Is it possible to install libtorch without CUDA / cudnn / GPU?

[HoLue]

Copy issue body from D:/di/Libraries/Vs/vcpkg/installed/vcpkg/issue_body.md

Is it possible to install without CUDA / cudnn / GPU?

I've done:
vcpkg install libtorch:x64-windows

I got:
Computing installation plan...
The following packages will be built and installed:

• blas:x64-windows@2023-04-14#1
• cpuinfo:x64-windows@2022-07-19#3
• cuda:[email protected]#13
• cudnn:[email protected]#14
• fftw3:[email protected]#9
• fmt:[email protected]#1
• foxi:x64-windows@2021-12-01
• fp16:x64-windows@2021-02-21#2
• fxdiv:x64-windows@2021-02-21#2
• gemmlowp:x64-windows@2021-09-28
• gflags:[email protected]#9
• glog:[email protected]
• lapack:x64-windows@2023-06-10#2
• lapack-reference[blas-select,core,noblas]:[email protected]#6
  libtorch[core,cuda,opencv,zstd]:[email protected]#7
• lmdb:[email protected]
• magma:[email protected]#1
• mimalloc:[email protected]#1
• nvidia-cutlass:[email protected]
• onnx:[email protected]
• onnx-optimizer:[email protected]
• openblas:[email protected]
• opencl:[email protected]
• opencv[core,default-features]:[email protected]#1
• pocketfft:x64-windows@2023-09-25
• psimd:x64-windows@2021-02-21#2
• pthreadpool:x64-windows@2020-04-10#3
• sleef:[email protected]#4
• vcpkg-gfortran:x64-windows@3#3
• xnnpack:x64-windows@2022-12-22
  Additional packages (*) will be modified to complete this operation.
  Detecting compiler hash for triplet x64-windows...
  Compiler found: C:/Program Files/Microsoft Visual Studio/2022/Community/VC/Tools/MSVC/14.42.34433/bin/Hostx64/x64/cl.exeRestored 0 package(s) from C:\Users\hl\AppData\Local\vcpkg\archives in 6.94 ms. Use --debug to see more details.
  Installing 1/30 openblas:[email protected]...
  Building openblas:[email protected]...
  -- Downloading https://patch-diff.githubusercontent.com/raw/OpenMathLib/OpenBLAS/pull/4926.diff?full_index=1 -> openblas-fix-arm64-windows-uwp.patch...
  -- Downloading https://github.com/OpenMathLib/OpenBLAS/archive/v0.3.28.tar.gz -> OpenMathLib-OpenBLAS-v0.3.28.tar.gz...
  -- Extracting source D:/di/Libraries/Vs/vcpkg/downloads/OpenMathLib-OpenBLAS-v0.3.28.tar.gz
  -- Applying patch uwp.patch
  -- Applying patch fix-redefinition-function.patch
  -- Applying patch install-tools.patch
  -- Applying patch gcc14.patch
  -- Applying patch D:/di/Libraries/Vs/vcpkg/downloads/openblas-fix-arm64-windows-uwp.patch
  -- Using source at D:/di/Libraries/Vs/vcpkg/buildtrees/openblas/src/v0.3.28-6f3e63eb38.clean
  -- Downloading https://github.com/StrawberryPerl/Perl-Dist-Strawberry/releases/download/SP_54001_64bit_UCRT/strawberry-perl-5.40.0.1-64bit-portable.zip -> strawberry-perl-5.40.0.1-64bit-portable.zip...
  -- Found external ninja('1.12.1').
  -- Configuring x64-windows
  -- Building x64-windows-dbg
  -- Building x64-windows-rel
  -- Fixing pkgconfig file: D:/di/Libraries/Vs/vcpkg/packages/openblas_x64-windows/lib/pkgconfig/openblas.pc
  -- Using cached msys2-mingw-w64-x86_64-pkgconf-12.3.0-1-any.pkg.tar.zst.
  -- Using cached msys2-msys2-runtime-3.5.4-2-x86_64.pkg.tar.zst.
  -- Using msys root at D:/di/Libraries/Vs/vcpkg/downloads/tools/msys2/21caed2f81ec917b
