app.py

import gradio as gr
import json
import base64
from io import BytesIO
import requests
import os

# HF token & model
HF_TOKEN = os.environ.get("HF_CV_ROBOT_TOKEN")
MODEL = "Qwen/Qwen2.5-VL-7B-Instruct"  # HF 支援列表裡的模型

if not HF_TOKEN:
    print("ERROR: HF_CV_ROBOT_TOKEN environment variable not set.")

# -------------------------------
# 主處理函數 (Main Processing Function)
# -------------------------------
def process(payload: dict):
    try:
        if not HF_TOKEN:
            return {"error": "Hugging Face token is missing. Please check Space secrets."}

        robot_id = payload.get("robot_id", "unknown")
        image_b64 = payload["image_b64"]

        # Base64 -> bytes -> 保存為 tmp.jpg
        tmp_path = "tmp.jpg"
        with open(tmp_path, "wb") as f:
            f.write(base64.b64decode(image_b64))

        # 上傳 image file 到 HF Router
        files = {"file": open(tmp_path, "rb")}
        upload_resp = requests.post(
            "https://huggingface.co/api/uploads",
            headers={"Authorization": f"Bearer {HF_TOKEN}"},
            files=files
        )
        files["file"].close()
        os.remove(tmp_path)

        if upload_resp.status_code != 200:
            return {"error": f"HF upload failed: {upload_resp.status_code}, {upload_resp.text}"}

        file_info = upload_resp.json()
        file_url = file_info.get("href")  # 取得 HF hosted file URL

        # JSON payload 放文字訊息 + image file reference
        data = {
            "model": MODEL,
            "messages": [
                {
                    "role": "user",
                    "content": [
                        {"type": "text", "text": f"![]({file_url}) Describe this image in detail."}
                    ]
                }
            ]
        }

        resp = requests.post(
            "https://router.huggingface.co/v1/chat/completions",
            headers={"Authorization": f"Bearer {HF_TOKEN}"},
            data={"payload": json.dumps(data)},
            timeout=60
        )

        if resp.status_code != 200:
            return {"error": f"VLM API error: {resp.status_code}, {resp.text}"}

        try:
            vlm_text = resp.json()["choices"][0]["message"]["content"][0]["text"]
        except (KeyError, IndexError, json.JSONDecodeError) as e:
            return {"error": f"Failed to parse VLM response: {e}, Response text: {resp.text}"}

        return {
            "received": True,
            "robot_id": robot_id,
            "vllm_analysis": vlm_text
        }

    except Exception as e:
        return {"error": str(e)}

# -------------------------------
# Gradio MCP Server
# -------------------------------
demo = gr.Interface(
    fn=process,
    inputs=gr.JSON(label="Input Payload (Dict format)"),
    outputs=gr.JSON(label="Reply to Jetson"),
    api_name="predict"
)

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