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"  # 確認此模型有支援 VLM (目前有)

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

# -------------------------------
# 主處理函數
# -------------------------------
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"]

        # ------------------------------------------------
        # ⭐ 1) Base64 → 圖檔並存成 temp.jpg
        # ------------------------------------------------
        img_bytes = base64.b64decode(image_b64)
        temp_path = "temp.jpg"
        with open(temp_path, "wb") as f:
            f.write(img_bytes)

        # ------------------------------------------------
        # ⭐ 2) JSON 部分（只放文字）
        # ------------------------------------------------
        data = {
            "model": MODEL,
            "messages": [
                {
                    "role": "user",
                    "content": [
                        {"type": "text", "text": "Describe this image in detail."}
                    ]
                }
            ]
        }

        # ------------------------------------------------
        # ⭐ 3) 用 multipart/form-data 傳送 image + JSON payload
        # ------------------------------------------------
        resp = requests.post(
            "https://router.huggingface.co/v1/chat/completions",
            headers={"Authorization": f"Bearer {HF_TOKEN}"},
            data={"payload": json.dumps(data)},
            files={"file": ("image.jpg", open(temp_path, "rb"), "image/jpeg")},
            timeout=60
        )

        # ------------------------------------------------
        # ⭐ 4) 處理回應
        # ------------------------------------------------
        if resp.status_code != 200:
            print(f"VLM API error: {resp.status_code}, {resp.text}")
            return {"error": f"VLM API error: {resp.status_code}, {resp.text}"}

        # 正常解析內容
        try:
            content = resp.json()["choices"][0]["message"]["content"]
            # content 是 array，找出 text
            vlm_text = ""
            for part in content:
                if part.get("type") == "text":
                    vlm_text += part["text"]
        except Exception 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.strip()
        }

    except Exception as e:
        print(f"Unexpected error: {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)