forked from phoenix/litellm-mirror
* fix - don't block proxy startup if not a premium user * test_litellm_proxy_server_config_with_prometheus * add test for proxy startup * fix remove unused test * fix startup test * add comment on bad-license
51 lines
1.7 KiB
Python
51 lines
1.7 KiB
Python
"""
|
|
This test ensures that the proxy starts and serves requests even with a bad license.
|
|
|
|
|
|
in ci/cd config.yml, we set the license to "bad-license"
|
|
"""
|
|
|
|
import pytest
|
|
import aiohttp
|
|
from typing import Optional
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_health_and_chat_completion():
|
|
"""
|
|
Test health endpoints and chat completion:
|
|
1. Check /health/readiness
|
|
2. Check /health/liveness
|
|
3. Make a chat completion call
|
|
"""
|
|
async with aiohttp.ClientSession() as session:
|
|
# Test readiness endpoint
|
|
async with session.get("http://0.0.0.0:4000/health/readiness") as response:
|
|
assert response.status == 200
|
|
readiness_response = await response.json()
|
|
assert readiness_response["status"] == "connected"
|
|
|
|
# Test liveness endpoint
|
|
async with session.get("http://0.0.0.0:4000/health/liveness") as response:
|
|
assert response.status == 200
|
|
liveness_response = await response.json()
|
|
print("liveness_response", liveness_response)
|
|
|
|
# Make a chat completion call
|
|
url = "http://0.0.0.0:4000/chat/completions"
|
|
headers = {
|
|
"Authorization": "Bearer sk-1234",
|
|
"Content-Type": "application/json",
|
|
}
|
|
data = {
|
|
"model": "gpt-4",
|
|
"messages": [
|
|
{"role": "system", "content": "You are a helpful assistant."},
|
|
{"role": "user", "content": "Hello!"},
|
|
],
|
|
}
|
|
|
|
async with session.post(url, headers=headers, json=data) as response:
|
|
assert response.status == 200
|
|
completion_response = await response.json()
|
|
assert "choices" in completion_response
|