Windows build fails with nvcc when compiling fused-ssim CUDA extension (PyTorch 2.10 + CUDA 12.9)

[Cedricyu]

I'm trying to build the CUDA extension in this repository using:
pip install --no-build-isolation --no-cache-dir .

Environment

• OS: Windows 10

• Python: 3.12

• PyTorch: 2.10.0+cu128

• CUDA Toolkit: 12.9

• GPU: (e.g. RTX 5060)

PS C:\Users\USER\Desktop\fused-ssim> pip install --no-build-isolation --no-cache-dir .
Defaulting to user installation because normal site-packages is not writeable
Processing ..
Preparing metadata (pyproject.toml) ... done
Requirement already satisfied: torch>=2.0.0 in C:\Users\USER\AppData\Roaming\Python\Python312\site-packages (from fused-ssim==1.0.0) (2.10.0+cu128)
Requirement already satisfied: filelock in C:\Users\USER\AppData\Roaming\Python\Python312\site-packages (from torch>=2.0.0->fused-ssim==1.0.0) (3.24.3)
Requirement already satisfied: typing-extensions>=4.10.0 in C:\Users\USER\AppData\Roaming\Python\Python312\site-packages (from torch>=2.0.0->fused-ssim==1.0.0) (4.15.0)
Requirement already satisfied: sympy>=1.13.3 in C:\Users\USER\AppData\Roaming\Python\Python312\site-packages (from torch>=2.0.0->fused-ssim==1.0.0) (1.14.0)
Requirement already satisfied: networkx>=2.5.1 in C:\Users\USER\AppData\Roaming\Python\Python312\site-packages (from torch>=2.0.0->fused-ssim==1.0.0) (3.6.1)
Requirement already satisfied: jinja2 in C:\Users\USER\AppData\Roaming\Python\Python312\site-packages (from torch>=2.0.0->fused-ssim==1.0.0) (3.1.6)
Requirement already satisfied: fsspec>=0.8.5 in C:\Users\USER\AppData\Roaming\Python\Python312\site-packages (from torch>=2.0.0->fused-ssim==1.0.0) (2026.2.0)
Requirement already satisfied: setuptools in C:\Users\USER\AppData\Roaming\Python\Python312\site-packages (from torch>=2.0.0->fused-ssim==1.0.0) (80.9.0)
Requirement already satisfied: mpmath<1.4,>=1.1.0 in C:\Users\USER\AppData\Roaming\Python\Python312\site-packages (from sympy>=1.13.3->torch>=2.0.0->fused-ssim==1.0.0) (1.3.0)
Requirement already satisfied: MarkupSafe>=2.0 in C:\Users\USER\AppData\Roaming\Python\Python312\site-packages (from jinja2->torch>=2.0.0->fused-ssim==1.0.0) (3.0.3)
Building wheels for collected packages: fused-ssim
Building wheel for fused-ssim (pyproject.toml) ... error
error: subprocess-exited-with-error

× Building wheel for fused-ssim (pyproject.toml) did not run successfully.
│ exit code: 1
╰─> [111 lines of output]
Compiling for CUDA.
Detected GPU architecture: sm_120
Compiling for CUDA.
Detected GPU architecture: sm_120
running bdist_wheel
W0222 01:22:15.643000 12300 torch\utils\cpp_extension.py:659] Attempted to use ninja as the BuildExtension backend but we could not find ninja.. Falling back to using the slow distutils backend.
running build
running build_py
copying fused_ssim_init_.py -> build\lib.win-amd64-cpython-312\fused_ssim
running egg_info
writing fused_ssim.egg-info\PKG-INFO
writing dependency_links to fused_ssim.egg-info\dependency_links.txt
writing requirements to fused_ssim.egg-info\requires.txt
writing top-level names to fused_ssim.egg-info\top_level.txt
reading manifest file 'fused_ssim.egg-info\SOURCES.txt'
adding license file 'LICENSE'
writing manifest file 'fused_ssim.egg-info\SOURCES.txt'
running build_ext
W0222 01:22:15.683000 12300 torch\utils\cpp_extension.py:484] Error checking compiler version for cl: [WinError 2] 系統找不到指定的檔案。
W0222 01:22:15.692000 12300 torch\utils\cpp_extension.py:525] The detected CUDA version (12.9) has a minor version mismatch with the version that was used to compile PyTorch (12.8). Most likely this shouldn't be a problem.

