[Frontend] Move api_server.py out openai folder (#52131)
Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
This commit is contained in:
@@ -1,570 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""Integration tests for shutdown behavior, timeout, and signal handling."""
|
||||
|
||||
import asyncio
|
||||
import signal
|
||||
import subprocess
|
||||
import sys
|
||||
import time
|
||||
from dataclasses import dataclass, field
|
||||
from pathlib import Path
|
||||
|
||||
import httpx
|
||||
import openai
|
||||
import psutil
|
||||
import pytest
|
||||
|
||||
from tests.utils import RemoteOpenAIServer
|
||||
from vllm.utils.network_utils import get_open_port
|
||||
|
||||
MODEL_NAME = "hmellor/tiny-random-LlamaForCausalLM"
|
||||
|
||||
# GPU initialization might take take longer
|
||||
_SERVER_STARTUP_TIMEOUT = 120
|
||||
_PROCESS_EXIT_TIMEOUT = 15
|
||||
_SHUTDOWN_DETECTION_TIMEOUT = 10
|
||||
_CHILD_CLEANUP_TIMEOUT = 10
|
||||
_INFLIGHT_REQUEST_START_TIMEOUT = 5
|
||||
_INFLIGHT_REQUEST_POLL_INTERVAL = 0.1
|
||||
_ABORT_CLIENT_TIMEOUT = 3
|
||||
|
||||
|
||||
def _get_child_pids(parent_pid: int) -> list[int]:
|
||||
try:
|
||||
parent = psutil.Process(parent_pid)
|
||||
return [c.pid for c in parent.children(recursive=True)]
|
||||
except psutil.NoSuchProcess:
|
||||
return []
|
||||
|
||||
|
||||
async def _assert_children_cleaned_up(
|
||||
child_pids: list[int],
|
||||
timeout: float = _CHILD_CLEANUP_TIMEOUT,
|
||||
):
|
||||
"""Wait for child processes to exit and fail if any remain."""
|
||||
if not child_pids:
|
||||
return
|
||||
|
||||
deadline = time.time() + timeout
|
||||
while time.time() < deadline:
|
||||
still_alive = []
|
||||
for pid in child_pids:
|
||||
try:
|
||||
p = psutil.Process(pid)
|
||||
if p.is_running() and p.status() != psutil.STATUS_ZOMBIE:
|
||||
still_alive.append(pid)
|
||||
except psutil.NoSuchProcess:
|
||||
pass
|
||||
if not still_alive:
|
||||
return
|
||||
await asyncio.sleep(0.5)
|
||||
|
||||
pytest.fail(
|
||||
f"Child processes {still_alive} still alive after {timeout}s. "
|
||||
f"Process cleanup may not be working correctly."
|
||||
)
|
||||
|
||||
|
||||
@dataclass
|
||||
class ShutdownState:
|
||||
got_503: bool = False
|
||||
got_500: bool = False
|
||||
requests_after_sigterm: int = 0
|
||||
aborted_requests: int = 0
|
||||
connection_errors: int = 0
|
||||
inflight_requests: int = 0
|
||||
stop_requesting: bool = False
|
||||
errors: list[str] = field(default_factory=list)
|
||||
|
||||
|
||||
async def _concurrent_request_loop(
|
||||
client: openai.AsyncOpenAI,
|
||||
state: ShutdownState,
|
||||
sigterm_sent: asyncio.Event | None = None,
|
||||
concurrency: int = 10,
|
||||
):
|
||||
"""Run multiple concurrent requests to keep the server busy."""
