Win11+RTX3090 亲测 · ComfyUI Hunyuan3D 全程实录 ③：diso 源码编译实战（CUDA 13.1 零降级）
环境：Windows 11 23H2 | Intel Ultra 9 285K | RTX 3090 24 GB | CUDA 13.1 | Python 3.12.11 | PyTorch 2.7.1+cu126 | VS2022 17.12
系列：全程实录第 ③ 篇（第 ② 篇见nvdiffrast 编译实战）
一、前言：diso 作用与编译必要性
Differentiable Iso-Surface Extraction Package (DISO)
diso（Differentiable Iso-Surface Extraction）为 Hunyuan3DWrapper 提供可微等值面提取能力，
PyPI无 Windows wheel，必须本地编译_C.pyd才能启用Quad Remesh / Fast Decimation节点。
本文继续零降级 CUDA 驱动，不改 PyTorch 版本，一次生成可用 wheel。
二、环境 checklist（与第 ② 篇一致）
| 项目 | 本机示例 | 最低要求 |
|---|---|---|
| OS | Windows 11 23H2 | Win10 21H2+ |
| GPU | RTX 3090 24 GB | Compute Capability ≥ 8.6 |
| 驱动 | 595.02 / CUDA 13.1 | ≥ 12.6 |
| Python | 3.12.11 64-bit | 3.10-3.12 |
| PyTorch | 2.7.1+cu126 | 2.5.0+cu118+ |
| VS Build Tools | 17.12 / MSVC 14.44 | 2019/2022 任意 |
⚠️终端要求：开始菜单 →“x64 Native Tools Command Prompt for VS 2022”→ 右键以管理员身份运行
三、直接安装 vs 本地编译 对比
| 方式 | 命令 | 结果 | 本机日志片段 |
|---|---|---|---|
| 直接 pip | pip install diso | ❌ CUDA 版本检查失败 | RuntimeError: CUDA version (13.1) mismatches PyTorch (12.6) |
| 本地编译 | python setup_patch.py bdist_wheel | ✅ 生成可用 wheel | creating 'dist\diso-0.1.4-cp312-cp312-win_amd64.whl' |
下文全程基于第二种。
四、一键脚本（失败→成功全流程）
保存为build_diso.bat，在“x64 Native Tools Command Prompt for VS 2022”中运行：
@echo off title diso-Windows-Build cd /d "%~dp0" call .venv\Scripts\activate echo [1/4] 克隆源码... git clone https://github.com/SarahWeiii/diso.git cd diso echo [2/4] 应用 MonkeyPatch... copy setup.py setup.py.bak python -c "import torch.utils.cpp_extension as _ext;_ext._check_cuda_version=lambda *a,**k:None" setup.py bdist_wheel echo [3/4] 安装 wheel... pip install dist\diso-0.1.4-cp312-cp312-win_amd64.whl echo [4/4] 验证... python -c "import diso;print('✅',diso.__file__)" pause五、setup.py 修改细节（核心）
在 文件最顶部插入两行即可绕过版本检查：
# 跳过 CUDA 驱动版本检查（必须放最前）importtorch.utils.cpp_extensionas_ext _ext._check_cuda_version=lambda*args,**kwargs:None完整setup_patch.py（已含 RTX30 架构优化）：
importglobimportosimporttorchfromsetuptoolsimportfind_packages,setupfromtorch.utils.cpp_extensionimport(CUDA_HOME,BuildExtension,CppExtension,CUDAExtension,)# 1. 强制跳过 Torch 内部 CUDA 驱动版本检查importtorch.utils.cpp_extensionas_ext _ext._check_cuda_version=lambda*args,**kwargs:Nonedefget_extensions():"""Refer to torchvision."""main_file=[os.path.join("src","pybind.cpp")]source_cuda=glob.glob(os.path.join("src","*.cu"))sources=main_file extension=CppExtension define_macros=[]extra_compile_args={}if(torch.cuda.is_available()and(CUDA_HOMEisnotNone))oros.getenv("FORCE_CUDA","0")=="1":extension=CUDAExtension sources+=source_cuda