✅ Available in this repository
Interactive project page and full configuration matrix	✅ Available here
WaveForcing-1.3B, 5-step	✅ 5step/ on Hugging Face
WaveForcing-1.3B, 4-step	✅ 4step/ on Hugging Face
Additional checkpoints	⏳ More checkpoints coming soon
Reference training code	S1/S2/S3 source and recipes
Paper / arXiv	✅ arXiv:2611.00859

Note

Both public checkpoints live in one repository, SJTU-DENG-Lab/WaveForcing-1.3B: 5step/ and 4step/, each a native Wan model.safetensors (BF16, EMA weights). A mirror is on ModelScope.

Quick start: WaveForcing-1.3B

The example below uses the 5-step checkpoint with an eight-GPU 5+2 pipeline: five denoising stages, one clean-KV store rank, and two streaming VAE stages. The runtime is currently validated with Python 3.12, PyTorch 2.11.0+cu128, the SGLang diffusion runtime, and Hopper GPUs. See pyproject.toml for the pinned Python stack.

1. Download the base model and WaveForcing-1.3B

python -m pip install -U "huggingface_hub[cli]"

hf download Wan-AI/Wan2.1-T2V-1.3B-Diffusers \
  --local-dir ckpts/Wan2.1-T2V-1.3B-Diffusers

hf download SJTU-DENG-Lab/WaveForcing-1.3B \
  --local-dir ckpts/WaveForcing-1.3B

WaveForcing-1.3B supplies the distilled DiT weights. Text encoder, tokenizer, VAE, and scheduler components are loaded from the Wan2.1 base model.

2. Launch WaveRT

python -m wave_rt \
  --model-path ckpts/Wan2.1-T2V-1.3B-Diffusers \
  --gen-ckpt ckpts/WaveForcing-1.3B/5step/model.safetensors \
  --rf-step 5 \
  --wp-size 6 \
  --denoising-step-list 1000,800,600,400,200 \
  --num-frames 24 \
  --height 480 \
  --width 832 \
  --vae-stages 2 \
  --vae-partition time \
  --kv-context causal \
  --exchange-mode paged \
  --attention-backend torch_sdpa \
  --cuda-visible-devices 0,1,2,3,4,5,6,7 \
  --prompt "A cinematic shot of a fluffy corgi running on a sunny beach, waves in the background." \
  --task quickstart \
  --run-tag corgi

The generated video and metrics are written to:

outputs/quickstart/corgi/video.mp4
outputs/quickstart/corgi/metrics.json

--num-frames counts latent frames and must be divisible by three. Use --num-frames 399 to match the long-video benchmark protocol below; this produces 1593 output frames (approximately 100 seconds at 16 FPS).

3. Run the 4-step checkpoint

Point --gen-ckpt at ckpts/WaveForcing-1.3B/4step/model.safetensors and set --rf-step 4 --wp-size 5 --denoising-step-list 1000,750,500,250 --vae-stages 3. This is the 4+3 layout used for the headline numbers below: four denoising stages, one store rank, and three streaming VAE stages on the same eight GPUs. The quick start above is the 5+2 layout, which the results table also lists.

Reference training

The training guide documents the S1/S2/S3 recipe, asset preparation, launch, and resume commands. Training uses the same Python 3.12 / PyTorch 2.11.0+cu128 environment as WaveRT, under a separate academic-use license.

Current results

The following numbers are measured on 8× NVIDIA H200 GPUs at 832×480 with 399 latent frames (1593 decoded frames). Steady-state values pool per-tick timings across three repetitions and report p50; DiT FPS includes pipeline fill/drain, and E2E FPS additionally includes the streaming VAE. Loading and file encoding are excluded.

Model	Steps	Topology	Runtime configuration	Steady FPS (p50)	DiT FPS	E2E FPS
1.3B	4	4+3	BF16/FA3 · causal · paged	152.1	147.6	134.8
1.3B	4	4+3	Sage · causal · paged	153.2	147.8	134.8
1.3B	4	4+3	Sage + W8A8 · causal · paged	157.8	145.5	132.9
1.3B	5	5+2	BF16/FA3 · causal · paged	99.6	104.8	96.5
14B system prototype¹	4	4+3	Sage + W8A8 · causal · paged	33.9	33.6	33.5

The best 1.3B 4+3 run decodes 1593 frames in approximately 11.8 seconds, or 134.8 E2E FPS, versus the matched 19.06 FPS single-GPU baseline (7.07× speedup). The maximum steady-state throughput is 157.8 FPS.

¹ The 14B row uses shape-accurate random-initialized weights and measures system scaling only; it is not a generation-quality result. The 1.3B rows use the released 4-step and 5-step checkpoints.

For every configuration, precise metric definitions, and the exchange-mode ablation, see the interactive results matrix. The machine-readable summary is available at page/static/data/results_summary.json.

Runtime layout

wave_rt/
├── denoiser/              # wavefront schedule, diagnostics, and KV exchange
│   └── exchange/          # sync, overlap, one-sided, staggered, relay, paged
├── runtime/               # SGLang integration, attention backends, FP8
├── distributed/           # process-group compatibility and CUDA IPC
├── pipelines/             # streaming VAE pipeline
├── config.py              # public configuration and CLI arguments
└── launcher.py            # one-shot and resident serving launchers

Citation

Please cite our arXiv paper:

@misc{denglab2026waveforcing,
  title        = {Wave Forcing: Towards Speed-of-Light Streaming Video Generation},
  author       = {Jin, Yijie and Liu, Yuxuan and Xu, Jiajun and Huang, Ye and
                  Wang, Chuzhi and Li, Jiayang and Chen, Ruiyang and Hong, Chenling and
                  Wang, Xu and Deng, Zhijie},
  year         = {2026},
  eprint       = {2611.00859},
  archivePrefix = {arXiv},
  url          = {https://arxiv.org/abs/2611.00859}
}

License

The WaveRT code is released under the Apache License 2.0. The RollingForcing-derived reference trainer retains its separate academic-use license.