Hot-swap LoRA with updated llama.cpp

[ltoniazzi]

Describe the feature

Feature

Since the PRs linked below in llama.cpp, it is possible to how-swap lora adapters. This allows to personalise NPCs and other GenAi game features by fine-tuning adapters that then can be quickly swapped on the same base model in memory.

Since llama.cpp has is being updated in #209. It would be nice to check how easily one can integrate this feature.

Todo list

• bin files for adapters are now deprecated in favour of new gguf files, which should be ameneded in the documentation, if appearing.
• Add in documentation link to how to convert adapters to gguf.
• Add example on performing hot-swap
• Test on using the new gguf formats should be run (I suspect the new adapter will automatically be used in hot-swapping mode)
• Test using multiple adapters and hot swapping them, add code if needed, and add an example in the examples folder.

Related links

Hot lora PRs in llama.cpp:

• Cli: Refactor lora adapter support ggml-org/llama.cpp#8332
• Server: server : add lora hotswap endpoint ggml-org/llama.cpp#8857

Discord threads:

• Discussion on lora

[ltoniazzi]

I am happy to have a go at it 😃

[amakropoulos]

Wow thanks for the detailed request ⭐!
I implemented multiple loras and hot swapping in #210 . Do you want to have a look at it and let me know what you think?
I haven't added a sample or description in the Readme yet.

[amakropoulos]

I added a unit test for lora at #219 .
I'm adding some documentation and will release once I fix some build issue.