|
||||
|
||||
async def single_request():
|
||||
while not state.stop_requesting:
|
||||
try:
|
||||
state.inflight_requests += 1
|
||||
response = await client.completions.create(
|
||||
model=MODEL_NAME,
|
||||
prompt="Write a story: ",
|
||||
max_tokens=200,
|
||||
)
|
||||
if sigterm_sent is not None and sigterm_sent.is_set():
|
||||
state.requests_after_sigterm += 1
|
||||
# Check if any choice has finish_reason='abort'
|
||||
if any(choice.finish_reason == "abort" for choice in response.choices):
|
||||
state.aborted_requests += 1
|
||||
except openai.APIStatusError as e:
|
||||
if e.status_code == 503:
|
||||
state.got_503 = True
|
||||
elif e.status_code == 500:
|
||||
state.got_500 = True
|
||||
else:
|
||||
state.errors.append(f"API error: {e}")
|
||||
except (openai.APIConnectionError, httpx.RemoteProtocolError):
|
||||
state.connection_errors += 1
|
||||
if sigterm_sent is not None and sigterm_sent.is_set():
|
||||
break
|
||||
except Exception as e:
|
||||
state.errors.append(f"Unexpected error: {e}")
|
||||
break
|
||||
finally:
|
||||
state.inflight_requests -= 1
|
||||
await asyncio.sleep(0.01)
|
||||
|
||||
tasks = [asyncio.create_task(single_request()) for _ in range(concurrency)]
|
||||
try:
|
||||
await asyncio.gather(*tasks, return_exceptions=True)
|
||||
finally:
|
||||
for t in tasks:
|
||||
if not t.done():
|
||||
t.cancel()
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_shutdown_on_engine_failure(tmp_path: Path):
|
||||
"""Verify that API returns connection error when server process is killed.
|
||||
|
||||
Starts a vLLM server, kills it to simulate a crash, then verifies that
|
||||
subsequent API calls fail appropriately.
|
||||
"""
|
||||
|
||||
port = get_open_port()
|
||||
|
||||
# Redirect server output to a file rather than a pipe: nothing drains the
|
||||
# pipes while we poll for startup, so a server that writes more than the
|
||||
# pipe buffer holds would block on write and never become ready.
|
||||
server_log = tmp_path / "server.log"
|
||||
with server_log.open("wb") as log_file:
|
||||
proc = subprocess.Popen(
|
||||
[
|
||||
# dtype, max-len etc set so that this can run in CI
|
||||
sys.executable,
|
||||
"-m",
|
||||
"vllm.entrypoints.openai.api_server",
|
||||
"--model",
|
||||
MODEL_NAME,
|
||||
"--dtype",
|
||||
"bfloat16",
|
||||
"--max-model-len",
|
||||
"128",
|
||||
"--enforce-eager",
|
||||
"--port",
|
||||
str(port),
|
||||
"--gpu-memory-utilization",
|
||||
"0.05",
|
||||
"--max-num-seqs",
|
||||
"2",
|
||||
],
|
||||
stdout=log_file,
|
||||
stderr=subprocess.STDOUT,
|
||||
preexec_fn=lambda: signal.signal(signal.SIGINT, signal.SIG_IGN),
|
||||
)
|
||||
|
||||
# Wait for server startup
|
||||
start_time = time.time()
|
||||
client = openai.AsyncOpenAI(
|
||||
base_url=f"http://localhost:{port}/v1",
|
||||
api_key="dummy",
|
||||
max_retries=0,
|
||||
timeout=10,
|
||||
)
|
||||
|
||||
# Poll until server is ready
|
||||
while time.time() - start_time < _SERVER_STARTUP_TIMEOUT:
|
||||
try:
|
||||
await client.completions.create(
|
||||
model=MODEL_NAME, prompt="Hello", max_tokens=1
|
||||
)
|
||||
break
|
||||
except Exception:
|
||||
time.sleep(0.5)
|
||||
if proc.poll() is not None:
|
||||
pytest.fail(
|
||||
f"Server died during startup: {proc.returncode}\n"
|
||||
f"{server_log.read_text(errors='replace')}"
|
||||
)
|
||||
else:
|
||||
proc.terminate()
|
||||
proc.wait(timeout=_PROCESS_EXIT_TIMEOUT)
|
||||
pytest.fail(
|
||||
f"Server failed to start in {_SERVER_STARTUP_TIMEOUT} seconds\n"
|
||||
f"{server_log.read_text(errors='replace')}"
|
||||
)
|
||||
|
||||
# Kill server to simulate crash
|
||||
proc.terminate()
|
||||
time.sleep(1)
|
||||
|
||||
# Verify API calls now fail
|
||||
with pytest.raises((openai.APIConnectionError, openai.APIStatusError)):
|
||||
await client.completions.create(
|
||||
model=MODEL_NAME, prompt="This should fail", max_tokens=1
|
||||
)
|
||||
|
||||
return_code = proc.wait(timeout=_PROCESS_EXIT_TIMEOUT)
|
||||
assert return_code is not None
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_wait_timeout_completes_requests():
|
||||
"""Verify wait timeout: new requests rejected, in-flight requests complete."""
