Feature/add android support

[AmSmart]

Description of Changes

• Updated compile Android to re-enable build and fix pipeline errors
• Added Android native libs to the CPU backend .nuspec file
• Updated .targets file to include new Android additions
• Added an extra condition to skip custom DLL resolution on Android because Android auto resolves libs based on generated apk
• Add Maui test Project (only Android enabled at the moment, will enable iOS in future PR)

Please let me know if any updates are required.

[AmSmart]

cc: @martindevans

[martindevans]

Thanks for the hard work @AmSmart! I'll try to find some time over the weekend to give this a proper in depth review.

Looking at the CI failures, it looks like some new workloads might need to be installed to compile the project. That'll require some changes in this file I would guess.

[amirvenus]

Is it CPU inference only or can we have GPU acceleration as well?

[AmSmart]

It's only CPU in this PR.

As far as I know, Mobile GPU inference is still quite slow at the moment. Saw it was even slower than CPU inference in some cases. Besides, it takes a bit of effort building CLBlast from source and trying to get that to work with the Android builds so I didn't consider it.

[martindevans]

Can these be split into a separate Backend package for Android (e.g. Backend.CPU.Android)?

[martindevans]

I've given this a deeper look, overall it looks good!

Left to resolve:

• Fix those CI errors (workloads must be installed: maui-android)
• Change nuget package to split out Android binaries from default ones
• Merge conflict
  • If you're splitting up the nuget this should be easy, just revert all changes to the backend.CPU.nuspec
• Test run (actually just found this run, so it's done)

[AmSmart]

Great, thanks for the review. Will follow up with the requested changes later today.

[AmSmart]

Hello @martindevans, I've made updates to the PR. Looks like we're good to go. Not sure how you trigger the Unit Test pipelines in this repo, though, to get a finite confirmation.

[martindevans]

Thanks @AmSmart, I'll be doing an update to the llama.cpp version soon (which will involve generating new binaries). I'll be basing that PR off this, so it should all be merged in then.

[AmSmart]

Great. Glad to hear it

[amirvenus]

Will this include VLM inference as well for multimodal models such as Qwen2.5-VL 0.5B or Gemma3 3b?

Thanks

[amirvenus]

Don't we also need an AndroidManifest.xml file?

[martindevans]

@AmSmart I've merged this work into a WIP branch and done a test run. Unfortunately one of the Android targets failed, could you take a look at it (here)? if you have any fixes to PR please target them at that branch.

[AmSmart]

I'll take a look at it

[martindevans]

It looks like clip.cpp is incompatible with 32 bit architectures, so we might have to disable the x86 target. That doesn't seem like a huge loss, I can't imagine there are many x86 Android devices out there?

[AmSmart]

Ahh, I see. Not surprising at all. Majority of Android devices are arm64, so you're right about it mostly being an acceptable/negligible loss.

PS: I even had to disable building some of the binary examples because it didn't support x86. It's mostly a relic of the past at this point.

[martindevans]

Ok I disabled the x86 target and started a new test run: https://github.com/SciSharp/LLamaSharp/actions/runs/14580598934