Current vendored llama.cpp won't build with MSVC 17.13

[Janzert]

Expected Behavior

llama-cpp-python correctly builds with latest MSVC compiler.

Current Behavior

The build dies with errors building llama.cpp due to issues with std::chrono in log.cpp.

Environment and Context

Windows 11
Visual Studio Community Edition v17.13.1

Steps to Reproduce

Follow standard or development (pip install -e .) build in readme.md

Failure Logs

      (ClCompile target) ->
        ...\llama-cpp-python\vendor\llama.cpp\common\log.cpp(28,79): error C2039: 'system_clock': is not a member of 'std::chrono' [...\build\vendor\llama.cpp\common\common.vcxproj]
        ...\llama-cpp-python\vendor\llama.cpp\common\log.cpp(28,79): error C3083: 'system_clock': the symbol to the left of a '::' must be a type [...\build\vendor\llama.cpp\common\common.vcxproj]
        ...\llama-cpp-python\vendor\llama.cpp\common\log.cpp(28,93): error C2039: 'now': is not a member of 'std::chrono' [...\build\vendor\llama.cpp\common\common.vcxproj]
        ...\llama-cpp-python\vendor\llama.cpp\common\log.cpp(28,93): error C3861: 'now': identifier not found [...\build\vendor\llama.cpp\common\common.vcxproj]
        ...\llama-cpp-python\vendor\llama.cpp\common\log.cpp(28,25): error C2672: 'std::chrono::duration_cast': no matching overloaded function found [...\build\vendor\llama.cpp\common\common.vcxproj]

Cause and resolution

This is apparently caused by a breaking change in MSVC (see ggml-org/llama.cpp#11834) and is fixed in PR ggml-org/llama.cpp#11836 with commit tinglou/llama.cpp@4e3db1e

Running git submodule update --remote vendor\llama.cpp locally fixes the issue.

[ethrx]

Hi @Janzert , having the same issue

Fresh-est install possible, latest build-tools and using python 3.10

Can you explain how you resolved it a bit more?

[Janzert]

To build with a newer llama.cpp you have to build from a git checkout, not from a wheel. So basically follow the instructions given under development in the readme. But after checking out llama-cpp-python run the above git submodule update command before doing the pip install or building a wheel if needed.

So the set of commands needed looks something like this:

git clone --recurse-submodules https://github.com/abetlen/llama-cpp-python.git
cd llama-cpp-python

# update to latest llama.cpp (this pulls in all llama.cpp changes not just the fix needed for MSVC, so may have other effects as well)
git submodule update --remote vendor\llama.cpp

# Upgrade pip (required for editable mode)
pip install --upgrade pip

# Install with pip
pip install -e .

[tjohnman]

Good to know there's a more "proper" fix. For the sake of completeness: I added #include <chrono> to vendor\llama.cpp\common\log.cpp, and the build completed successfully as well.

[Janzert]

Latest updates include the fix for this.