app.py

import spaces
import json
import subprocess
from llama_cpp import Llama
from llama_cpp_agent import LlamaCppAgent, MessagesFormatterType
from llama_cpp_agent.providers import LlamaCppPythonProvider
from llama_cpp_agent.chat_history import BasicChatHistory
from llama_cpp_agent.chat_history.messages import Roles
import gradio as gr
from huggingface_hub import hf_hub_download

# 下載Sakura-32B模型
hf_hub_download(
    repo_id="SakuraLLM/Sakura-32B-Qwen2beta-v0.10pre1-GGUF",
    filename="sakura-32b-qwen2beta-v0.10pre1-q4km.gguf",
    local_dir="./models"
)

llm = None
llm_model = None

@spaces.GPU(duration=120)
def respond(
    message,
    history: list[tuple[str, str]],
    model='sakura-32b-qwen2beta-v0.10pre1-q4km.gguf'
):
    chat_template = MessagesFormatterType.GEMMA_2
    system_message = "你是一个轻小说翻译模型，可以流畅通顺地使用给定的术语表以日本轻小说的风格将日文翻译成简体中文，并联系上下文正确使用人称代词，注意不要混淆使役态和被动态的主语和宾语，不要擅自添加原文中没有的代词，也不要擅自增加或减少换行。"

    global llm
    global llm_model
    
    if llm is None or llm_model != model:
        llm = Llama(
            model_path=f"models/{model}",
            flash_attn=True,
            n_gpu_layers=81,
            n_batch=1024,
            n_ctx=8192,
        )
        llm_model = model

    provider = LlamaCppPythonProvider(llm)

    agent = LlamaCppAgent(
        provider,
        system_prompt=f"{system_message}",
        predefined_messages_formatter_type=chat_template,
        debug_output=True
    )
    
    settings = provider.get_provider_default_settings()
    settings.temperature = 0.1
    settings.top_p = 0.3
    settings.do_sample = True
    settings.num_beams = 1
    settings.repetition_penalty = 1
    settings.max_new_tokens = 512
    settings.min_new_tokens = 1
    settings.stream = True

    messages = BasicChatHistory()

    for msn in history:
        user = {
            'role': Roles.user,
            'content': msn[0]
        }
        assistant = {
            'role': Roles.assistant,
            'content': msn[1]
        }
        messages.add_message(user)
        messages.add_message(assistant)
    
    stream = agent.get_chat_response(
        message,
        llm_sampling_settings=settings,
        chat_history=messages,
        returns_streaming_generator=True,
        print_output=False
    )
    
    outputs = ""
    for output in stream:
        outputs += output
        yield outputs

description = """<p align="center">Defaults to Sakura-32B-Qwen2beta</p>
<p><center>
<a href="https://huggingface.co/SakuraLLM/Sakura-32B-Qwen2beta-v0.10pre1-GGUF" target="_blank">[Sakura-32B-Qwen2beta Model]</a>
</center></p>
"""

demo = gr.ChatInterface(
    respond,
    retry_btn="Retry",
    undo_btn="Undo",
    clear_btn="Clear",
    submit_btn="Send",
    title="Chat with Sakura 32B using llama.cpp", 
    description=description,
    chatbot=gr.Chatbot(
        scale=1, 
        likeable=False,
        show_copy_button=True
    )
)

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