Cannot build static windows libraries

[dnns92]

🐛 Describe the bug

🐛 Describe the bug

The torch_cpu.dll is ~200 Mb, which is way too large for us. Therefore, we'd like to build torch statically. I followed the procedures hinted in the README.

These are the commands I used:

git clone https://github.com/pytorch/pytorch.git
cd pytorch && git checkout tags/v2.0.1
git submodule update --init --recursive
python -m venv .venv && .venv\Scripts\activate
pip install -r requirements.txt
pip install mkl mkl-include cmake


# additionally as found in the link above
curl https://s3.amazonaws.com/ossci-windows/mkl_2020.2.254.7z -k -O
mkdir mkl
# move 7z to mkl dir and right click, unpack here
set CMAKE_INCLUDE_PATH=D:\00_src\pytorch-fresh-install\mkl\include
set "LIB=D:\00_src\pytorch-fresh-install\mkl\lib"
# end

cd .. && mkdir build-pytorch && cd build-pytorch
cmake -DBUILD_SHARED_LIBS:BOOL=OFF -DUSE_CUDA=0 -DCMAKE_BUILD_TYPE:STRING=MinSizeRel -DCMAKE_INSTALL_PREFIX:PATH=../pytorch-install ../pytorch
# use visual studio 2019 and ALL_BUILD with MinSizeRel to build

This builds for about an hour and fails with 1200 linker errors. Is building static libs not possible anymore?

Versions

Windows 11 21H2, x64
using pip for python

cc @malfet @seemethere @peterjc123 @mszhanyi @skyline75489 @nbcsm @vladimir-aubrecht @iremyux @Blackhex @cristianPanaite

[dnns92]

Also, there are currently no static versions downloadable for torch 2.0.1 btw.

[kit1980]

Can you give more info about your use case for PyTorch on Windows?
Foe example, can you use Windows Subsystem for Linux?

[dnns92]

In this case I want to run an ai model in a windows app. Yes, we might also do that with openVino or onnxruntime, but we like the flexibility of torch and dont really care about the execution time.

[kit1980]

In this case I want to run an ai model in a windows app. Yes, we might also do that with openVino or onnxruntime, but we like the flexibility of torch and dont really care about the execution time.

Do you need GPU support?

[malfet]

@dnns92 please note that static builds are unlikely going to be smaller than dynamic ones. What you probably want is lite interpreter, i.e. version of PyTorch that supports inference only (i.e. all gradient ops will be stripped) as well as limits what operators you are building PyTorch with, see this tutorial which is a bit offtopic but might help you get where you want: https://pytorch.org/tutorials/prototype/lite_interpreter.html

[dnns92]

In this case I want to run an ai model in a windows app. Yes, we might also do that with openVino or onnxruntime, but we like the flexibility of torch and dont really care about the execution time.

Do you need GPU support?

no

@dnns92 please note that static builds are unlikely going to be smaller than dynamic ones. What you probably want is lite interpreter, i.e. version of PyTorch that supports inference only (i.e. all gradient ops will be stripped) as well as limits what operators you are building PyTorch with, see this tutorial which is a bit offtopic but might help you get where you want: https://pytorch.org/tutorials/prototype/lite_interpreter.html

okay, thank you for your advise, we will have a look into it. Nevertheless, 250 Mbytes is really large, even for such an app, so static linking might enable us only using the parts that we really need and ending up with a much smaller filesize (at least thats the idea...)

So having this as a static library might still be valuable to other users.