Compiling TensorSharp with Visual Studio 2026

[vvdb-architecture]

I'm trying to compile the latest version of TensorSharp (forking the repository, loading the solution TensorSharp.slnx and building it).

I already found a first issue for a PR: the Magick.NET-Q8-AnyCPU must be updated to at least version 14.16.0 since the existing one has a security vulnerability. I know it's silly and trivial, but a guy has to start somewhere.

Alas, not much success so far.
Two things I struggle with:

Thing 1

I have all the prerequisites: curl is part of my system. It turns out that cmake and git are also installed as part of Visual Studio 2026, and you just need the paths:

• C:\Program Files\Microsoft Visual Studio\18\Community\Common7\IDE\CommonExtensions\Microsoft\TeamFoundation\Team Explorer\Git\mingw64\bin for git version 2.54.0.2
• C:\Program Files\Microsoft Visual Studio\18\Community\Common7\IDE\CommonExtensions\Microsoft\CMake\CMake\bin for cmake version 4.3.1.0.

I've also installed the NVidia Toolkit 12.6, because version 13.3 wasn't working correctly. With the NVidia Toolkit 13.3, I had:

No CMAKE_CUDA_COMPILER could be found.
Tell CMake where to find the compiler by setting either the environment
variable "CUDACXX" or the CMake cache entry CMAKE_CUDA_COMPILER to the full
path to the compiler, or to the compiler name if it is in the PATH.

C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.6/bin/nvcc.exe
...

Try as I might, adding the CUDACXX environment variable to point to C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\bin\nvcc.exe was not being picked up. So I just uninstalled 13.3 and installed 12.6, even though by doing that I lose Visual Studio 2026 integration.

But now I get the following:

