fix(build): locate CUDA static libs on distro-packaged toolkits

[passabilities]

Description

./build-gpu.sh fails on Linux machines that use a distro-packaged CUDA toolkit:

error: could not find native static library `cudart_static`, perhaps an -L flag is missing?

The build gets as far as Detected GPU feature: cuda and then dies while compiling llama-cpp-sys-2.

Root cause

llama-cpp-sys-2/build.rs emits cargo:rustc-link-lib=static=cudart_static (plus cublas_static, culibos) and delegates -L path discovery to find_cuda_helper. That crate's find_cuda_lib_dirs() only probes $CUDA_LIBRARY_PATH, /opt/cuda/lib64 and /usr/local/cuda*/lib64.

Distro packages such as Ubuntu's nvidia-cuda-toolkit install into /usr, putting nvcc at /usr/bin/nvcc and the libraries in /usr/lib/x86_64-linux-gnu. None of the probed paths exist, so rustc receives the -l static= flags with zero -L paths and the link fails.

This reproduces in isolation: a two-line crate with #[link(name = "cudart_static", kind = "static")] fails identically without -L, and succeeds with -L native=/usr/lib/x86_64-linux-gnu.

Fix

Both frontend/build-gpu.sh and frontend/dev-gpu.sh now locate the CUDA static libraries before building and export a matching -L native= into RUSTFLAGS.

The block only runs when TAURI_GPU_FEATURE = cuda on linux-gnu, and probes in this order:

1. $CUDA_PATH/lib64
2. $CUDA_HOME/lib64
3. nvcc-relative ../lib64
4. /usr/local/cuda/lib64
5. /opt/cuda/lib64
6. /usr/lib/$(uname -m)-linux-gnu
7. /usr/lib64

The first directory containing libcudart_static.a wins. Conventional locations are checked before the multiarch path, so existing /usr/local/cuda and /opt/cuda installs keep their current behaviour. RUSTFLAGS is appended to rather than overwritten.

If no candidate matches, the script prints a warning pointing at CUDA_PATH and continues rather than aborting, since the link step may still succeed on setups this heuristic does not cover.

Related Issue

No existing issue covers this specific link failure. Related Linux build reports: #305, #587.

Type of Change

• Bug fix
• New feature
• Documentation update
• Performance improvement
• Code refactoring
• Other (please describe)

Testing

• Unit tests added/updated
• Manual testing performed
• All tests pass

On Ubuntu 24.04 with nvidia-cuda-toolkit (nvcc 12.0 at /usr/bin/nvcc, no /usr/local/cuda*):

• Before: the build fails at llama-cpp-sys-2 with the error above.
• After: the full build succeeds and produces both .deb and .AppImage bundles.
• ldd on the resulting binary confirms CUDA is genuinely linked: libcudart.so.12, libcublas.so.12, libcublasLt.so.12, libcuda.so.1.

Shell-script only; no effect on macOS or Windows, or on Linux builds that are not using the cuda feature.

Documentation

• Documentation updated
• No documentation needed

Checklist

• Code follows project style
• Self-reviewed the code
• Added comments for complex code
• Updated README if needed
• Branch is up to date with devtest
• No merge conflicts

Screenshots (if applicable)

N/A — build script change only.

Additional Notes

The probe order deliberately prefers the conventional CUDA install locations so that machines with a vendor toolkit under /usr/local/cuda are unaffected. The multiarch path is only reached when nothing more specific matched.