lora: make sure model keep track of associated adapters

[ngxson]

Ref: #18469 (comment)

I'm extracting this change to a dedicated PR for visibility

[ggerganov]

If we delegate it to the llama_model to free the adapters, then we should remove the llama_adapter_lora_free() from the public API.

For example, with the current change I believe we will be double freeing the adapters - once on model destruction and once more here for all llama_adapter_lora_ptr instances:

llama.cpp/include/llama-cpp.h

Lines 23 to 26 in cd78e57

	struct llama_adapter_lora_deleter {
	void operator()(llama_adapter_lora * adapter) { llama_adapter_lora_free(adapter); }
	};

[ngxson]

Yes that makes sense. llama_adapter_lora_free exists because we wanted to allow hot-loading / hot-removing of lora adapter. But I don't think it's actually used that way in reality.

By removing llama_adapter_lora_free, we will now assume that lora adapters are an "extension" to a model, and must be loaded before the model can be used (i.e. context creation from a model)

[ggerganov]

@ngxson This is probably good to merge?

[ngxson]

Ah yes, thanks. Merging it now