|
||||
server_args = [
|
||||
"--dtype",
|
||||
"bfloat16",
|
||||
"--max-model-len",
|
||||
"256",
|
||||
"--enforce-eager",
|
||||
"--gpu-memory-utilization",
|
||||
"0.05",
|
||||
"--max-num-seqs",
|
||||
"4",
|
||||
"--shutdown-timeout",
|
||||
"30",
|
||||
]
|
||||
|
||||
with RemoteOpenAIServer(MODEL_NAME, server_args) as remote_server:
|
||||
client = remote_server.get_async_client()
|
||||
proc = remote_server.proc
|
||||
child_pids = _get_child_pids(proc.pid)
|
||||
|
||||
state = ShutdownState()
|
||||
sigterm_sent = asyncio.Event()
|
||||
|
||||
request_task = asyncio.create_task(
|
||||
_concurrent_request_loop(client, state, sigterm_sent, concurrency=10)
|
||||
)
|
||||
|
||||
await asyncio.sleep(0.5)
|
||||
proc.send_signal(signal.SIGTERM)
|
||||
sigterm_sent.set()
|
||||
|
||||
try:
|
||||
await asyncio.wait_for(request_task, timeout=_SHUTDOWN_DETECTION_TIMEOUT)
|
||||
except asyncio.TimeoutError:
|
||||
pass
|
||||
finally:
|
||||
state.stop_requesting = True
|
||||
if not request_task.done():
|
||||
request_task.cancel()
|
||||
await asyncio.gather(request_task, return_exceptions=True)
|
||||
|
||||
# wait timeout should complete in-flight requests
|
||||
assert state.requests_after_sigterm > 0, (
|
||||
f"Wait timeout should complete in-flight requests. "
|
||||
f"503: {state.got_503}, 500: {state.got_500}, "
|
||||
f"conn_errors: {state.connection_errors}, errors: {state.errors}"
|
||||
)
|
||||
# server must stop accepting new requests (503, 500, or connection close)
|
||||
assert state.got_503 or state.got_500 or state.connection_errors > 0, (
|
||||
f"Server should stop accepting requests. "
|
||||
f"completed: {state.requests_after_sigterm}, errors: {state.errors}"
|
||||
)
|
||||
|
||||
await _assert_children_cleaned_up(child_pids)
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
@pytest.mark.parametrize("wait_for_engine_idle", [0.0, 2.0])
|
||||
async def test_abort_timeout_exits_quickly(wait_for_engine_idle: float):
|
||||
server_args = [
|
||||
"--dtype",
|
||||
"bfloat16",
|
||||
"--max-model-len",
|
||||
"256",
|
||||
"--enforce-eager",
|
||||
"--gpu-memory-utilization",
|
||||
"0.05",
|
||||
"--max-num-seqs",
|
||||
"4",
|
||||
"--shutdown-timeout",
|
||||
"0",
|
||||
]
|
||||
|
||||
with RemoteOpenAIServer(MODEL_NAME, server_args) as remote_server:
|
||||
proc = remote_server.proc
|
||||
child_pids = _get_child_pids(proc.pid)
|
||||
|
||||
if wait_for_engine_idle > 0:
|
||||
client = remote_server.get_async_client()
|
||||
# Send requests to ensure engine is fully initialized
|
||||
for _ in range(2):
|
||||
await client.completions.create(
|
||||
model=MODEL_NAME,
|
||||
prompt="Test request: ",
|
||||
max_tokens=10,
|
||||
)
|
||||
# Wait for engine to become idle
|
||||
await asyncio.sleep(wait_for_engine_idle)
|
||||
|
||||
start_time = time.time()
|
||||
proc.send_signal(signal.SIGTERM)
