Steps to Build and Install llama-cpp-python 0.3.7 w/CUDA on Windows 11 [06/03/2025]

[VigneshRajan-AMRC]

After reviewing multiple GitHub issues, forum discussions, and guides from other Python packages, I was able to successfully build and install llama-cpp-python 0.3.7 with CUDA on Windows 11. Below are the exact steps I followed, which may help others facing similar build issues.

---

Errors Encountered During Build

I encountered the following build errors even after installing the right build tools. The issue was resolved by updating the llama.cpp submodule.

D:\llama-cpp-python\vendor\llama.cpp\common\log.cpp(28,79): error C2039: 'system_clock': is not a member of 'std::chrono' [%localappdata%\Temp\tmp0zkelev6\build\vendor\llama.cpp\common\common.vcxproj]

D:\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 [%localappdata%\Temp\tmp0zkelev6\build\vendor\llama.cpp\common\common.vcxproj]

D:\llama-cpp-python\vendor\llama.cpp\common\log.cpp(28,93): error C2039: 'now': is not a member of 'std::chrono' [%localappdata%\Temp\tmp0zkelev6\build\vendor\llama.cpp\common\common.vcxproj]

D:\llama-cpp-python\vendor\llama.cpp\common\log.cpp(28,93): error C3861: 'now': identifier not found [%localappdata%\Temp\tmp0zkelev6\build\vendor\llama.cpp\common\common.vcxproj]

D:\llama-cpp-python\vendor\llama.cpp\common\log.cpp(28,25): error C2672: 'std::chrono::duration_cast': no matching overloaded function found [%localappdata%\Temp\tmp0zkelev6\build\vendor\llama.cpp\common\common.vcxproj]

---

Hardware Setup

• OS: Windows 11
• CPU: Intel i9-9900K
• GPUs: RTX 3080, RTX A4000

Software Setup

• Git Bash: 2.48.1
• Python: 3.12.9 (likely works with 3.11 and 3.13)
• CMake: 4.0.0-rc3
• Visual Studio 2022: 17.13.2 with the following workloads installed -
  • Python development
  • Desktop development with C++
  • Linux and embedded development with C++
• CUDA Toolkit: 12.8.0 (installed without modifying PhysX, Display, and HD Audio drivers, as they were already installed by the latest Studio Driver Package v572.60 that I'm using)

---

Installation Steps

Make sure to use Command Prompt and not PowerShell, for the following steps.

1. Clone the Repository

Navigate to the directory where you want to build the package and run:

git clone --recursive -j%NUMBER_OF_PROCESSORS% https://github.com/abetlen/llama-cpp-python.git

2. Navigate to the Package Directory

cd llama-cpp-python

3. Update the llama.cpp Submodule

git submodule update --remote vendor\llama.cpp

4. Set Required Environment Variables

Configure the necessary environment variables:

set FORCE_CMAKE=1
set CMAKE_ARGS=-DGGML_CUDA=ON

# Optional variables to speed up the build process
# set CMAKE_BUILD_PARALLEL_LEVEL=%NUMBER_OF_PROCESSORS%
# set MAKEFLAGS=-j%NUMBER_OF_PROCESSORS%

These are only temporary and will be unset when you exit the current CMD session.

5. Build and Install

Run the following command to build and install the package:

python -m pip install . --no-cache-dir --force-reinstall --upgrade

---

This setup worked for me, and I hope it helps others facing similar build issues.

[VigneshRajan-AMRC]

Leaving the issue open for visibility. Repo maintainers can close this issue when the documentation is updated or these steps aren't relevant any more.

[Jamiegammon1979]

Didn't work for me. I got an error that it couldn't find Curl.

[Fredrik-Petersson]

Didn't work for me. I got an error that it couldn't find Curl.

I managed to resolve the Curl error by settings this Environmental Variable:
set CMAKE_ARGS=-DGGML_CUDA=ON -DLLAMA_CURL=OFF

After this everything worked but the build time took a very long time (15 minutes or so) as I didn´t set the CPU flags to run on more than one core.

Did some quick tests and GPU Acceleration works for Gemma2-9b and Gemma3-12b.

This is the first time I have been able to run the Gemma3 models with GPU acceleration so big thanks to @VigneshRajan-AMRC for posting this guide! Hope they will release the pre-built binaries for 0.3.7 to save others from this work.

[Granddyser]

I created a guide that solves most common llama-cpp-python CUDA issues on Windows: https://github.com/Granddyser/windows-llama-cpp-python-cuda-guide