Add create_chat_completion call for LlamaCpp cover

[Sm1Ling]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

Hello everyone!

Recently noticed that the class langchain_community.llms.llamacpp.LlamaCpp uses only _call or _stream functions to communicate with original llama_cpp.llama.Llama class.

Inside of _call function it makes forward calling of llama_cpp.llama.Llama
Example: result = self.client(prompt=prompt, **params) where client is llama_cpp.llama.Llama object

Such relization blocks opportunity to use chat models with given ChatPromptValue as input. The ChatPromptTemplate simply turns into string and passes to _call.

Probably it worth adding handling of llama_cpp.llama.Llama.create_chat_completion method to create support for the Chat models

Motivation

• LlamaCpp cover for llama_cpp.llama.Llama currently can't properly handle Chat messages
• That causes troubles with models designed to parse chat messages (for instance, istruction tuned models)

Proposal (If applicable)

The parent BaseLLM class of LlamaCpp
makes invoke call thread: LLMChain.invoke -> BaseLLM.invoke -> BaseLLM.generate_prompt -> BaseLLM.generate_prompt -> BaseLLM._generate_helper -> LLM._generate -> LlamaCpp._call -> llama_cpp.llama.Llama.__call__

I have noticed that in BaseLLM.generate_prompt every prompt just turns into string. And that is common for every single model. So every model looses its opportunity to parse sequence of messages by its own (many models on huggingface have their Chat template parsing)

I suppose. It is more global problem than it seemed. It will be very usefull to leave ability of LLMs to parse chat models by their own and not parse messages inside of LangChain into string

[ThanhNguye-n]

#22589 Hope help you 👍

[Sm1Ling]

Looking forward for your progress!