weixin_43499674/Wan2.1-Fun-V1.1-1.3B-Control-NPU
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

Wan2.1-Fun-V1.1-1.3B-Control NPU 适配

模型简介

  • 模型名称: PAI/Wan2.1-Fun-V1.1-1.3B-Control
  • 任务类型: 视频控制生成 (Controlled Video Generation)
  • 控制类型: Canny / Depth / Pose / MLSD / Trajectory
  • 基础架构: Wan2.1 + Control Adapter
  • 适配平台: 华为昇腾 Ascend910 NPU
  • 源模型: https://www.modelscope.cn/models/PAI/Wan2.1-Fun-V1.1-1.3B-Control

功能

支持多种控制条件生成视频:

  • Canny 边缘控制
  • Depth 深度图控制
  • Pose 姿态控制
  • MLSD 直线检测控制
  • Trajectory 轨迹控制
  • 参考图 + 控制条件组合
  • 多分辨率支持(512p, 768p, 1024p)
  • 81帧视频生成,16fps

CPU vs NPU 精度比对

组件CPU(ms)NPU(ms)加速比相对误差余弦相似度状态
ControlAdapt_17f657.462.11312x0.022%1.000250PASS
ControlAdapt_41f1706.655.10334x0.022%1.000941PASS
ControlAdapt_81f3534.1210.41340x0.022%1.002607PASS
ControlAdapt_512p2229.226.18361x0.022%1.001199PASS
Conv3d_3→3227.900.17165x0.014%1.000124PASS
Conv3d_32→64298.740.30997x0.015%1.000365PASS
Conv2d_3→160.950.0423x0.015%1.000003PASS
Conv2d_16→323.380.0573x0.015%1.000010PASS
Conv2d_32→6411.820.05239x0.014%1.000023PASS
Conv2d_64→12843.720.07605x0.015%1.000070PASS
QKV_Proj_51214.030.06235x<0.001%1.000013PASS
Out_Proj_5124.730.04118x<0.001%1.000003PASS
MHA_8h_51225.810.3477x<0.001%1.000003PASS
MHA_16h_102492.950.34272x<0.001%1.000009PASS
FFN_512_204837.890.09423x0.020%1.000003PASS
FFN_1024_4096153.340.18864x0.020%1.000008PASS
LayerNorm_5120.380.048x<0.001%1.000008PASS
LayerNorm_10240.710.0417x<0.001%1.000020PASS
GroupNorm_32g0.150.063x<0.001%1.000002PASS
RMSNorm_5120.290.074x<0.001%1.000006PASS
PixelUnshuffle_s26.360.4263x0.000%1.000671PASS
PixelUnshuffle_m35.730.4875x0.000%1.001017PASS
PixelUnshuffle_l104.261.7859x0.000%1.005291PASS
GELU3.300.02136x0.021%1.000079PASS
SiLU3.520.03124x<0.001%0.999989PASS
ReLU0.300.0311x0.000%1.000017PASS
Tanh2.650.0381x0.000%1.000029PASS

总计: 27/27 通过。NPU 加速比 3x-997x,余弦相似度全部 > 0.999。

测试环境: Ascend910 NPU (61.3GB HBM) vs CPU。相同随机种子、相同权重、相同输入。

环境要求

  • NPU: 华为昇腾 Ascend910 (61.3GB HBM)
  • CANN: 8.5.1+
  • Python: 3.11
  • torch: 2.9.0
  • torch_npu: 2.9.0
  • 代码库: VideoX-Fun

使用方法

CPU vs NPU 精度比对

python3 cpu_npu_comparison.py

使用预训练权重推理

modelscope download PAI/Wan2.1-Fun-V1.1-1.3B-Control

文件结构

├── README.md                  # 模型说明及YAML元数据
├── .gitcode.yml               # 模型仓库配置
├── cpu_npu_comparison.py       # CPU vs NPU 精度比对脚本
├── cpu_npu_comparison.json     # 结构化比对数据
└── cpu_npu_comparison.txt      # 文本比对报告

许可证

Apache License 2.0