bug/feat: Please provide Cuda 13.3 support in the settings drop down menu

[kulbhushanchand]

Problem Description

As of now the settings menu does not have option for Cuda 13.3.
Because of this the Open WebUI always downloads the Cuda 12.4 version instead of Cuda 13.3.
See the screenshots below.

Open WebUI settings -

llama.cpp release download options -

Proposed Solution

Add CUDA 13.3 option in the settings menu, with relevant selector URL link parser.

Alternatives Considered

As of now I have to manually download the correct files (for e.g. https://github.com/ggml-org/llama.cpp/releases/download/b10655/llama-b10655-bin-win-cuda-13.3-x64.zip), extract and copy paste in the llama.cpp folder

[kulbhushanchand]

Proposed fix

The CUDA variant list in the Inference Runtime settings is currently hardcoded and out of date. I am happy to open a PR if useful, or a maintainer can take this directly.

Where the problem is

Two files need updating:

src/renderer/src/lib/components/Main/Settings/InferenceRuntime.svelte, the Settings dropdown, currently only offers CUDA 12.4 and CUDA 13.1
src/main/utils/llamacpp.ts, getAssetPattern(), which maps each variant to the exact llama.cpp release asset filename to download

What's missing

Checked llama.cpp's actual release history and the official Windows CUDA builds have moved through 12.4 → 13.1 → 13.3 → 13.4 as the underlying CUDA toolkit versions changed. Only 12.4 and 13.1 are wired up in the app, so anyone on a release that shipped 13.3/13.4 (e.g. b10666) can't select it and falls back to CPU-only.

Proposed change

Add cuda-13.3 and cuda-13.4 as additional pinned entries alongside the existing cuda-12.4 and cuda-13.1

[kulbhushanchand]

Opened #265 with the fix described above — adds cuda-13.3 and cuda-13.4 as pinned variant options in both InferenceRuntime.svelte (Settings dropdown) and llamacpp.ts (getAssetPattern() asset resolution).

Existing cuda-12.4 and cuda-13.1 entries are untouched, so pinning an older release tag that only shipped 13.1 still resolves correctly.

Ready for review whenever a maintainer has a chance to take a look. Happy to adjust the approach (e.g. resolving CUDA assets dynamically by major version instead of pinning each minor) if preferred.