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-T2V-1.3B-5step	✅ Preview checkpoint on Hugging Face
Additional checkpoints	⏳ More checkpoints coming soon
Training code	⏳ Coming soon
Paper / arXiv	⏳ Coming soon

Note

The Hugging Face repository currently contains the only public checkpoint: a preview 1.3B model trained for five denoising steps. The 4-step and larger checkpoints used in parts of the system study are not public yet.

Quick start: 1.3B 5-step Preview

The example below uses the released 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 Preview checkpoint

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-T2V-1.3B-5step-Preview \
  --local-dir ckpts/WaveForcing-T2V-1.3B-5step-Preview

The Preview repository 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-T2V-1.3B-5step-Preview/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 preview \
  --run-tag corgi

The generated video and metrics are written to:

outputs/preview/corgi/video.mp4
outputs/preview/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 1596 output frames (approximately 100 seconds at 16 FPS).

Current results

The following numbers are measured on 8× NVIDIA H200 GPUs with 399 latent frames / 1596 output frames. Steady-state numbers pool per_tick_ms[12:-4] across three repetitions and report p50. DiT FPS includes pipeline fill/drain; E2E FPS additionally includes the streaming VAE.

Model	Topology	Runtime configuration	Steady FPS (p50)	DiT FPS	E2E FPS
1.3B	4+3	Sage · causal · paged	125.7	124.1	117.7
1.3B	4+3	Sage + W8A8 · causal · paged	126.5	121.5	115.9
1.3B Preview	5+2	BF16 / SDPA · causal · paged	86.5	84.9	84.2
14B system prototype¹	4+3	Sage + W8A8 · causal · paged	28.5	28.1	28.1

The best 1.3B 4+3 run generates 1596 frames in approximately 13.56 seconds, or 117.7 E2E FPS, versus the matched 15.0 FPS single-GPU baseline (7.9× speedup). The maximum valid steady-state throughput is 126.5 FPS.

The released five-step Preview uses the 5+2 topology. Its two-stage VAE is too coarse to hide all decoding work: the best reliable BF16 configuration reaches 84.2 E2E FPS, even though faster DiT configurations exceed 120 FPS. The full three-stage VAE in 4+3 follows the DiT cadence more closely and avoids this decoder tail.

¹ The 14B row uses shape-accurate random-initialized weights and measures system scaling only; it is not a generation-quality result.

For every configuration, precise metric definitions, and the synchronized vs. overlapped comparison, 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

The arXiv entry is not public yet. For now, please cite the project page:

@misc{denglab2026waveforcing,
  title        = {Wave Forcing: Towards Speed-of-Light Streaming Video Generation},
  author       = {{DENG Lab MLSys Team}},
  year         = {2026},
  howpublished = {Project blog},
  organization = {Shanghai Jiao Tong University},
  url          = {https://sjtu-deng-lab.github.io/WaveForcing/}
}

License

Released under the Apache License 2.0.