fix: preserve disconnect events in api key middleware (#17253)

This commit is contained in:
shuwenn
2026-01-27 11:48:24 +08:00
committed by GitHub
parent fd3b179ffd
commit 57e432d951
3 changed files with 94 additions and 24 deletions

View File

@@ -155,31 +155,54 @@ def add_api_key_middleware(
"""Add middleware for three endpoint auth levels: normal/admin_optional/admin_force."""
# Import lazily so `decide_request_auth()` can be unit-tested without FastAPI installed.
from fastapi.responses import ORJSONResponse
from starlette.requests import Request
@app.middleware("http")
async def authentication(request, call_next):
path = request.url.path
authz = request.headers.get("Authorization")
level = _get_auth_level_from_app_and_scope(request.app, request.scope)
decision = decide_request_auth(
method=request.method,
path=path,
authorization_header=authz,
api_key=api_key,
admin_api_key=admin_api_key,
auth_level=level,
)
class _ApiKeyASGIMiddleware:
"""ASGI-native middleware to preserve client disconnect events."""
if not decision.allowed:
return ORJSONResponse(
content={
"error": (
"Unauthorized"
if decision.error_status_code == 401
else "Forbidden"
)
},
status_code=decision.error_status_code,
def __init__(self, app, *, api_key, admin_api_key, fastapi_app):
self.app = app
self.api_key = api_key
self.admin_api_key = admin_api_key
self.fastapi_app = fastapi_app
async def __call__(self, scope, receive, send):
if scope["type"] != "http":
await self.app(scope, receive, send)
return
request = Request(scope, receive=receive)
path = request.url.path
authz = request.headers.get("Authorization")
level = _get_auth_level_from_app_and_scope(self.fastapi_app, scope)
decision = decide_request_auth(
method=request.method,
path=path,
authorization_header=authz,
api_key=self.api_key,
admin_api_key=self.admin_api_key,
auth_level=level,
)
return await call_next(request)
if not decision.allowed:
response = ORJSONResponse(
content={
"error": (
"Unauthorized"
if decision.error_status_code == 401
else "Forbidden"
)
},
status_code=decision.error_status_code,
)
await response(scope, receive, send)
return
await self.app(scope, receive, send)
app.add_middleware(
_ApiKeyASGIMiddleware,
api_key=api_key,
admin_api_key=admin_api_key,
fastapi_app=app,
)

View File

@@ -1662,6 +1662,7 @@ def run_and_check_memory_leak(
disable_overlap,
chunked_prefill_size,
assert_has_abort,
api_key: Optional[str] = None,
):
other_args = [
"--chunked-prefill-size",
@@ -1689,6 +1690,7 @@ def run_and_check_memory_leak(
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
return_stdout_stderr=(stdout, stderr),
api_key=api_key,
)
# Launch a thread to stream the output

View File

@@ -63,6 +63,51 @@ class TestAbort(CustomTestCase):
)
class TestAbortWithApiKey(CustomTestCase):
def workload_func(self, base_url, model, api_key: str):
def process_func():
def run_one(_):
prompt = """
System: You are a helpful assistant.
User: What is the capital of France?
Assistant: The capital of France is
"""
response = requests.post(
f"{base_url}/generate",
json={
"text": prompt,
"sampling_params": {
"temperature": 0,
"max_new_tokens": 2048,
},
},
headers={"Authorization": f"Bearer {api_key}"},
)
response.json()
with ThreadPoolExecutor(16) as executor:
list(executor.map(run_one, list(range(16))))
p = multiprocessing.Process(target=process_func)
p.start()
time.sleep(0.5)
p.terminate()
time.sleep(10)
def test_memory_leak_with_api_key(self):
api_key = "test-api-key"
run_and_check_memory_leak(
lambda base_url, model: self.workload_func(base_url, model, api_key),
disable_radix_cache=False,
enable_mixed_chunk=False,
disable_overlap=False,
chunked_prefill_size=8192,
assert_has_abort=True,
api_key=api_key,
)
class TestAbortAll(CustomTestCase):
@classmethod
def setUpClass(cls):