|
||||
|
||||
# abort timeout (0) should stop the server promptly.
|
||||
try:
|
||||
proc.wait(timeout=4.0)
|
||||
except subprocess.TimeoutExpired:
|
||||
proc.kill()
|
||||
proc.wait(timeout=5)
|
||||
pytest.fail("Process did not exit after SIGTERM with abort timeout")
|
||||
|
||||
exit_time = time.time() - start_time
|
||||
assert exit_time < 4.1, f"Default shutdown took too long: {exit_time:.1f}s"
|
||||
assert proc.returncode in (0, -15, None), f"Unexpected: {proc.returncode}"
|
||||
|
||||
await _assert_children_cleaned_up(child_pids)
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_wait_timeout_with_short_duration():
|
||||
"""Verify server exits cleanly with a short wait timeout."""
|
||||
wait_timeout = 3
|
||||
server_args = [
|
||||
"--dtype",
|
||||
"bfloat16",
|
||||
"--max-model-len",
|
||||
"256",
|
||||
"--enforce-eager",
|
||||
"--gpu-memory-utilization",
|
||||
"0.05",
|
||||
"--max-num-seqs",
|
||||
"4",
|
||||
"--shutdown-timeout",
|
||||
str(wait_timeout),
|
||||
]
|
||||
|
||||
with RemoteOpenAIServer(MODEL_NAME, server_args) as remote_server:
|
||||
client = remote_server.get_async_client()
|
||||
proc = remote_server.proc
|
||||
child_pids = _get_child_pids(proc.pid)
|
||||
|
||||
state = ShutdownState()
|
||||
request_task = asyncio.create_task(
|
||||
_concurrent_request_loop(client, state, concurrency=3)
|
||||
)
|
||||
|
||||
await asyncio.sleep(0.5)
|
||||
|
||||
start_time = time.time()
|
||||
proc.send_signal(signal.SIGTERM)
|
||||
|
||||
# server should exit within wait_timeout + buffer
|
||||
max_wait = wait_timeout + 15
|
||||
for _ in range(int(max_wait * 10)):
|
||||
if proc.poll() is not None:
|
||||
break
|
||||
time.sleep(0.1)
|
||||
|
||||
exit_time = time.time() - start_time
|
||||
|
||||
state.stop_requesting = True
|
||||
if not request_task.done():
|
||||
request_task.cancel()
|
||||
await asyncio.gather(request_task, return_exceptions=True)
|
||||
|
||||
if proc.poll() is None:
|
||||
proc.kill()
|
||||
proc.wait(timeout=5)
|
||||
pytest.fail(f"Process did not exit within {max_wait}s after SIGTERM")
|
||||
|
||||
assert exit_time < wait_timeout + 10, (
|
||||
f"Took too long to exit ({exit_time:.1f}s), expected <{wait_timeout + 10}s"
|
||||
)
|
||||
assert proc.returncode in (0, -15, None), f"Unexpected: {proc.returncode}"
|
||||
|
||||
await _assert_children_cleaned_up(child_pids)
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_abort_timeout_fails_inflight_requests():
|
||||
"""Verify abort timeout (0) immediately aborts in-flight requests."""
