test_serve_cli.py

# SPDX-License-Identifier: Apache-2.0
import subprocess

import pytest
import os

from ..utils import RemoteOpenAIServer, models_path_prefix

MODEL_NAME = os.path.join(models_path_prefix, "meta-llama/Llama-3.2-1B-Instruct")


@pytest.fixture(scope="module")
def server():
    args = [
        "--max-model-len", "1024", "--enforce-eager", "--load-format", "dummy"
    ]

    with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
        yield remote_server


@pytest.mark.benchmark
def test_bench_serve(server):
    command = [
        "vllm",
        "bench",
        "serve",
        "--model",
        MODEL_NAME,
        "--host",
        server.host,
        "--port",
        str(server.port),
        "--random-input-len",
        "32",
        "--random-output-len",
        "4",
        "--num-prompts",
        "5",
    ]
    result = subprocess.run(command, capture_output=True, text=True)
    print(result.stdout)
    print(result.stderr)

    assert result.returncode == 0, f"Benchmark failed: {result.stderr}"