app.py

import streamlit as st
from huggingface_hub import hf_hub_download
from llama_cpp import Llama
import time

# --- 1. 页面基础配置 ---
st.set_page_config(
    page_title="Llama 3.2 AI Assistant",
    page_icon="🤖",
    layout="wide",  # 使用宽屏模式
    initial_sidebar_state="expanded"
)

# --- 2. 自定义 CSS (美化界面) ---
st.markdown("""
<style>
    /* 隐藏 Streamlit 默认的汉堡菜单和页脚 */
    #MainMenu {visibility: hidden;}
    footer {visibility: hidden;}
    header {visibility: hidden;}
    
    /* 调整主容器的顶部 padding，让内容更紧凑 */
    .block-container {
        padding-top: 2rem;
        padding-bottom: 2rem;
    }
    
    /* 美化侧边栏 */
    section[data-testid="stSidebar"] {
        background-color: #f7f9fc; /* 浅灰蓝背景 */
    }
    
    /* 自定义标题样式 */
    .title-text {
        font-family: 'Helvetica Neue', sans-serif;
        font-weight: 700;
        font-size: 2.5rem;
        color: #1E88E5; /* 科技蓝 */
        text-align: center;
        margin-bottom: 20px;
    }
    
    .subtitle-text {
        font-family: 'Helvetica Neue', sans-serif;
        font-weight: 400;
        font-size: 1.1rem;
        color: #666;
        text-align: center;
        margin-bottom: 40px;
    }
</style>
""", unsafe_allow_html=True)

# --- 3. 标题区域 ---
st.markdown('<div class="title-text">🤖 Llama 3.2-3B AI Assistant</div>', unsafe_allow_html=True)
st.markdown('<div class="subtitle-text">Powered by Marcus719/Llama-3.2-3B-changedata-Lab2-GGUF</div>', unsafe_allow_html=True)

# --- 4. 侧边栏 (控制面板) ---
with st.sidebar:
    st.image("https://huggingface.co/front/assets/huggingface_logo-noborder.svg", width=50)
    st.header("⚙️ 控制面板")
    
    # 参数设置
    temperature = st.slider("Temperature (创造性)", min_value=0.1, max_value=1.5, value=0.7, step=0.1, help="值越高，回答越随机；值越低，回答越严谨。")
    max_tokens = st.slider("Max Tokens (最大长度)", min_value=64, max_value=2048, value=512, step=64)
    
    st.divider()
    
    # 系统提示词 (System Prompt)
    system_prompt = st.text_area(
        "系统设定 (System Prompt)", 
        value="You are a helpful and polite AI assistant.",
        height=100
    )
    
    st.divider()
    
    # 清除历史按钮
    if st.button("🗑️ 清除对话历史", use_container_width=True):
        st.session_state.messages = []
        st.rerun()

    st.markdown("---")
    st.markdown("Optimization: **Unsloth Q4_K_M**")

# --- 5. 模型加载逻辑 ---
REPO_ID = "Marcus719/Llama-3.2-3B-changedata-Lab2-GGUF"
FILENAME = "unsloth.Q4_K_M.gguf"

@st.cache_resource
def load_model():
    model_path = hf_hub_download(repo_id=REPO_ID, filename=FILENAME)
    llm = Llama(
        model_path=model_path,
        n_ctx=4096,
        n_threads=2, # HF Spaces free tier limit
        verbose=False
    )
    return llm

try:
    if "llm" not in st.session_state:
        with st.spinner("🚀 正在启动 AI 引擎，请稍候..."):
            st.session_state.llm = load_model()
except Exception as e:
    st.error(f"模型加载失败: {e}")

# --- 6. 聊天逻辑 ---

# 初始化历史
if "messages" not in st.session_state:
    st.session_state.messages = []

# 显示历史消息
for message in st.session_state.messages:
    # 设置不同的头像
    avatar = "🧑‍💻" if message["role"] == "user" else "🤖"
    with st.chat_message(message["role"], avatar=avatar):
        st.markdown(message["content"])

# 处理用户输入
if prompt := st.chat_input("在此输入您的问题..."):
    # 1. 显示用户输入
    st.session_state.messages.append({"role": "user", "content": prompt})
    with st.chat_message("user", avatar="🧑‍💻"):
        st.markdown(prompt)

    # 2. 生成 AI 回复
    with st.chat_message("assistant", avatar="🤖"):
        message_placeholder = st.empty()
        full_response = ""
        
        # 构建带 System Prompt 的消息列表
        messages_payload = [{"role": "system", "content": system_prompt}] + [
            {"role": m["role"], "content": m["content"]}
            for m in st.session_state.messages
        ]
        
        stream = st.session_state.llm.create_chat_completion(
            messages=messages_payload,
            stream=True,
            max_tokens=max_tokens,
            temperature=temperature
        )
        
        for chunk in stream:
            if "content" in chunk["choices"][0]["delta"]:
                token = chunk["choices"][0]["delta"]["content"]
                full_response += token
                # 模拟打字机效果，稍微平滑一点显示
                message_placeholder.markdown(full_response + "▌")
        
        message_placeholder.markdown(full_response)
    
    # 3. 保存回复
    st.session_state.messages.append({"role": "assistant", "content": full_response})