YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

HunYuan-Image-3-Comfy-W4A8

Hunyuan Image 3(混元生图 3.0)的 ComfyUI W4A8 量化模型(4-bit 权重 / 8-bit 激活)。

文件结构

diffusion_models/
  hunyuan_image_3_instruct_moe_w4a8.safetensors   # W4A8 量化扩散模型
vae/
  hunyuan_image_3_vae_fp32.safetensors            # FP32 VAE
convert_hyi3_experts.py                           # 专家层量化转换脚本
evaluate_hyi3_expert_quant.py                     # 专家层量化精度评估脚本
hyi3_w4a8_checkpoint_error.json                   # 完整逐层量化误差报告

使用方法

需要搭配插件 eastmoe/Comfyui-Hunyuan-Image3 使用。

  • diffusion_models/hunyuan_image_3_instruct_moe_w4a8.safetensors 放入 ComfyUI/models/diffusion_models/
  • vae/hunyuan_image_3_vae_fp32.safetensors 放入 ComfyUI/models/vae/

量化精度

相对 RMSE(按专家/层统计的百分位分布):

指标 p50 相对 RMSE p95 p99 最大值
Gate/Up 权重 7.314% 7.316% 7.316% 7.318%
Down 权重 7.312% 7.315% 7.315% 7.317%
完整专家前向 12.547% 13.028% 13.192% 13.370%

最差完整专家的余弦相似度为 0.99102

详细逐层数据见 hyi3_w4a8_checkpoint_error.json,可通过 evaluate_hyi3_expert_quant.py 复现。

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support