  -- Fixing pkgconfig file: D:/di/Libraries/Vs/vcpkg/packages/openblas_x64-windows/debug/lib/pkgconfig/openblas.pc
  -- Installing: D:/di/Libraries/Vs/vcpkg/packages/openblas_x64-windows/share/openblas/copyright
  -- Performing post-build validation
  Stored binaries in 1 destinations in 1.3 s.
  Elapsed time to handle openblas:x64-windows: 3.1 min
  openblas:x64-windows package ABI: 399907692d18d6dea31219f03cdeaf916ed84fdd9e8468351197628ee836fecd
  Installing 2/30 blas:x64-windows@2023-04-14#1...
  Building blas:x64-windows@2023-04-14#1...
  -- Skipping post-build validation due to VCPKG_POLICY_EMPTY_PACKAGE
  Stored binaries in 1 destinations in 35.6 ms.
  Elapsed time to handle blas:x64-windows: 169 ms
  blas:x64-windows package ABI: 6e5e9aebd99702046d09b866f7df2c21d4d2f47ffdeb1a52c846de1cd2633bc8
  Installing 3/30 cpuinfo:x64-windows@2022-07-19#3...
  Building cpuinfo:x64-windows@2022-07-19#3...
  -- Note: cpuinfo only supports static library linkage. Building static library.
  -- Downloading https://github.com/pytorch/cpuinfo/archive/5e63739504f0f8e18e941bd63b2d6d42536c7d90.tar.gz -> pytorch-cpuinfo-5e63739504f0f8e18e941bd63b2d6d42536c7d90.tar.gz...
  -- Extracting source D:/di/Libraries/Vs/vcpkg/downloads/pytorch-cpuinfo-5e63739504f0f8e18e941bd63b2d6d42536c7d90.tar.gz
  -- Using source at D:/di/Libraries/Vs/vcpkg/buildtrees/cpuinfo/src/42536c7d90-2a15ca16d1.clean
  -- Found external ninja('1.12.1').
  -- Configuring x64-windows
  -- Building x64-windows-dbg
  -- Building x64-windows-rel
  -- Fixing pkgconfig file: D:/di/Libraries/Vs/vcpkg/packages/cpuinfo_x64-windows/lib/pkgconfig/libcpuinfo.pc
  -- Using cached msys2-mingw-w64-x86_64-pkgconf-12.3.0-1-any.pkg.tar.zst.
  -- Using cached msys2-msys2-runtime-3.5.4-2-x86_64.pkg.tar.zst.
  -- Using msys root at D:/di/Libraries/Vs/vcpkg/downloads/tools/msys2/21caed2f81ec917b
  -- Fixing pkgconfig file: D:/di/Libraries/Vs/vcpkg/packages/cpuinfo_x64-windows/debug/lib/pkgconfig/libcpuinfo.pc
  -- Installing: D:/di/Libraries/Vs/vcpkg/packages/cpuinfo_x64-windows/share/cpuinfo/copyright
  -- Performing post-build validation
  Stored binaries in 1 destinations in 88.4 ms.
  Elapsed time to handle cpuinfo:x64-windows: 5 s
  cpuinfo:x64-windows package ABI: a740dc2c5897837227473b59a949eca99516b7a8052aad6fec422d06bc735115
  Installing 4/30 cuda:[email protected]#13...
  Building cuda:[email protected]#13...
  -- Found CUDA V11.0.221
  -- Installing: D:/di/Libraries/Vs/vcpkg/packages/cuda_x64-windows/share/cuda/copyright
  -- Performing post-build validation
  Stored binaries in 1 destinations in 29.6 ms.
  Elapsed time to handle cuda:x64-windows: 217 ms
  cuda:x64-windows package ABI: bb864d1c37aa56ab8e7f1f33e40f0b58954fdf5ed2127ae2a5888a16de76928a
  Installing 5/30 cudnn:[email protected]#14...
  Building cudnn:[email protected]#14...
  -- Found CUDA V11.0.221
  -- CUDNN_INCLUDE_DIR: CUDNN_INCLUDE_DIR-NOTFOUND
  -- CUDNN_LIBRARY: CUDNN_LIBRARY-NOTFOUND
  CMake Error at ports/cudnn/portfile.cmake:62 (message):
  Please download CUDNN from official sources
  (https://developer.nvidia.com/cudnn) and install it
  Call Stack (most recent call first):
  scripts/ports.cmake:192 (include)