|
||||
server_args = [
|
||||
"--dtype",
|
||||
"bfloat16",
|
||||
"--max-model-len",
|
||||
"256",
|
||||
"--enforce-eager",
|
||||
"--gpu-memory-utilization",
|
||||
"0.05",
|
||||
"--max-num-seqs",
|
||||
"4",
|
||||
"--shutdown-timeout",
|
||||
"0",
|
||||
]
|
||||
|
||||
with RemoteOpenAIServer(MODEL_NAME, server_args) as remote_server:
|
||||
client = remote_server.get_async_client(timeout=_ABORT_CLIENT_TIMEOUT)
|
||||
proc = remote_server.proc
|
||||
child_pids = _get_child_pids(proc.pid)
|
||||
|
||||
state = ShutdownState()
|
||||
sigterm_sent = asyncio.Event()
|
||||
|
||||
request_task = asyncio.create_task(
|
||||
_concurrent_request_loop(client, state, sigterm_sent, concurrency=10)
|
||||
)
|
||||
|
||||
deadline = time.time() + _INFLIGHT_REQUEST_START_TIMEOUT
|
||||
while state.inflight_requests == 0 and time.time() < deadline:
|
||||
await asyncio.sleep(_INFLIGHT_REQUEST_POLL_INTERVAL)
|
||||
assert state.inflight_requests > 0
|
||||
|
||||
proc.send_signal(signal.SIGTERM)
|
||||
sigterm_sent.set()
|
||||
|
||||
try:
|
||||
await asyncio.wait_for(request_task, timeout=5)
|
||||
except asyncio.TimeoutError:
|
||||
pass
|
||||
finally:
|
||||
state.stop_requesting = True
|
||||
if not request_task.done():
|
||||
request_task.cancel()
|
||||
await asyncio.gather(request_task, return_exceptions=True)
|
||||
|
||||
# With abort timeout (0), requests should be aborted (finish_reason='abort')
|
||||
# or rejected (connection errors or API errors)
|
||||
assert (
|
||||
state.aborted_requests > 0
|
||||
or state.connection_errors > 0
|
||||
or state.got_500
|
||||
or state.got_503
|
||||
), (
|
||||
f"Abort timeout should cause request aborts or failures. "
|
||||
f"aborted: {state.aborted_requests}, "
|
||||
f"503: {state.got_503}, 500: {state.got_500}, "
|
||||
f"conn_errors: {state.connection_errors}, "
|
||||
f"completed: {state.requests_after_sigterm}"
|
||||
)
|
||||
|
||||
# Verify fast shutdown
|
||||
start_time = time.time()
|
||||
for _ in range(100):
|
||||
if proc.poll() is not None:
|
||||
break
|
||||
time.sleep(0.1)
|
||||
|
||||
exit_time = time.time() - start_time
|
||||
assert exit_time < 10, f"Abort timeout shutdown took too long: {exit_time:.1f}s"
|
||||
|
||||
await _assert_children_cleaned_up(child_pids)
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_request_rejection_during_shutdown():
|
||||
"""Verify new requests are rejected with error during shutdown."""
|
||||
server_args = [
|
||||
"--dtype",
|
||||
"bfloat16",
|
||||
"--max-model-len",
|
||||
"256",
|
||||
"--enforce-eager",
|
||||
"--gpu-memory-utilization",
|
||||
"0.05",
|
||||
"--max-num-seqs",
|
||||
"4",
|
||||
"--shutdown-timeout",
|
||||
"30",
|
||||
]
|
||||
|
||||
with RemoteOpenAIServer(MODEL_NAME, server_args) as remote_server:
|
||||
client = remote_server.get_async_client()
|
||||
proc = remote_server.proc
|
||||
child_pids = _get_child_pids(proc.pid)
|
||||
|
||||
proc.send_signal(signal.SIGTERM)
|
||||
|
||||
await asyncio.sleep(1.0)
|
||||
|
||||
# Try to send new requests - they should be rejected
|
||||
rejected_count = 0
|
||||
for _ in range(10):
|
||||
try:
|
||||
await client.completions.create(
|
||||
model=MODEL_NAME, prompt="Hello", max_tokens=10
|
||||
)
|
||||
except (
|
||||
openai.APIStatusError,
|
||||
openai.APIConnectionError,
|
||||
httpx.RemoteProtocolError,
|
||||
):
|
||||
rejected_count += 1
|
||||
await asyncio.sleep(0.1)
|
||||
|
||||
assert rejected_count > 0, (
|
||||
f"Expected requests to be rejected during shutdown, "
|
||||
f"but {rejected_count} were rejected out of 10"
|
||||
)
|
||||
|
||||
await _assert_children_cleaned_up(child_pids)
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_multi_api_server_shutdown():
|
||||
"""Verify shutdown works with multiple API servers."""
