"""Real local Gradio HTTP/API round trip with the tiny architecture, CPU only.""" import json from pathlib import Path import torch from gradio_client import Client import app from engine import generate, POLICIES from test_engine import Tokenizer from test_model import tiny_model def test_http_generate_and_download(monkeypatch): monkeypatch.setenv("NO_PROXY", "localhost,127.0.0.1,::1") monkeypatch.setenv("no_proxy", "localhost,127.0.0.1,::1") monkeypatch.setenv("GRADIO_ANALYTICS_ENABLED", "False") # All application callbacks, serialization, queueing and file download remain real. # Only the inference device is changed; no 9B checkpoint is loaded. monkeypatch.setattr(app, "generate", lambda m, t, ids, settings: generate(m, t, ids, settings, device="cpu")) torch.set_num_threads(1) demo = app.create_demo(tiny_model(), Tokenizer(), {"test_only": True}) try: _, url, _ = demo.queue().launch(server_name="127.0.0.1", prevent_thread_lock=True, quiet=True) client = Client(url, verbose=False) response, metrics, table, report = client.predict("hello", "", 17, 16, 2, POLICIES[0], 112, True, api_name="/generate") assert isinstance(response, str) assert 0 < metrics["output_tokens"] <= 17 assert len(table["data"]) >= 2 saved = json.loads(Path(report).read_text()) assert saved["prompt"] == "hello" assert saved["environment"]["test_only"] is True assert saved["settings"]["max_tokens"] == 17 finally: demo.close()