1>------ Build started: Project: TensorSharp.Backends.GGML, Configuration: Debug Any CPU ------
1>  Building GGML native library (incremental); CMake recompiles only changed sources.
1>  vulkan-toolchain: already populated at D:\source\repos\TensorSharp\ExternalProjects\vulkan-toolchain
1>  Importing MSVC environment from C:\Program Files\Microsoft Visual Studio\18\Community\VC\Auxiliary\Build\vcvars64.bat
1>  Configuring TensorSharp.GGML.Native (CUDA=ON, CUDA_ARCHITECTURES=86-real, VULKAN=ON, TESTS=OFF, GENERATOR=Ninja, PARALLEL=32)
1>  ggml: updating existing checkout to master (https://github.com/ggml-org/ggml.git)
1>  From https://github.com/ggml-org/ggml
1>   * branch            master     -> FETCH_HEAD
1>  HEAD is now at 90951f9 ggml : bump version to 0.18.1 (#1578)
1>  ggml: now at 90951f9
1>  CMake Error at C:/Program Files/Microsoft Visual Studio/18/Community/Common7/IDE/CommonExtensions/Microsoft/CMake/CMake/share/cmake-4.3/Modules/CMakeDetermineCompilerId.cmake:934 (message):
1>    Compiling the CUDA compiler identification source file
1>    "CMakeCUDACompilerId.cu" failed.
1>
1>    Compiler: C:/Program Files/NVIDIA GPU Computing
1>    Toolkit/CUDA/v12.6/bin/nvcc.exe
1>
1>    Build flags:
1>
1>    Id flags: --keep;--keep-dir;tmp -v
1>
1>
1>
1>    The output was:
1>
1>    (this part has been abbreviated, it's just an endless list of environment variablesd)
1>
1>
1>
1>
1>
1>    #$ INCLUDES="-IC:/Program Files/NVIDIA GPU Computing
1>    Toolkit/CUDA/v12.6/bin/../include"
1>
1>    #$ LIBRARIES= "/LIBPATH:C:/Program Files/NVIDIA GPU Computing
1>    Toolkit/CUDA/v12.6/bin/../lib/x64"
1>
1>    #$ CUDAFE_FLAGS=
1>
1>    #$ PTXAS_FLAGS=
1>
1>    #$ erase tmp/a_dlink.reg.c
1>
1>    CMakeCUDACompilerId.cu
1>
1>    #$ resource file tmp/CMakeCUDACompilerId.cpp4.ii.res:
1>    [-D__CUDA_ARCH_LIST__=520 -D__NV_LEGACY_LAUNCH -nologo -E -TP -EHsc
1>    -D__CUDACC__ -D__NVCC__ "-IC:/Program Files/NVIDIA GPU Computing
1>    Toolkit/CUDA/v12.6/bin/../include" -D__CUDACC_VER_MAJOR__=12
1>    -D__CUDACC_VER_MINOR__=6 -D__CUDACC_VER_BUILD__=20
1>    -D__CUDA_API_VER_MAJOR__=12 -D__CUDA_API_VER_MINOR__=6
1>    -D__NVCC_DIAG_PRAGMA_SUPPORT__=1 -FI "cuda_runtime.h"
1>    "CMakeCUDACompilerId.cu" ]
1>
1>    #$ cl.exe @"tmp/CMakeCUDACompilerId.cpp4.ii.res" >
1>    "tmp/CMakeCUDACompilerId.cpp4.ii"
1>
1>    CMakeCUDACompilerId.cu
1>
1>    C:\Program Files\NVIDIA GPU Computing
1>Toolkit\CUDA\v12.6\include\crt/host_config.h(170): fatal error C1189:
1>    #error: -- unsupported Microsoft Visual Studio version! Only the versions
1>    between 2017 and 2022 (inclusive) are supported! The nvcc flag
1>    '-allow-unsupported-compiler' can be used to override this version check;
1>    however, using an unsupported host compiler may cause compilation failure
1>    or incorrect run time execution.  Use at your own risk.
1>
1>    # --error 0x2 --
1>
1>
1>
1>  -- Configuring incomplete, errors occurred!
1>
1>
1>  Call Stack (most recent call first):
1>    C:/Program Files/Microsoft Visual Studio/18/Community/Common7/IDE/CommonExtensions/Microsoft/CMake/CMake/share/cmake-4.3/Modules/CMakeDetermineCompilerId.cmake:8 (CMAKE_DETERMINE_COMPILER_ID_BUILD)
1>    C:/Program Files/Microsoft Visual Studio/18/Community/Common7/IDE/CommonExtensions/Microsoft/CMake/CMake/share/cmake-4.3/Modules/CMakeDetermineCompilerId.cmake:53 (__determine_compiler_id_test)
1>    C:/Program Files/Microsoft Visual Studio/18/Community/Common7/IDE/CommonExtensions/Microsoft/CMake/CMake/share/cmake-4.3/Modules/CMakeDetermineCUDACompiler.cmake:163 (CMAKE_DETERMINE_COMPILER_ID)
1>    CMakeLists.txt:94 (enable_language)
1>
1>
1>  cmake configure failed with exit code 1
1>  At D:\source\repos\TensorSharp\TensorSharp.GGML.Native\build-windows.ps1:429 char:28
1>  + ... DE -ne 0) { throw "cmake configure failed with exit code $LASTEXITCOD ...
1>  +                 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1>      + CategoryInfo          : OperationStopped: (cmake configure failed with exit code 1:String) [], RuntimeException
1>      + FullyQualifiedErrorId : cmake configure failed with exit code 1
1>
1>d:\source\repos\TensorSharp\TensorSharp.Backends.GGML\TensorSharp.Backends.GGML.csproj(103,5): error MSB3073: The command "powershell -NoProfile -ExecutionPolicy Bypass -File "d:\source\repos\TensorSharp\TensorSharp.Backends.GGML/../TensorSharp.GGML.Native/build-windows.ps1" " exited with code 1.

It seems that I need to add the -allow-unsupported-compiler flag if I want to keep using the NVidia toolkit 12.6.
I switched back to the 13.3 version (since it has Visual Studio 2026 integration and it seems to be a requirement), but I don't know how to make this version work without changing the provided TensorSharp files.
I'll try to shamelessly copy the C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v13.3/bin/ to the C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.6/bin/ directory and see what happens when I get a chance.

Thing 2
On a slightly related matter, I notice that compilation creates files in several directories which become part of the changes because they are not explicitly excluded:

• the .vs directory
• the TensorSharp.Backends.Cuda.native\ptx directory

I also notice that on Visual Studio 2026, there is a launchSettings.json file created in the TensorSharp.Server\Properties directory.

