chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 14 Packages Projects Releases Wiki Activity

[CI] feat: add early exit to wait_for_server when process dies (#18602)

Browse Source
This commit is contained in:
shuwenn
2026-02-13 16:46:09 -08:00
committed by GitHub
parent eccf875d49
commit 3299c4f9c1
33 changed files with 229 additions and 179 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/manual/hicache/test_disaggregation_hicache.py
+2 -2
View File
@@ -33,8 +33,8 @@ class DisaggregationHiCacheBase(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
test/manual/piecewise_cudagraph/test_disaggregation_piecewise_cuda_graph.py
+2 -2
View File
@@ -25,8 +25,8 @@ class TestDisaggregationPiecewiseCudaGraph(PDDisaggregationServerBase):
cls.start_decode()
# Wait for both to be ready
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
test/manual/test_mori_transfer_engine_e2e.py
+18 -6
View File
@@ -52,10 +52,14 @@ class TestMoriTransferEngineE2E(PDDisaggregationServerBase):
cls.start_decode()
cls.wait_server_ready(
cls.prefill_url + "/health", timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
cls.prefill_url + "/health",
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
process=cls.process_prefill,
)
cls.wait_server_ready(
cls.decode_url + "/health", timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
cls.decode_url + "/health",
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
process=cls.process_decode,
)
cls.launch_lb()
@@ -85,7 +89,9 @@ class TestMoriTransferEngineE2E(PDDisaggregationServerBase):
print("Starting load balancer:", " ".join(lb_command))
cls.process_lb = subprocess.Popen(lb_command, stdout=None, stderr=None)
cls.wait_server_ready(
cls.lb_url + "/health", timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
cls.lb_url + "/health",
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
process=cls.process_lb,
)
@classmethod
@@ -190,10 +196,14 @@ class TestMoriTransferEngineTPMismatchE2E(PDDisaggregationServerBase):
cls.start_decode()
cls.wait_server_ready(
cls.prefill_url + "/health", timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
cls.prefill_url + "/health",
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
process=cls.process_prefill,
)
cls.wait_server_ready(
cls.decode_url + "/health", timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
cls.decode_url + "/health",
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
process=cls.process_decode,
)
cls.launch_lb()
@@ -222,7 +232,9 @@ class TestMoriTransferEngineTPMismatchE2E(PDDisaggregationServerBase):
print("Starting load balancer:", " ".join(lb_command))
cls.process_lb = subprocess.Popen(lb_command, stdout=None, stderr=None)
cls.wait_server_ready(
cls.lb_url + "/health", timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
cls.lb_url + "/health",
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
process=cls.process_lb,
)
@classmethod
test/registered/amd/disaggregation/test_disaggregation_basic.py
+6 -6
View File
@@ -44,8 +44,8 @@ class TestDisaggregationAccuracy(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
@@ -238,8 +238,8 @@ class TestDisaggregationMooncakeFailure(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
@@ -346,8 +346,8 @@ class TestDisaggregationSimulatedRetract(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
test/registered/amd/disaggregation/test_disaggregation_pp.py
+6 -6
View File
@@ -40,8 +40,8 @@ class TestDisaggregationPrefillPPAccuracy(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
@@ -129,8 +129,8 @@ class TestDisaggregationPrefillPPDynamicChunkAccuracy(PDDisaggregationServerBase
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
@@ -219,8 +219,8 @@ class TestDisaggregationDecodePPAccuracy(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
test/registered/disaggregation/test_disaggregation_basic.py
+8 -8
View File
@@ -34,8 +34,8 @@ class TestDisaggregationAccuracy(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
