tests/test_switch_models.py

import os
os.environ["MOCK_INFERENCE"] = "true"

from fastapi.testclient import TestClient
from app import app
import model_service

client = TestClient(app)

def test_switch_models():
    # Load model A
    response = client.post("/v1/models/load", json={"model": "plain-qwen-0.5b"})
    assert response.status_code == 200
    assert model_service.active_model_id == "plain-qwen-0.5b"
    
    # Load model B
    response = client.post("/v1/models/load", json={"model": "plain-llama-3.2-3b"})
    assert response.status_code == 200
    assert model_service.active_model_id == "plain-llama-3.2-3b"
    
    # Reset and unload
    response = client.post("/v1/reset", json={"unload_model": True})
    assert response.status_code == 200
    assert model_service.active_model_id is None