Pre-built wheels for Python 3.14 and 3.14 free-threaded

[clemlesne]

Is your feature request related to a problem? Please describe.

Missing pre-built wheel requires long computation and slowdowns.

Describe the solution you'd like

Pre-build the wheels in GitHub Actions:

llama-cpp-python/.github/workflows/build-and-release.yaml

Line 78 in c37132b

CIBW_BUILD: "cp38-* cp39-* cp310-* cp311-* cp312-*"

cibuildwheel is compatible with CPython 3.14.0 Final since October 12.

Describe alternatives you've considered

Keep it like that. Let all the community lose hundreds of thousands of CPU minutes to rebuild the wheel again and again.

[rookiemann]

I have a new wheel for llama-cpp-python 0.3.16 if interested, it's not a free-threaded one but it is for Python 3.14.2 using CUDA toolkit 13.1

https://github.com/rookiemann/llama-cpp-python-py314-cuda131-wheel-or-python314-llama-cpp-gpu-wheel

[ngoldbaum]

I'm happy to help out with support for free-threaded Python. I've been working at Quansight for the past couple of years on ecosystem support for free-threaded Python.

I'm coming across this because llama-cpp-python is a test dependency of llguidance (see guidance-ai/llguidance#256), which itself is a dependency of vLLM (see vllm-project/vllm#28762).

Currently a priority for my team at Quansight is to improve support for vLLM on the free-threaded build, so this can be done on funded time in support of that.

Not sure if a maintainer could weigh in and indicate they're supportive - it looks like this repo hasn't had any recent commits.

[doniaalani-alt]

By the way how do I use bigfloat in python? I have 3.14.3