_test_inference.py

"""
Test inference endpoint for PregoPal MiniCPM-o API on Modal

Usage:
    python _test_inference.py
    
    # Or with custom URL:
    python _test_inference.py https://andrew-jiabin--prego-pal-minicpm-serve.modal.run
"""
import sys
import json
import httpx
import os

os.environ["PYTHONIOENCODING"] = "utf-8"

# Default URL - update after deployment
URL = "https://andrew-jiabin--prego-pal-minicpm-serve.modal.run"

if len(sys.argv) > 1:
    URL = sys.argv[1].rstrip("/")

payload = {
    "model": "MiniCPM-o-4_5",
    "messages": [{"role": "user", "content": "你好，请用一句话回答：1+1等于几？"}],
    "max_tokens": 50,
    "temperature": 0.7
}

print("=" * 60)
print("PregoPal MiniCPM-o API 测试")
print("=" * 60)
print(f"URL: {URL}/v1/chat/completions")
print("注意：冷启动需要 2-10 分钟加载模型...")
print()

try:
    r = httpx.post(
        URL + "/v1/chat/completions",
        json=payload,
        timeout=600  # 10 minutes for cold start
    )
    print(f"状态码: {r.status_code}")
    if r.status_code == 200:
        data = r.json()
        content = data["choices"][0]["message"]["content"]
        print(f"\n=== 模型响应 ===")
        print(content)
        print("=" * 20)
    else:
        print(f"\n响应 (前500字符):")
        print(r.text[:500])
except httpx.TimeoutException:
    print("\n请求超时 (600s) - 容器仍在初始化，检查 Modal 日志")
    print("  modal app logs $(modal app list | grep prego-pal | awk '{print $1}')")
except Exception as e:
    print(f"\n错误: {e}")