feat: sync llama.cpp

[a-ghorbani]

Sync llama.cpp with the b3972 release.

This will close #77.

[a-ghorbani]

I’m almost done, just running the final tests.
Mostly based on @Vali-98's work: https://github.com/Vali-98/cui-llama.rn

@Vali-98 , if you have a moment, a quick review would be much appreciated to catch anything I might have missed!

[Vali-98]

It looks okay from a glance, can't confirm for the IOS code though.

Just to confirm, I applied changes for cui-llama.rn from my fork of llama.cpp: https://github.com/Vali-98/llama.cpp/tree/cui-llama.rn. @a-ghorbani , did you swap to this fork or translated my changes to the relevant patch files?

The current method of patching changes to llama.rn is the main reason why I wanted my own fork instead, as patching the files tends to fail spectacularly when syncing llama.cpp. After 2-3 sync's, I decided that it was a major maintainability issue. I assume this is the reason why the CI is failing too.

Ideally for release builds, we would include prebuilt binaries instead of needing to redundantly build rn-llama.cpp.

[a-ghorbani]

It looks okay from a glance, can't confirm for the IOS code though.

Cool! For the iOS part, I took some ideas from Android-related changes.

Just to confirm, I applied changes for cui-llama.rn from my fork of llama.cpp: https://github.com/Vali-98/llama.cpp/tree/cui-llama.rn. @a-ghorbani , did you swap to this fork or translated my changes to the relevant patch files?

I translated the changes into relevant patch files to maintain consistency with the approach in this repo.

The current method of patching changes to llama.rn is the main reason why I wanted my own fork instead, as patching the files tends to fail spectacularly when syncing llama.cpp. After 2-3 sync's, I decided that it was a major maintainability issue. I assume this is the reason why the CI is failing too.

Yep, maintaining these patches is a challenge. usually, I revert the changes and create the patches from scratch again.
(the ci was actually due to an issue with the submodule update had an extra --remote that led to fetching the latest rather than using the one pinned here; it should be fixed now).

[Vali-98]

You might want to mention that this adds the vocab_only field as it isn't actually a part of gpt_params on llama.cpp.

[jhen0409]

This is awesome, thanks for the contribution!

It looks like n_gpu_layers = 0 for disable GPU is broken (see ggerganov/llama.cpp#10089), I'll checkout for that before merge this.