app.py

import os
import gradio as gr
import copy
from llama_cpp import Llama
from huggingface_hub import hf_hub_download  


llm = Llama(
    model_path=hf_hub_download(
        repo_id=os.environ.get("REPO_ID", "TheBloke/Llama-2-7b-Chat-GGUF"),
        filename=os.environ.get("MODEL_FILE", "llama-2-7b-chat.Q5_0.gguf"),
    ),
    n_ctx=2048,
    n_gpu_layers=50, # change n_gpu_layers if you have more or less VRAM 
)


def generate_text(
    message,
    history: list[tuple[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
):
    temp = ""
    input_prompt = f"[INST] <<SYS>>\n{system_message}\n<</SYS>>\n\n "
    for interaction in history:
        input_prompt = input_prompt + str(interaction[0]) + " [/INST] " + str(interaction[1]) + " </s><s> [INST] "

    input_prompt = input_prompt + str(message) + " [/INST] "

    output = llm(
        input_prompt,
        temperature=temperature,
        top_p=top_p,
        top_k=40,
        repeat_penalty=1.1,
        max_tokens=max_tokens,
        stop=[
            "<|prompter|>",
            "<|endoftext|>",
            "<|endoftext|> \n",
            "ASSISTANT:",
            "USER:",
            "SYSTEM:",
        ],
        stream=True,
    )
    for out in output:
        stream = copy.deepcopy(out)
        temp += stream["choices"][0]["text"]
        yield temp


demo = gr.ChatInterface(
    generate_text,
    title="llama-cpp-python on GPU",
    description="Running LLM with https://github.com/abetlen/llama-cpp-python",
    examples=[
        ['How to setup a human base on Mars? Give short answer.'],
        ['Explain theory of relativity to me like I’m 8 years old.'],
        ['What is 9,000 * 9,000?'],
        ['Write a pun-filled happy birthday message to my friend Alex.'],
        ['Justify why a penguin might make a good king of the jungle.']
    ],
    cache_examples=False,
    retry_btn=None,
    undo_btn="Delete Previous",
    clear_btn="Clear",
)
demo.queue(concurrency_count=1, max_size=5)

if __name__ == "__main__":
    demo.launch()