I don't want these files to be committed to the repository, but in order to do that, I must explicitly add them to .gitignore, but doing so also changes the content of the existing .gitignore.

[zhongkaifu]

Thanks for the feedback. I will investigate it.

[vvdb-architecture]

I feeel like an idiot but I still can't compile TensorSharp's GGML files, even with -no-cuda. I get this (abbreviated to show only the compilation errors):

PS D:\Source\repos\TensorSharp\TensorSharp.GGML.Native> .\build-windows.ps1 --no-cuda
vulkan-toolchain: already populated at D:\Source\repos\TensorSharp\ExternalProjects\vulkan-toolchain
MSVC environment already active (VCToolsInstallDir=C:\Program Files\Microsoft Visual Studio\18\Community\VC\Tools\MSVC\14.51.36231\); not importing vcvars.
Configuring TensorSharp.GGML.Native (CUDA=OFF, CUDA_ARCHITECTURES=n/a, VULKAN=ON, TESTS=OFF, GENERATOR=Ninja, PARALLEL=32)
ggml: updating existing checkout to master (https://github.com/ggml-org/ggml.git)
remote: Total 0 (delta 0), reused 0 (delta 0), pack-reused 0 (from 0)
From https://github.com/ggml-org/ggml
 * branch            master     -> FETCH_HEAD
HEAD is now at d99724f readme : update following llama.cpp style (#1599)
ggml: now at d99724f
-- CMAKE_SYSTEM_PROCESSOR: AMD64
-- CMAKE_GENERATOR_PLATFORM:
-- GGML_SYSTEM_ARCH: x86
-- Including CPU backend
-- x86 detected
-- Adding CPU backend variant ggml-cpu: /arch:AVX GGML_AVX
-- Vulkan found
-- GL_KHR_cooperative_matrix supported by glslc
-- GL_NV_cooperative_matrix2 supported by glslc
-- GL_NV_cooperative_matrix_decode_vector supported by glslc
-- GL_EXT_integer_dot_product supported by glslc
-- GL_EXT_bfloat16 supported by glslc
-- GL_EXT_float_e2m1 supported by glslc
-- GL_EXT_float_e4m3 supported by glslc
-- Including Vulkan backend
-- ggml version: 0.21.0
-- ggml commit:  d99724f
-- Configuring done (0.7s)
-- Generating done (0.1s)
-- Build files have been written to: D:/Source/repos/TensorSharp/TensorSharp.GGML.Native/build-windows
[0/2] Re-checking globbed directories...
[1/8] Performing build step for 'vulkan-shaders-gen'
ninja: no work to do.
[2/8] Performing install step for 'vulkan-shaders-gen'
-- Up-to-date: D:/Source/repos/TensorSharp/TensorSharp.GGML.Native/build-windows/Release/./vulkan-shaders-gen.exe
[4/8] Building CXX object g\src\ggml-vulkan\CMakeFiles\ggml-vulkan.dir\ggml-vulkan.cpp.obj
FAILED: [code=2] g/src/ggml-vulkan/CMakeFiles/ggml-vulkan.dir/ggml-vulkan.cpp.obj
C:\PROGRA~1\MICROS~1\18\COMMUN~1\VC\Tools\MSVC\1451~1.362\bin\Hostx86\x86\cl.exe  /nologo /TP -DGGML_SCHED_MAX_COPIES=4 -DGGML_VULKAN_BFLOAT16_GLSLC_SUPPORT -DGGML_VULKAN_COOPMAT2_DECODE_VECTOR_GLSLC_SUPPORT -DGGML_VULKAN_COOPMAT2_GLSLC_SUPPORT -DGGML_VULKAN_COOPMAT_GLSLC_SUPPORT -DGGML_VULKAN_FLOAT_E2M1_GLSLC_SUPPORT -DGGML_VULKAN_FLOAT_E4M3_GLSLC_SUPPORT -DGGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT -D_CRT_SECURE_NO_WARNINGS -D_XOPEN_SOURCE=600 -ID:\Source\repos\TensorSharp\ExternalProjects\ggml\src\ggml-vulkan\.. -ID:\Source\repos\TensorSharp\TensorSharp.GGML.Native\build-windows\g\src\ggml-vulkan -ID:\Source\repos\TensorSharp\ExternalProjects\ggml\src\..\include -external:ID:\Source\repos\TensorSharp\ExternalProjects\vulkan-toolchain\Vulkan-Headers\include -external:ID:\Source\repos\TensorSharp\ExternalProjects\vulkan-toolchain\spirv-headers-install\include -external:W0 /DWIN32 /D_WINDOWS /EHsc /O2 /Ob2 /DNDEBUG -std:c++17 -MD   /MP /bigobj /showIncludes /Fog\src\ggml-vulkan\CMakeFiles\ggml-vulkan.dir\ggml-vulkan.cpp.obj /Fdg\src\ggml-vulkan\CMakeFiles\ggml-vulkan.dir\ggml-vulkan.pdb /FS -c D:\Source\repos\TensorSharp\ExternalProjects\ggml\src\ggml-vulkan\ggml-vulkan.cpp
D:\Source\repos\TensorSharp\ExternalProjects\ggml\src\ggml-vulkan\ggml-vulkan.cpp(2529): error C2678: binary '<<': no operator found which takes a left-hand operand of type 'std::basic_ostream<char,std::char_traits<char>>' (or there is no acceptable conversion)
C:\Program Files\Microsoft Visual Studio\18\Community\VC\Tools\MSVC\14.51.36231\include\__msvc_ostream.hpp(476): note: could be 'std::basic_ostream<char,std::char_traits<char>> &std::basic_ostream<char,std::char_traits<char>>::operator <<(std::basic_streambuf<char,std::char_traits<char>> *)'
D:\Source\repos\TensorSharp\ExternalProjects\ggml\src\ggml-vulkan\ggml-vulkan.cpp(2529): note: 'std::basic_ostream<char,std::char_traits<char>> &std::basic_ostream<char,std::char_traits<char>>::operator <<(std::basic_streambuf<char,std::char_traits<char>> *)': cannot convert argument 2 from 'vk::Buffer' to 'std::basic_streambuf<char,std::char_traits<char>> *'