@@ -204,8 +204,8 @@ class TestDisaggregationMooncakeFailure(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
@@ -306,8 +306,8 @@ class TestDisaggregationMooncakeSpec(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
@@ -375,8 +375,8 @@ class TestDisaggregationSimulatedRetract(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
test/registered/distributed/test_disaggregation_aarch64.py
+2 -2
View File
@@ -29,8 +29,8 @@ class TestDisaggregationMooncakeAARCH64Accuracy(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
test/registered/distributed/test_disaggregation_different_tp.py
+8 -8
View File
@@ -32,8 +32,8 @@ class TestDisaggregationMooncakePrefillLargerTP(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
@@ -103,8 +103,8 @@ class TestDisaggregationMooncakeDecodeLargerTP(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
@@ -174,8 +174,8 @@ class TestDisaggregationMooncakeMHAPrefillLargerTP(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
@@ -245,8 +245,8 @@ class TestDisaggregationMooncakeMHADecodeLargerTP(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
test/registered/distributed/test_disaggregation_dp_attention.py
+2 -2
View File
@@ -34,8 +34,8 @@ class TestDisaggregationDPAttention(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
test/registered/distributed/test_disaggregation_hybrid_attention.py
+6 -6
View File
@@ -29,8 +29,8 @@ class TestDisaggregationHybridAttentionMamba(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
@@ -97,8 +97,8 @@ class TestDisaggregationHybridAttentionMambaExtraBuffer(PDDisaggregationServerBa
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
@@ -175,8 +175,8 @@ class TestDisaggregationHybridAttentionMambaDPDecode(PDDisaggregationServerBase)
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
test/registered/distributed/test_disaggregation_pp.py
+6 -6
View File
@@ -28,8 +28,8 @@ class TestDisaggregationPrefillPPAccuracy(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
@@ -101,8 +101,8 @@ class TestDisaggregationPrefillPPDynamicChunkAccuracy(PDDisaggregationServerBase
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
@@ -175,8 +175,8 @@ class TestDisaggregationDecodePPAccuracy(PDDisaggregationServerBase):
cls.start_decode()
# Block until both
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
test/registered/distributed/test_epd_disaggregation.py
+7 -7
View File
@@ -44,9 +44,9 @@ class TestEPDDisaggregationOneEncoder(PDDisaggregationServerBase):
decode_thread.join()
# Wait for all servers to be ready
cls.wait_server_ready(cls.encode_url + "/health")
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.encode_url + "/health", process=cls.process_encode)
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
@@ -249,10 +249,10 @@ class TestEPDDisaggregationMultiEncoders(PDDisaggregationServerBase):
prefill_thread.join()
decode_thread.join()
cls.wait_server_ready(cls.encode_url1 + "/health")
cls.wait_server_ready(cls.encode_url2 + "/health")
cls.wait_server_ready(cls.prefill_url + "/health")
cls.wait_server_ready(cls.decode_url + "/health")
cls.wait_server_ready(cls.encode_url1 + "/health", process=cls.process_encode1)
cls.wait_server_ready(cls.encode_url2 + "/health", process=cls.process_encode2)
cls.wait_server_ready(cls.prefill_url + "/health", process=cls.process_prefill)
cls.wait_server_ready(cls.decode_url + "/health", process=cls.process_decode)
cls.launch_lb()
test/registered/hicache/test_hicache_storage_file_backend.py
+9 -12
View File
@@ -29,6 +29,7 @@ from sglang.test.test_utils import (
is_in_ci,
popen_launch_server,
)
from sglang.utils import wait_for_http_ready
register_cuda_ci(est_time=200, suite="stage-b-test-large-2-gpu")
register_amd_ci(est_time=526, suite="stage-b-test-large-2-gpu-amd")
@@ -53,7 +54,7 @@ class HiCacheStorageBaseMixin:
# Launch server with HiCache enabled and cache report
cls.process = cls._launch_server_with_hicache()
cls._wait_for_server_ready()
cls._wait_for_server_ready(process=cls.process)