define_macros+=[("WITH_CUDA",None)]nvcc_flags=os.getenv("NVCC_FLAGS","")ifnvcc_flags=="":nvcc_flags=["-O3"]else:nvcc_flags=nvcc_flags.split(" ")extra_compile_args={"cxx":["-O3"],"nvcc":nvcc_flags,}sources=[sforsinsources]include_dirs=["src"]print("sources:",sources)ext_modules=[extension("diso._C",sources,include_dirs=include_dirs,define_macros=define_macros,extra_compile_args=extra_compile_args,)]returnext_modules setup(name="diso",version="0.1.4",author_email="xiwei@ucsd.edu",keywords="differentiable iso-surface extraction",description="Differentiable Iso-Surface Extraction Package",classifiers=["Operating System :: POSIX :: Linux","Operating System :: Microsoft :: Windows","Intended Audience :: Developers","Intended Audience :: Education","Intended Audience :: Other Audience","Intended Audience :: Science/Research","Natural Language :: English","Framework :: Robot Framework :: Tool","Programming Language :: Python :: 3.6","Programming Language :: Python :: 3.7","Programming Language :: Python :: 3.8","Programming Language :: Python :: 3.9","Programming Language :: Python :: 3.10","Programming Language :: Python :: 3.11","Topic :: Software Development :: Libraries :: Python Modules","Topic :: Utilities",],license="CC BY-NC 4.0",packages=find_packages(exclude=["tests"]),python_requires=">=3.6",install_requires=["trimesh"],ext_modules=get_extensions(),cmdclass={"build_ext":BuildExtension.with_options(no_python_abi_suffix=True),},zip_safe=False)六、编译成功现场（日志片段）
[3/3] Linking build\lib.win-amd64-cpython-312\diso\_C.cp312-win_amd64.pyd creating 'dist\diso-0.1.4-cp312-cp312-win_amd64.whl' Successfully installed diso-0.1.4七、安装与验证
pipinstalldist\diso-0.1.4-cp312-cp312-win_amd64.whl python -c"import diso; print('✅', diso.__file__)"输出示例：
✅ H:\YourComfyUI\.venv\Lib\site-packages\diso\__init__.py八、常见报错对照表（收藏级）
| 报错关键词 | 原因 | 一键修复 |
|---|---|---|
CUDA version (13.1) mismatches PyTorch (12.6) | 驱动 vs 编译版本检查 | 本文 MonkeyPatch |
cl.exe not found | 未用 VS2022 x64 终端 | 开始菜单 → x64 Native Tools |
MSVC/cl.exe with traditional preprocessor is used | 仅警告，可忽略 | 已加/WX-不视为错误 |
九、一键脚本（失败→成功全流程）
保存为build_diso.bat，在“x64 Native Tools Command Prompt for VS 2022”中运行：
@echo off title diso-Windows-Build cd /d "%~dp0" call .venv\Scripts\activate echo [1/4] 克隆源码... git clone https://github.com/SarahWeiii/diso.git cd diso echo [2/4] 应用 MonkeyPatch... copy setup.py setup.py.bak python -c "import torch.utils.cpp_extension as _ext;_ext._check_cuda_version=lambda *a,**k:None" setup.py bdist_wheel echo [3/4] 安装 wheel... pip install dist\diso-0.1.4-cp312-cp312-win_amd64.whl echo [4/4] 验证... python -c "import diso;print('✅',diso.__file__)" pause十、系列交叉引用
- 第 0 篇（已发）：自定义光栅化器编译全记录
- 第 ① 篇（已发）：依赖安装完全指南
- 第 ② 篇（已发）：nvdiffrast 编译实战
转载注明出处 → 博客标题 + 链接即可。
ComfyUI, Hunyuan3D, diso, CUDA13.1, PyTorch12.6, 源码编译, setup.py, MonkeyPatch, RTX3090, Windows11
