数字人 LAM 部署笔记

目录

windos踩坑

---

GitHub - aigc3d/LAM: [SIGGRAPH 2025] LAM: Large Avatar Model for One-shot Animatable Gaussian Head

windos踩坑

nvidia-smi


环境  cuda 11.8

conda install -c "nvidia/label/cuda-11.8.0" cuda-toolkit
conda create --name cuda11.8 -y python=3.10

conda activate cuda11.8


set PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\bin;%PATH%


指定虚拟环境cuda11.8

set CUDA_HOME=E:\miniconda\envs\cuda11.8\Lib\site-packages\torch\lib

使用命令：

set CUDA_HOME=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8
set PATH=%CUDA_HOME%\bin;%PATH%
set LIBRARY_PATH=%CUDA_HOME%\libnvvp;%LIBRARY_PATH%
set LD_LIBRARY_PATH=%CUDA_HOME%\libnvvp;%LD_LIBRARY_PATH%


pytorch 3d   踩坑：

1   不可使用虚拟环境的精简版cuda11.8，要安装完整版cuda11.8

2    需要cmd进入E:\work_now\suanfa_diaoyan\LAM\external\landmark_detection\FaceBoxesV2\utils 目录

    执行命令  python build.py build_ext --inplace


3 报错 File "nms/cpu_nms.pyx", line 25, in nms.cpu_nms.cpu_nms ValueError: Buffer dtype mismatch, expected 'int_t' but got 'long long'

解决办法：用文本编辑器打开fast_rcnn_root/lib/utils/nms.pyx，将第25行的np.int_t修改为np.intp_t。（重点：要重新用fastrcnn的setup build）