|
||||
server_args = [
|
||||
"--dtype",
|
||||
"bfloat16",
|
||||
"--max-model-len",
|
||||
"256",
|
||||
"--enforce-eager",
|
||||
"--gpu-memory-utilization",
|
||||
"0.05",
|
||||
"--max-num-seqs",
|
||||
"4",
|
||||
"--shutdown-timeout",
|
||||
"30",
|
||||
"--api-server-count",
|
||||
"2",
|
||||
]
|
||||
|
||||
with RemoteOpenAIServer(MODEL_NAME, server_args, auto_port=True) as remote_server:
|
||||
client = remote_server.get_async_client()
|
||||
proc = remote_server.proc
|
||||
child_pids = _get_child_pids(proc.pid)
|
||||
|
||||
assert len(child_pids) >= 2, (
|
||||
f"Expected at least 2 child processes, got {len(child_pids)}"
|
||||
)
|
||||
|
||||
state = ShutdownState()
|
||||
sigterm_sent = asyncio.Event()
|
||||
|
||||
# Start concurrent requests across both API servers
|
||||
request_task = asyncio.create_task(
|
||||
_concurrent_request_loop(client, state, sigterm_sent, concurrency=8)
|
||||
)
|
||||
|
||||
await asyncio.sleep(0.5)
|
||||
|
||||
# Send SIGTERM to parent - should propagate to all children
|
||||
proc.send_signal(signal.SIGTERM)
|
||||
sigterm_sent.set()
|
||||
|
||||
try:
|
||||
await asyncio.wait_for(request_task, timeout=_SHUTDOWN_DETECTION_TIMEOUT)
|
||||
except asyncio.TimeoutError:
|
||||
pass
|
||||
finally:
|
||||
state.stop_requesting = True
|
||||
if not request_task.done():
|
||||
request_task.cancel()
|
||||
await asyncio.gather(request_task, return_exceptions=True)
|
||||
|
||||
for _ in range(300): # up to 30 seconds
|
||||
if proc.poll() is not None:
|
||||
break
|
||||
time.sleep(0.1)
|
||||
|
||||
if proc.poll() is None:
|
||||
proc.kill()
|
||||
proc.wait(timeout=5)
|
||||
pytest.fail("Process did not exit after SIGTERM")
|
||||
|
||||
await _assert_children_cleaned_up(child_pids)
|
||||
@@ -1,175 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
import asyncio
|
||||
import os
|
||||
|
||||
import openai # use the official client for correctness check
|
||||
import pytest
|
||||
import pytest_asyncio
|
||||
|
||||
from tests.utils import RemoteOpenAIServer
|
||||
from tests.v1.utils import check_request_balancing
|
||||
|
||||
MODEL_NAME = "hmellor/tiny-random-LlamaForCausalLM"
|
||||
|
||||
DP_SIZE = os.getenv("DP_SIZE", "1")
|
||||
|
||||
|
||||
@pytest.fixture(scope="module")
|
||||
def default_server_args():
|
||||
return [
|
||||
# use half precision for speed and memory savings in CI environment
|
||||
"--dtype",
|
||||
"bfloat16",
|
||||
"--max-model-len",
|
||||
"2048",
|
||||
"--max-num-seqs",
|
||||
"128",
|
||||
"--enforce-eager",
|
||||
"--api-server-count",
|
||||
"4",
|
||||
"--data_parallel_size",
|
||||
DP_SIZE,
|
||||
]
|
||||
|
||||
|
||||
@pytest.fixture(scope="module")
|
||||
