March 2025 Update

[martindevans]

Updated to llama.cpp be7c3034108473beda214fd1d7c98fd6a7a3bdf5, built with https://github.com/martindevans/LLamaSharp/actions/runs/13838284757.

This includes the musl changes introduced by @Coopaguard in #1100.

• Windows CPU
• Windows CUDA
• Windows Vulkan
• Linux CPU
• Linux CPU (musl)
• Linux CUDA
• Linux Vulkan
• MacOS

[LoicDagnas]

Hello @martindevans, how difficult would it be to update the binaries up to the work added by this PR ggml-org/llama.cpp#12322 to have image embedders back on GPU ? 🙏🏼

[m0nsky]

I haven't started testing the other platforms yet, I would be fine with it. The difference seems to be only 2 days so hopefully not too difficult on the C# side.

[martindevans]

There have been no changes to llama.h since 6fefc05a7a4e676780ae10b0a4d0728e5281f367, which is what this PR is based on, therefore it should be as simple as running another build with a newer commit. I'll start one running now.

I'll wait a few hours and run one later this evening. I didn't realise this PR had only just been merged so it doesn't have a corresponding llama.cpp release yet.

[martindevans]

New build run started: https://github.com/martindevans/LLamaSharp/actions/runs/13838284757

[martindevans]

@Coopaguard @LoicDagnas @m0nsky: I've updated to be7c3034108473beda214fd1d7c98fd6a7a3bdf5.

[martindevans]

On my PC all the tests pass except the kernel memory ones. You can probably go ahead with testing the various platforms while I look into that

[SignalRT]

On my PC all the tests pass except the kernel memory ones. You can probably go ahead with testing the various platforms while I look into that

I have the same issue running the KernelMemory tests, but that brings me to another question. Why does this library exist in LlamaSharp? Isn’t there a connector in KernelMemory to use LlamaSharp? It seems to duplicate work.

[martindevans]

That's a good point. I don't really know what the history is there or how they compare though. I've only ever touched the km integration to fix breaking changes (like this PR).

[martindevans]

KM tests should be fixed now

[SignalRT]

KM tests should be fixed now

It's fixed.

I get a warning on Rider IDE related with a security issue on "SixLabors.ImageSharp" Version="3.1.5" and I tested also version 3.1.7 without problems. Reference: https://osv.dev/vulnerability/GHSA-2cmq-823j-5qj8

I do not commit the change to upgrade the SixLabors.ImageSharp package.

[martindevans]

I'll have a look at the ImageSharp thing in a separate PR, thanks for testing it.

[m0nsky]

Unit tests passed on Windows CUDA & Linux CUDA.

Test application is running fine on:

• Windows CPU
• Windows CUDA
• Windows Vulkan
• Linux CPU
• Linux CUDA
• Linux Vulkan