error: building cudnn:x64-windows failed with: BUILD_FAILED
See https://learn.microsoft.com/vcpkg/troubleshoot/build-failures?WT.mc_id=vcpkg_inproduct_cli for more information.
Elapsed time to handle cudnn:x64-windows: 139 ms
Please ensure you're using the latest port files with git pull and vcpkg update.
Then check for known issues at:
https://github.com/microsoft/vcpkg/issues?q=is%3Aissue+is%3Aopen+in%3Atitle+cudnn
You can submit a new issue at:
https://github.com/microsoft/vcpkg/issues/new?title=[cudnn]+Build+error+on+x64-windows&body=Copy+issue+body+from+D%3A%2Fdi%2FLibraries%2FVs%2Fvcpkg%2Finstalled%2Fvcpkg%2Fissue_body.md

[github-actions]

Please see #30604 for how to properly report a build failure.

[HoLue]

Sorry thought *.md was included automatically. here it is:

Package: cudnn:[email protected]#14

Host Environment

Host: x64-windows
Compiler: MSVC 19.42.34435.0
vcpkg-tool version: 2024-10-18-e392d7347fe72dff56e7857f7571c22301237ae6 vcpkg-scripts version: 9d163f061 2024-11-05 (6 weeks ago)

To Reproduce

vcpkg install libtorch:x64-windows

Failure logs

-- Found CUDA V11.0.221
-- CUDNN_INCLUDE_DIR: CUDNN_INCLUDE_DIR-NOTFOUND
-- CUDNN_LIBRARY: CUDNN_LIBRARY-NOTFOUND
CMake Error at ports/cudnn/portfile.cmake:62 (message):
Please download CUDNN from official sources
(https://developer.nvidia.com/cudnn) and install it
Call Stack (most recent call first):
scripts/ports.cmake:192 (include)

[HoLue]

There is a recommodation what to do, but that's somehow very difficult on my system. therefore the question if I can install without cuda, cudnn, gpu.

[JoergAtGithub]

I've the same issue, libtorch is a generic library, which should build without the GPU drivers of a third-party vendor. My system does not have a NVIDIA GPU, why should I install their drivers to build libtorch?

[dg0yt]

CUDA is required for the port's default feature cuda. To suppress default features at the command line, use the core pseudo-feature and explicitly include all desired features, e.g.

vcpkg install libtorch[core,zstd]

In manifest mode, use a libtorch dependency with "default-features": false.

[HoLue]

@dg0yt Thanks!! May I ask which features are default? (Where can I find out? Actually as I am no expert therein I want to install all default features except cuda.)

[dg0yt]

The immediate source is the port's manifest:

vcpkg/ports/libtorch/vcpkg.json

Lines 54 to 61 in f06267d

	"default-features": [
	{
	"name": "cuda",
	"platform": "windows & !x86"
	},
	"opencv",
	"zstd"
	],

There is no "global" list, but the active default features vary subject to platform expressions. On vcpkg.link, you will find some features marked with Default Feature and the relevant expression:
https://vcpkg.link/ports/libtorch

On a linux console, the following also works for me:

vcpkg install --dry-run libtorch | grep libtorch
    libtorch[core,opencv,zstd]:[email protected]#8

(You see the effect of the windows & !x86 expression.)

[HoLue]

@dg0yt Thanks a lot! This owrks somehow in principle. Now I have an error compiling onnx-optimizer, something with protobuf.... I created another ticket therefore:
#42854