Model Details

This model is an MXFP8 (OCP microscaling FP8, group_size 32) weight-quantized model of Wan-AI/Wan2.2-S2V-14B generated by intel/auto-round using zero-shot RTN. Please follow the license of the original model.

Quantized (MXFP8): diffusion transformer backbone
Kept in BF16: audio encoder, VAE, embeddings, norms

How to Use

Environment

uv pip install vllm==0.29.0
uv pip install git+https://github.com/vllm-project/vllm-omni.git@1c7476ec19899f1e61838e23e9fad3505403b9b9

vLLM Usage

vllm serve intel-ai/Wan2.2-S2V-14B-MXFP8-AutoRound \
    --omni

To generate speech-to-video, send a request to /v1/videos with input audio and guidance prompt.

Evaluation

Benchmark	BF16	MXFP8	Recovery
VBench Total Score	83.50	83.10	99.52%
Quality Score	82.50	82.20	99.64%
Motion Smoothness	97.60	97.30	99.69%
Dynamic Degree	74.20	73.80	99.46%
Audio-Lip Sync	95.80	95.50	99.69%
Subject Consistency	95.50	95.20	99.69%
Average	88.18	87.85	99.62%

Video Generation Output

Demo video generated by the quantized checkpoint:

Download Demo Video

Generate the Model

auto-round \
    --model Wan-AI/Wan2.2-S2V-14B \
    --scheme mxfp8 \
    --iters 0 \
    --disable_opt_rtn \
    --output_dir Wan2.2-S2V-14B-MXFP8-AutoRound

Ethical Considerations and Limitations

The model can produce factually incorrect output, and should not be relied on to produce factually accurate information. Because of the limitations of the pretrained model and the finetuning datasets, it is possible that this model could generate lewd, biased or otherwise offensive outputs. Therefore, before deploying any applications of the model, developers should perform safety testing.

Caveats and Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Here are a couple of useful links to learn more about Intel's AI software:

Disclaimer

The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please consult an attorney before using this model for commercial purposes.

Cite

@article{cheng2023optimize,
  title={Optimize weight rounding via signed gradient descent for the quantization of llms},
  author={Cheng, Wenhua and Zhang, Weiwei and Shen, Haihao and Cai, Yiyang and He, Xin and Lv, Kaokao and Liu, Yi},
  journal={arXiv preprint arXiv:2309.05516},
  year={2023}
}

arxiv github

Signed-off-by: Wang, Chang chang1.wang@intel.com