app.py

import gradio as gr
import spaces
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

import operator
from typing import Annotated, Literal
from typing_extensions import TypedDict

from langchain_core.messages import AIMessage, AnyMessage, SystemMessage, HumanMessage, ToolMessage
from langgraph.graph import StateGraph, END


# 定义图的状态
class GraphState(TypedDict):
    messages: Annotated[list[AnyMessage], operator.add]


# 只加载一次模型和分词器
MODEL_NAME = "inclusionAI/Ring-mini-2.0"
device = "cuda" if torch.cuda.is_available() else "cpu"
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    torch_dtype=torch.float16 if device == "cuda" else torch.float32,
    trust_remote_code=True
).to(device)


# 定义图的节点
def call_model(state: GraphState):
    """模型调用节点"""
    messages = state["messages"]

    # 拼接 prompt
    prompt = ""
    for msg in messages:
        if msg.type == "system":
            prompt += f"{msg.content}\n"
        elif msg.type == "human":
            prompt += f"User: {msg.content}\n"
        elif msg.type == "ai":
            prompt += f"Assistant: {msg.content}\n"
    prompt += "Assistant:"

    input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
    output_ids = model.generate(
        input_ids,
        max_new_tokens=512, # 暂时硬编码
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id,
    )
    output = tokenizer.decode(output_ids[0][input_ids.shape[1]:], skip_special_tokens=True)
    
    return {"messages": [AIMessage(content=output)]}

# 构建图
workflow = StateGraph(GraphState)
workflow.add_node("llm", call_model)
workflow.set_entry_point("llm")
workflow.add_edge("llm", END)

# 编译图
app = workflow.compile()
@spaces.GPU
def respond(message, history, system_message, hf_token: gr.OAuthToken = None):
    """Gradio 接口的响应函数，调用 LangGraph 应用"""
    
    # 将 Gradio 的 history 格式转换为 LangChain 消息格式
    messages = []
    if system_message:
        messages.append(SystemMessage(content=system_message))
    
    for turn in history:
        user_message, bot_message = turn
        if user_message:
            messages.append(HumanMessage(content=user_message))
        if bot_message:
            messages.append(AIMessage(content=bot_message))
            
    messages.append(HumanMessage(content=message))

    # 使用 invoke 方法进行一次性调用
    inputs = {"messages": messages}
    final_state = app.invoke(inputs)
    
    # 从最终状态中提取最后一条消息
    final_response = final_state["messages"][-1].content
    
    return final_response

# 重新定义 ChatInterface
chatbot = gr.ChatInterface(
    respond,
    type="messages", # 改为 messages 类型以更好地匹配 LangChain
    additional_inputs=[
        gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
    ],
)

with gr.Blocks() as demo:
    gr.Markdown("# HuggingFace Running")
    with gr.Sidebar():
        gr.LoginButton()
    chatbot.render()


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