  ==================================================
  Building with GPU architecture: sm_120
  ==================================================

  building 'fused_ssim_cuda' extension
  "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.41.34120\bin\HostX86\x64\cl.exe" /c /nologo /O2 /W3 /GL /DNDEBUG /MD -IC:\Users\USER\AppData\Roaming\Python\Python312\site-packages\torch\include -IC:\Users\USER\AppData\Roaming\Python\Python312\site-packages\torch\include\torch\csrc\api\include "-IC:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include" -IC:\Python312\include -IC:\Python312\Include "-IC:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.41.34120\include" "-IC:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.41.34120\ATLMFC\include" "-IC:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\VS\include" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\ucrt" "-IC:\Program Files (x86)\Windows Kits\10\\include\10.0.22621.0\\um" "-IC:\Program Files (x86)\Windows Kits\10\\include\10.0.22621.0\\shared" "-IC:\Program Files (x86)\Windows Kits\10\\include\10.0.22621.0\\winrt" "-IC:\Program Files (x86)\Windows Kits\10\\include\10.0.22621.0\\cppwinrt" /EHsc /Tpext.cpp /Fobuild\temp.win-amd64-cpython-312\Release\ext.obj /MD /wd4819 /wd4251 /wd4244 /wd4267 /wd4275 /wd4018 /wd4190 /wd4624 /wd4067 /wd4068 /EHsc -O3 -DFUSED_SSIM_CUDA -DTORCH_API_INCLUDE_EXTENSION_H -DTORCH_EXTENSION_NAME=fused_ssim_cuda /std:c++17
  cl : 命令列 warning D9002 : 忽略未知的選項 '-O3'
  ext.cpp
  "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\bin\nvcc" -c ssim.cu -o build\temp.win-amd64-cpython-312\Release\ssim.obj -IC:\Users\USER\AppData\Roaming\Python\Python312\site-packages\torch\include -IC:\Users\USER\AppData\Roaming\Python\Python312\site-packages\torch\include\torch\csrc\api\include "-IC:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include" -IC:\Python312\include -IC:\Python312\Include "-IC:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.41.34120\include" "-IC:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.41.34120\ATLMFC\include" "-IC:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\VS\include" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\ucrt" "-IC:\Program Files (x86)\Windows Kits\10\\include\10.0.22621.0\\um" "-IC:\Program Files (x86)\Windows Kits\10\\include\10.0.22621.0\\shared" "-IC:\Program Files (x86)\Windows Kits\10\\include\10.0.22621.0\\winrt" "-IC:\Program Files (x86)\Windows Kits\10\\include\10.0.22621.0\\cppwinrt" -Xcompiler /EHsc -Xcompiler /wd4068 -Xcompiler /wd4067 -Xcompiler /wd4624 -Xcompiler /wd4190 -Xcompiler /wd4018 -Xcompiler /wd4275 -Xcompiler /wd4267 -Xcompiler /wd4244 -Xcompiler /wd4251 -Xcompiler /wd4819 -Xcompiler /MD -Xcudafe --diag_suppress=dll_interface_conflict_dllexport_assumed -Xcudafe --diag_suppress=dll_interface_conflict_none_assumed -Xcudafe --diag_suppress=field_without_dll_interface -Xcudafe --diag_suppress=base_class_has_different_dll_interface -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr -O3 -DFUSED_SSIM_CUDA --maxrregcount=32 --use_fast_math -arch=sm_120 -DTORCH_API_INCLUDE_EXTENSION_H -DTORCH_EXTENSION_NAME=fused_ssim_cuda -std=c++17 --use-local-env
  C:/Program Files (x86)/Windows Kits/10//include/10.0.22621.0//um\winnt.h(24437): warning #174-D: expression has no effect
        (CallbackEnviron);
        ^

  Remark: The warnings can be suppressed with "-diag-suppress <warning-number>"

  C:/Program Files (x86)/Windows Kits/10//include/10.0.22621.0//um\winuser.h(14668): warning #108-D: signed bit field of length 1
        BOOL fBarFocused:1;
             ^

  C:/Program Files (x86)/Windows Kits/10//include/10.0.22621.0//um\winuser.h(14669): warning #108-D: signed bit field of length 1
        BOOL fFocused:1;
             ^

