llama-cpp-pascal-cuda-windows

CUDA-accelerated llama-cpp-python compile guides for legacy NVIDIA Pascal GPUs (GTX 1050/1060/1070/1080 series) on Windows, where modern Visual Studio refuses to compile against older CUDA toolkits out of the box.

Two verified toolchain paths are documented here, depending on which Visual Studio setup you'd rather deal with:

Path	CUDA	Visual Studio	Folder
A	12.1	2019 Build Tools (installed alongside 2022)	CUDA_12.1/
B	12.6	2022 (no second VS install needed)	CUDA_12.6/

Both are tested end-to-end on a GTX 1050 Ti with comparable inference performance (~27-29 tok/s either way) — pick whichever avoids more hassle on your machine. Path A needs a second VS install but keeps CUDA on an older, longer-verified version. Path B avoids installing VS2019 but requires the newer CUDA toolkit.

Path A — CUDA 12.1 + VS2019

1. CUDA_12.1/COMPILE_REQUIREMENTS.md — hardware/software versions this was tested on, why each version is pinned, and measured performance vs. Ollama.
2. CUDA_12.1/COMPILE_GUIDE.md — the actual step-by-step compile process, with the VS 2019 Build Tools workaround and flag explanations. Read the requirements first — it explains why the versions in the guide are pinned the way they are, which makes the guide itself easier to follow if something doesn't match your setup.

Path B — CUDA 12.6 + VS2022

1. CUDA_12.6/COMPILE_REQUIREMENTS.md
2. CUDA_12.6/COMPILE_GUIDE.md Same hardware target, no VS2019 required, includes its own benchmark data.