print(f"Test server launched successfully at {cls.base_url}")
print(f"Cache directory: {cls.temp_dir}")
@@ -128,18 +129,14 @@ class HiCacheStorageBaseMixin:
)
@classmethod
def _wait_for_server_ready(cls, timeout: int = 60) -> bool:
def _wait_for_server_ready(cls, timeout: int = 60, process=None) -> bool:
"""Wait for server to be ready"""
start_time = time.time()
while time.time() - start_time < timeout:
try:
response = requests.get(f"{cls.base_url}/health", timeout=5)
if response.status_code == 200:
return True
except requests.RequestException:
pass
time.sleep(2)
raise TimeoutError("Server failed to start within timeout")
wait_for_http_ready(
url=f"{cls.base_url}/health",
timeout=timeout,
process=process,
)
return True
def send_request(
self, prompt: str, max_tokens: int = 100, temperature: float = 0.0
test/registered/hicache/test_hicache_storage_runtime_attach_detach.py
+12 -13
View File
@@ -26,6 +26,7 @@ from sglang.test.test_utils import (
find_available_port,
popen_launch_server,
)
from sglang.utils import wait_for_http_ready
register_cuda_ci(est_time=200, suite="stage-b-test-large-2-gpu")
@@ -68,17 +69,15 @@ class TestHiCacheStorageRuntimeAttachDetach(CustomTestCase):
shutil.rmtree(cls.temp_dir, ignore_errors=True)
@classmethod
def _wait_for_server_ready(cls, base_url: str, timeout: int = 60) -> bool:
start_time = time.time()
while time.time() - start_time < timeout:
try:
code, _body = cls._http_get(f"{base_url}/health", timeout=5)
if code == 200:
return True
except Exception:
pass
time.sleep(2)
raise TimeoutError("Server failed to start within timeout")
def _wait_for_server_ready(
cls, base_url: str, timeout: int = 60, process=None
) -> bool:
wait_for_http_ready(
url=f"{base_url}/health",
timeout=timeout,
process=process,
)
return True
@staticmethod
def _http_get(url: str, timeout: int = 10, headers: dict | None = None):
@@ -204,7 +203,7 @@ class TestHiCacheStorageRuntimeAttachDetach(CustomTestCase):
env=self.env,
)
try:
self._wait_for_server_ready(self.base_url)
self._wait_for_server_ready(self.base_url, process=process1)
code_info, _body_info = self._http_get(
f"{self.base_url}/hicache/storage-backend", timeout=10
@@ -234,7 +233,7 @@ class TestHiCacheStorageRuntimeAttachDetach(CustomTestCase):
env=self.env,
)
try:
self._wait_for_server_ready(base_url2)
self._wait_for_server_ready(base_url2, process=process2)
# 1) Initially disabled (but unauthorized without admin key)
code_info2_unauth, _ = self._http_get(
test/srt/experiment_runner.py
+19 -14
View File
@@ -11,9 +11,10 @@ from datetime import datetime
from typing import List, Optional, Tuple
import psutil
import requests
import yaml
from sglang.utils import wait_for_http_ready
@dataclass
class ServerConfig:
@@ -182,18 +183,20 @@ class ExperimentRunner:
self.process_manager = ProcessManager()
self.logger = logging.getLogger(__name__)
def wait_for_server(self, port: int, timeout: int = 300) -> bool:
start_time = time.perf_counter()
while time.perf_counter() - start_time < timeout:
try:
response = requests.get(f"http://localhost:{port}/health")
if response.status_code == 200:
self.logger.debug(f"Server ready on port {port}")
return True
except requests.RequestException:
time.sleep(2)
return False
def wait_for_server(
self, port: int, timeout: int = 300, process: Optional[subprocess.Popen] = None
) -> bool:
try:
wait_for_http_ready(
url=f"http://localhost:{port}/health",
timeout=timeout,
process=process,
)
self.logger.debug(f"Server ready on port {port}")
return True
except (RuntimeError, TimeoutError) as e:
self.logger.error("Server failed to become ready: %s", e)
return False
def run_task(self, config: TaskConfig) -> TaskResult:
start_time = time.perf_counter()
@@ -216,7 +219,9 @@ class ExperimentRunner:
self.process_manager.start_process(config.server_cmd, "SERVER")
)
if not self.wait_for_server(port):
if not self.wait_for_server(
port, process=self.process_manager.server_process
):
raise TimeoutError("Server startup timeout")
time.sleep(10)