  C:/Program Files (x86)/Windows Kits/10//include/10.0.22621.0//um\wincrypt.h(21836): warning #1835-D: attribute "dllimport" does not apply here
    typedef __declspec(dllimport) BOOL (__stdcall *PFN_CERT_IS_WEAK_HASH)(
                       ^

  C:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\shared\rpcndr.h(730): warning #108-D: signed bit field of length 1
        int                             fInDontFree       :1;
                                        ^

  C:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\shared\rpcndr.h(731): warning #108-D: signed bit field of length 1
        int                             fDontCallFreeInst :1;
                                        ^

  C:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\shared\rpcndr.h(732): warning #108-D: signed bit field of length 1
        int                             fUnused1          :1;
                                        ^

  C:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\shared\rpcndr.h(733): warning #108-D: signed bit field of length 1
        int                             fHasReturn        :1;
                                        ^

  C:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\shared\rpcndr.h(734): warning #108-D: signed bit field of length 1
        int                             fHasExtensions    :1;
                                        ^

  C:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\shared\rpcndr.h(735): warning #108-D: signed bit field of length 1
        int                             fHasNewCorrDesc   :1;
                                        ^

  C:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\shared\rpcndr.h(736): warning #108-D: signed bit field of length 1
        int                             fIsIn             :1;
                                        ^

  C:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\shared\rpcndr.h(737): warning #108-D: signed bit field of length 1
        int                             fIsOut            :1;
                                        ^

  C:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\shared\rpcndr.h(738): warning #108-D: signed bit field of length 1
        int                             fIsOicf           :1;
                                        ^

  C:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\shared\rpcndr.h(739): warning #108-D: signed bit field of length 1
        int                             fBufferValid      :1;
                                        ^

  C:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\shared\rpcndr.h(740): warning #108-D: signed bit field of length 1
        int                             fHasMemoryValidateCallback: 1;
                                        ^

  C:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\shared\rpcndr.h(741): warning #108-D: signed bit field of length 1
        int                             fInFree             :1;
                                        ^

  C:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\shared\rpcndr.h(742): warning #108-D: signed bit field of length 1
        int                             fNeedMCCP         :1;
                                        ^

  C:/Users/USER/AppData/Roaming/Python/Python312/site-packages/torch/include\c10/cuda/CUDACachingAllocator.h(212): error: invalid combination of type specifiers
      StreamSegmentSize(cudaStream_t s, bool char, size_t sz)
                                             ^

  C:/Users/USER/AppData/Roaming/Python/Python312/site-packages/torch/include\c10/cuda/CUDACachingAllocator.h(213): error: type name is not allowed
          : stream(s), is_small_pool(char), total_size(sz) {}
                                     ^

  2 errors detected in the compilation of "ssim.cu".
  ssim.cu
  error: command 'C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v12.9\\bin\\nvcc.EXE' failed with exit code 2
  [end of output]

note: This error originates from a subprocess, and is likely not a problem with pip.
ERROR: Failed building wheel for fused-ssim
Failed to build fused-ssim
error: failed-wheel-build-for-install

× Failed to build installable wheels for some pyproject.toml based projects
╰─> fused-ssim

[TheNerdonian88]

I don't know if it is the fix for your particular issue, but it is impossible to compile those older Pytorch versions on Windows, due to a known bug.

The file autograd.h has a bug where IValue is packed, but the solution is simple and basically linked in the code comment:

"// NB: the if-constexpr usage triggers compilation errors on Windows
// with certain compiler settings
// (see pytorch/pytorch#144707 for examples).
// It's not clear what the problem is, so we're going to ignore it for now."

At the end of the linked pull, some one found out that 2 lines of code, concerning std have to be commented out:

compiled_autograd.h

From:

else if constexpr (::std::is_same_v<T, ::std::string>) {
return at::StringType::get();

To:

//else if constexpr (::std::is_same_v<T, ::std::string>) {
//return at::StringType::get();

This is a needed change and it produces multiple error massages and a failed build other wise.

I just got fused-ssim to compile on Windows, this was the last hurdle.
On Win11 compiled with Visual studio 2019 and cuda 12.8 and the coresponding Pytorch between 2.4 and 2.8 I think.
Installed in a env and ran the compiled with the Miniforge console, using the last setup command on this repro.

Make sure to make manual entries into the environment variables in the Windows settings, so the MSVC compiler can be found, as well for the cuda toolkit. Also only install one version of VS

Hope it helps, but that change is necessary for a compile on Windows either way.