def server(default_server_args):
|
||||
with RemoteOpenAIServer(MODEL_NAME, default_server_args) as remote_server:
|
||||
yield remote_server
|
||||
|
||||
|
||||
@pytest_asyncio.fixture
|
||||
async def client(server):
|
||||
async with server.get_async_client() as async_client:
|
||||
yield async_client
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
@pytest.mark.parametrize(
|
||||
"model_name",
|
||||
[MODEL_NAME],
|
||||
)
|
||||
async def test_single_completion(
|
||||
client: openai.AsyncOpenAI, server: RemoteOpenAIServer, model_name: str
|
||||
) -> None:
|
||||
async def make_request():
|
||||
completion = await client.completions.create(
|
||||
model=model_name, prompt="Hello, my name is", max_tokens=10, temperature=1.0
|
||||
)
|
||||
|
||||
assert completion.id is not None
|
||||
assert completion.choices is not None and len(completion.choices) == 1
|
||||
|
||||
choice = completion.choices[0]
|
||||
# The exact number of tokens can vary slightly with temperature=1.0,
|
||||
# so we check for a reasonable minimum length.
|
||||
assert len(choice.text) >= 1
|
||||
# Finish reason might not always be 'length' if the model finishes early
|
||||
# or due to other reasons, especially with high temperature.
|
||||
# So, we'll accept 'length' or 'stop'.
|
||||
assert choice.finish_reason in ("length", "stop")
|
||||
|
||||
# Token counts can also vary, so we check they are positive.
|
||||
assert completion.usage.completion_tokens > 0
|
||||
assert completion.usage.prompt_tokens > 0
|
||||
assert completion.usage.total_tokens > 0
|
||||
return completion
|
||||
|
||||
# Test single request
|
||||
result = await make_request()
|
||||
assert result is not None
|
||||
|
||||
await asyncio.sleep(0.5)
|
||||
|
||||
# Send two bursts of requests
|
||||
num_requests = 100
|
||||
tasks = [make_request() for _ in range(num_requests)]
|
||||
results = await asyncio.gather(*tasks)
|
||||
assert len(results) == num_requests
|
||||
assert all(completion is not None for completion in results)
|
||||
|
||||
await asyncio.sleep(0.5)
|
||||
|
||||
tasks = [make_request() for _ in range(num_requests)]
|
||||
results = await asyncio.gather(*tasks)
|
||||
assert len(results) == num_requests
|
||||
assert all(completion is not None for completion in results)
|
||||
|
||||
# Check request balancing via Prometheus metrics if DP_SIZE > 1
|
||||
check_request_balancing(server, int(DP_SIZE))
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
@pytest.mark.parametrize(
|
||||
"model_name",
|
||||
[MODEL_NAME],
|
||||
)
|
||||
async def test_completion_streaming(
|
||||
client: openai.AsyncOpenAI, server: RemoteOpenAIServer, model_name: str
|
||||
) -> None:
|
||||
prompt = "What is an LLM?"