... lots of "notes'..

D:\Source\repos\TensorSharp\ExternalProjects\ggml\src\ggml-vulkan\ggml-vulkan.cpp(2545): error C2678: binary '<<': no operator found which takes a left-hand operand of type 'std::basic_ostream<char,std::char_traits<char>>' (or there is no acceptable conversion)

...

D:\Source\repos\TensorSharp\ExternalProjects\ggml\src\ggml-vulkan\ggml-vulkan.cpp(2548): error C2678: binary '<<': no operator found which takes a left-hand operand of type 'std::basic_ostream<char,std::char_traits<char>>' (or there is no acceptable conversion)

...
D:\Source\repos\TensorSharp\ExternalProjects\ggml\src\ggml-vulkan\ggml-vulkan.cpp(8456): error C2672: 'vk::CommandBuffer::copyBuffer': no matching overloaded function found
...
D:\Source\repos\TensorSharp\ExternalProjects\ggml\src\ggml-vulkan\ggml-vulkan.cpp(2529): error C2678: binary '<<': no operator found which takes a left-hand operand of type 'std::basic_ostream<char,std::char_traits<char>>' (or there is no acceptable conversion)




ninja: build stopped: subcommand failed.
cmake build failed with exit code 2
At D:\Source\repos\TensorSharp\TensorSharp.GGML.Native\build-windows.ps1:493 char:28
+ ... DE -ne 0) { throw "cmake build failed with exit code $LASTEXITCODE" } ...
+                 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    + CategoryInfo          : OperationStopped: (cmake build failed with exit code 2:String) [], RuntimeException
    + FullyQualifiedErrorId : cmake build failed with exit code 2

PS D:\Source\repos\TensorSharp\TensorSharp.GGML.Native>

The cuda builds fail as well, but I figure why bother if I can't get the -no-cuda stuff to compile anyway.

Any help would be appreciated.

[zhongkaifu]

No worries. Let me take a look and get back to you.

[zhongkaifu]

@vvdb-architecture I made a fix #170 for it and hopefully it willbe helpful.

[vvdb-architecture]

I pulled your latest version and tried it: it seems to compile OK.
I'm so happy I created a pull request to celebrate: #172