Model Details

This model is an MXFP4 (mixed MXFP4 experts / MXFP8 shared, attention kept in BF16) model of tencent/HunyuanImage-3.0-Instruct-Distil generated by intel/auto-round using RTN. Please follow the license of the original model.

Quantized (MXFP4): routed experts, shared MLP (MXFP8)
Kept in BF16: attention, router, vision tower, embeddings, norms

How to Use

Environment

pip install vllm==0.29.0
uv pip install git+https://github.com/vllm-project/vllm-omni.git@1c7476ec19899f1e61838e23e9fad3505403b9b9
uv pip install flashinfer-python==0.6.18

vLLM Usage

vllm serve intel-ai/HunyuanImage-3.0-Instruct-Distil-MXFP4-BF16Attn-AutoRound \
    --omni

Evaluation

DiT-Only Image Fidelity (TP2)

Candidate	Topology	PSNR (dB)	SSIM@1	SSIM@1/8
BF16 (Reference)	2 GPUs TP2	Baseline	1.0000	1.0000
MXFP8 (all 8-bit, INC)	2 GPUs TP2	32.77	0.9760	0.9813
MXFP4-ar (attention MXFP8)	2 GPUs TP2	19.51	0.8421	0.6627
This build (attention BF16)	2 GPUs TP2	19.60	0.8494	0.6699

Visual Comparisons

DiT-Only Pipeline Comparison (Figure A)

BF16 (DiT TP2)	MXFP8 (DiT TP2)	MXFP4-ar (DiT TP2)	This build (DiT TP2)

AR+DiT Pipeline Branches (Figure B)

Build	AR output (conditioning text)	Tokens	Conditioning aspect
MXFP8	`It's a<boi><img_size_1024><img_ratio_14>`	6	896×1152
MXFP4-ar	`It's a warm and healing style<boi><img_size_1024><img_ratio_14>`	10	896×1152
This build	`sitting upright on a light-colored wooden floor.<boi><img_size_1024><img_ratio_9>`	12	576×1472

Weight-Space Relative L2 Deviation vs BF16 Base

Module	Scheme	rel. L2 vs BF16	MXFP4-ar (attn@MXFP8)
self_attn.qkv_proj (L0) / self_attn.o_proj (L31)	BF16	0.0000	0.0267
mlp.shared_mlp.gate_and_up_proj (L0) / down_proj (L31)	MXFP8	0.0267	0.0267
mlp.experts.* (sampled: L0/L15/L31)	MXFP4	0.1113 – 0.1125	0.1113 – 0.1125
mlp.gate.wg router	BF16	0.0000	0.0000

Generate the Model

auto-round \
    --model_name tencent/HunyuanImage-3.0-Instruct-Distil \
    --model_free \
    --scheme MXFP8 \
    --layer_config "{.mlp.experts.:{bits:4,data_type:mx_fp}}" \
    --ignore_layers "vision,guidance_emb,timestep_emb,timestep_r_emb,final_layer,wte,self_attn.qkv_proj,self_attn.o_proj" \
    --format auto_round \
    --output_dir HunyuanImage-3.0-Instruct-Distil-MXFP4-BF16Attn-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: Lv, Kaokao kaokao.lv@intel.com