|
||||
|
||||
async def make_streaming_request():
|
||||
# Perform a non-streaming request to get the expected full output
|
||||
single_completion = await client.completions.create(
|
||||
model=model_name,
|
||||
prompt=prompt,
|
||||
max_tokens=5,
|
||||
temperature=0.0,
|
||||
)
|
||||
single_output = single_completion.choices[0].text
|
||||
|
||||
# Perform the streaming request
|
||||
stream = await client.completions.create(
|
||||
model=model_name, prompt=prompt, max_tokens=5, temperature=0.0, stream=True
|
||||
)
|
||||
chunks: list[str] = []
|
||||
finish_reason_count = 0
|
||||
last_chunk = None
|
||||
async for chunk in stream:
|
||||
chunks.append(chunk.choices[0].text)
|
||||
if chunk.choices[0].finish_reason is not None:
|
||||
finish_reason_count += 1
|
||||
last_chunk = chunk # Keep track of the last chunk
|
||||
|
||||
# finish reason should only return in the last block for OpenAI API
|
||||
assert finish_reason_count == 1, "Finish reason should appear exactly once."
|
||||
assert last_chunk is not None, "Stream should have yielded at least one chunk."
|
||||
assert last_chunk.choices[0].finish_reason == "length", (
|
||||
"Finish reason should be 'length'."
|
||||
)
|
||||
# Check that the combined text matches the non-streamed version.
|
||||
assert "".join(chunks) == single_output, (
|
||||
"Streamed output should match non-streamed output."
|
||||
)
|
||||
return True # Indicate success for this request
|
||||
|
||||
# Test single request
|
||||
result = await make_streaming_request()
|
||||
assert result is not None
|
||||
|
||||
await asyncio.sleep(0.5)
|
||||
|
||||
# Send two bursts of requests
|
||||
num_requests = 100
|
||||
tasks = [make_streaming_request() for _ in range(num_requests)]
|
||||
results = await asyncio.gather(*tasks)
|
||||
|
||||
assert len(results) == num_requests, (
|
||||
f"Expected {num_requests} results, got {len(results)}"
|
||||
)
|
||||
assert all(results), "Not all streaming requests completed successfully."
|
||||
|
||||
await asyncio.sleep(0.5)
|
||||
|
||||
tasks = [make_streaming_request() for _ in range(num_requests)]
|
||||
results = await asyncio.gather(*tasks)
|
||||
|
||||
assert len(results) == num_requests, (
|
||||
f"Expected {num_requests} results, got {len(results)}"
|
||||
)
|
||||
assert all(results), "Not all streaming requests completed successfully."
|
||||
|
||||
# Check request balancing via Prometheus metrics if DP_SIZE > 1
|
||||
check_request_balancing(server, int(DP_SIZE))
|
||||
@@ -1,47 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
from tempfile import TemporaryDirectory
|
||||
|
||||
import httpx
|
||||
import pytest
|
||||
|
||||
from vllm import envs
|
||||
from vllm.version import __version__ as VLLM_VERSION
|
||||
|
||||
from ...utils import RemoteOpenAIServer
|
||||
|
||||
MODEL_NAME = "Qwen/Qwen3-0.6B"
|
||||
|
||||
|
||||
@pytest.fixture(scope="module")
|
||||
def server():
|
||||
with TemporaryDirectory() as tmpdir:
|
||||
args = [
|
||||
# use half precision for speed and memory savings in CI environment
|
||||
"--dtype",
|
||||
"bfloat16",
|
||||
"--max-model-len",
|
||||
"8192",
|
||||
"--enforce-eager",
|
||||
"--max-num-seqs",
|
||||
"128",
|
||||
"--uds",
|
||||
f"{tmpdir}/vllm.sock",
|
||||
]
|
||||
|
||||
with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
|
||||
yield remote_server
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_show_version(server: RemoteOpenAIServer):
|
||||
transport = httpx.HTTPTransport(uds=server.uds)
|
||||
client = httpx.Client(transport=transport)
|
||||
response = client.get(server.url_for("version"))
|
||||
response.raise_for_status()
|
||||
|
||||
payload = response.json()
|
||||
assert payload["version"] == VLLM_VERSION
|
||||
if envs.VLLM_USE_RUST_FRONTEND:
|
||||
assert payload["rust_frontend_version"] == VLLM_VERSION
|
||||
Reference in New Issue
Block a user