app.py

import os
import sys
import time
import urllib.request
from huggingface_hub import snapshot_download

# --- 配置（可通过环境变量覆盖） ---
MODEL_REPO = os.getenv("MODEL_REPO", "lastmass/Qwen3_Medical_GRPO")
MODEL_FILE = os.getenv("MODEL_FILE", "Qwen3_Medical_GRPO.i1-Q4_K_M.gguf")
MODEL_DIR = os.getenv("MODEL_DIR", "/models")
MODEL_PATH = os.path.join(MODEL_DIR, MODEL_FILE)
HF_TOKEN = os.getenv("HF_TOKEN", None)  # 如果模型是私有的，请在 Space Secrets 设置
N_THREADS = int(os.getenv("N_THREADS", "8"))

os.makedirs(MODEL_DIR, exist_ok=True)

def download_via_http(url, dest_path, token=None, chunk_size=4*1024*1024):
    """使用 urllib 分块下载，支持 token（Bearer）"""
    print(f"Downloading via HTTP: {url} -> {dest_path}")
    req = urllib.request.Request(url)
    if token:
        req.add_header("Authorization", f"Bearer {token}")
    try:
        with urllib.request.urlopen(req, timeout=120) as resp:
            # 如果状态不是 200，抛错
            if resp.status not in (200, 302, 301):
                raise RuntimeError(f"HTTP download returned status {resp.status}")
            # 写入临时文件，下载完成后重命名
            tmp_dest = dest_path + ".part"
            with open(tmp_dest, "wb") as fh:
                while True:
                    chunk = resp.read(chunk_size)
                    if not chunk:
                        break
                    fh.write(chunk)
            os.replace(tmp_dest, dest_path)
            print("HTTP download finished.")
    except Exception as e:
        if os.path.exists(dest_path):
            os.remove(dest_path)
        raise

def download_model_if_missing():
    if os.path.exists(MODEL_PATH):
        print(f"Model already exists at {MODEL_PATH}")
        return

    print(f"Model not found at {MODEL_PATH}. Trying snapshot_download from {MODEL_REPO} ...")
    # 先尝试使用 huggingface_hub.snapshot_download（优先）
    try:
        outdir = snapshot_download(
            repo_id=MODEL_REPO,
            repo_type="model",
            local_dir=MODEL_DIR,
            token=HF_TOKEN,
            allow_patterns=[MODEL_FILE]  # 仅抓取我们需要的文件
        )
        # snapshot_download 有时会返回 download 目录；搜索目标文件
        found = None
        for root, _, files in os.walk(outdir):
            if MODEL_FILE in files:
                found = os.path.join(root, MODEL_FILE)
                break
        if found:
            # 如果找到了，把它移动到 MODEL_PATH（若已在正确位置则跳过）
            if os.path.abspath(found) != os.path.abspath(MODEL_PATH):
                print(f"Found model at {found}, moving to {MODEL_PATH}")
                os.replace(found, MODEL_PATH)
            print("snapshot_download succeeded.")
            return
        else:
            print("snapshot_download did not find the file (0 files). Will try direct HTTP download as fallback.")
    except Exception as e:
        print("snapshot_download failed / returned nothing:", e, file=sys.stderr)

    # 备用：直接构造 resolve URL 并下载
    direct_url = f"https://huggingface.co/{MODEL_REPO}/resolve/main/{MODEL_FILE}"
    try:
        download_via_http(direct_url, MODEL_PATH, token=HF_TOKEN)
        return
    except Exception as e:
        print("Direct HTTP download failed:", e, file=sys.stderr)
        # 最后再尝试在 MODEL_DIR 下搜索一遍（保险）
        for root, _, files in os.walk(MODEL_DIR):
            if MODEL_FILE in files:
                found = os.path.join(root, MODEL_FILE)
                print(f"Found model at {found} after fallback search; moving to {MODEL_PATH}")
                os.replace(found, MODEL_PATH)
                return
        raise RuntimeError(f"Model download finished but {MODEL_PATH} still not found. Check repo contents and network.")

# 执行下载
download_model_if_missing()
time.sleep(0.5)

# --- 导入并初始化 llama_cpp（确保模型存在） ---
try:
    from llama_cpp import Llama
except Exception as e:
    print("Failed to import llama_cpp. Ensure the wheel matches the runtime and required system libs are present.", file=sys.stderr)
    raise

if not os.path.exists(MODEL_PATH):
    raise RuntimeError(f"Model path does not exist after download: {MODEL_PATH}")

n_threads = max(1, N_THREADS)
llm = Llama(model_path=MODEL_PATH, n_ctx=4096, n_threads=n_threads)

# --- gradio 接口 ---
import gradio as gr

system_prompt = """You are given a problem.
Think about the problem and provide your working out.
Place it between <start_working_out> and <end_working_out>.
Then, provide your solution between <SOLUTION></SOLUTION>"""

def chat(user_input):
    try:
        prompt = system_prompt + "\n\nUser input: " + user_input + " <start_working_out>"
        response = llm(prompt, max_tokens=1024, temperature=0.7)  # 改为1024
        return response["choices"][0]["text"]
    except Exception as e:
        err_msg = f"Error while generating: {e}"
        print(err_msg, file=sys.stderr)
        return err_msg

with gr.Blocks() as demo:
    gr.Markdown("# 🏥 Qwen3 Medical GGUF Demo")
    
    # 添加警告和说明信息
    with gr.Row():
        with gr.Column():
            gr.HTML("""
            <div style="background-color: #fff3cd; border: 1px solid #ffeaa7; border-radius: 5px; padding: 15px; margin-bottom: 20px;">
                <h3 style="color: #856404; margin-top: 0;">⚠️ Performance Notice</h3>
                <p style="color: #856404; margin-bottom: 10px;">
                    This demo runs the <strong>lastmass/Qwen3_Medical_GRPO</strong> model (Q4_K_M quantized version) 
                    on Hugging Face's free CPU hardware. Inference is <strong>very slow</strong>.
                </p>
                <p style="color: #856404; margin-bottom: 0;">
                    For better performance, we recommend running inference <strong>locally</strong> with GPU acceleration. 
                    Please refer to the <a href="https://huggingface.co/lastmass/Qwen3_Medical_GRPO" target="_blank">model repository</a> 
                    for usage instructions. For optimal performance, use <strong>vLLM</strong> for inference.
                </p>
            </div>
            """)
    
    with gr.Row():
        with gr.Column():
            input_box = gr.Textbox(label="Input your question", placeholder="Please enter your medical question...")
            submit_btn = gr.Button("Generate Response")
        with gr.Column():
            output_box = gr.Textbox(label="Model Response", lines=10)
    
    submit_btn.click(fn=chat, inputs=input_box, outputs=output_box)

demo.launch(server_name="0.0.0.0", server_port=7860)