minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[CI][Disaggregation] Fix EADDRINUSE flake in test_disaggregation_dwdp_gpt_oss (#39611)

Browse Source 
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Shangming Cai
2026-09-15 23:05:05 +08:00
committed by GitHub
co-authored by Claude Opus 5
parent 1aeeb25e86
commit 47a157f257
1 changed files with 25 additions and 63 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/disaggregation/test_disaggregation_dwdp_gpt_oss.py
+17 -55
View File
@@ -6,10 +6,6 @@ from sglang.test.run_eval import run_eval
from sglang.test.server_fixtures.disaggregation_fixture import (
PDDisaggregationServerBase,
)
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
popen_launch_pd_server,
)
register_cuda_ci(est_time=170, stage="extra-b", runner_config="4-gpu-b200")
@@ -23,69 +19,35 @@ class TestDisaggregationDWDPGptOss(PDDisaggregationServerBase):
NUM_PREFILL_GPUS = 2
NUM_DECODE_GPUS = 2
@classmethod
def setUpClass(cls):
super().setUpClass()
cls.model = GPT_OSS_MODEL_PATH
# Drive the base fixture's launchers rather than reimplementing them, so
# both sides get the pinned --nccl-port. Deriving it from get_free_port()
# on each side races onto the same port and dies at init_process_group
# with EADDRINUSE.
prefill_tp_size = NUM_PREFILL_GPUS
decode_tp_size = NUM_DECODE_GPUS
decode_base_gpu_id = NUM_PREFILL_GPUS
cls.start_prefill()
cls.start_decode()
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()
@classmethod
def start_prefill(cls):
prefill_args = [
"--trust-remote-code",
"--disaggregation-mode",
"prefill",
"--disaggregation-bootstrap-port",
cls.bootstrap_port,
"--tp",
str(cls.NUM_PREFILL_GPUS),
extra_prefill_args = [
"--dwdp-size",
str(cls.NUM_PREFILL_GPUS),
str(NUM_PREFILL_GPUS),
"--disable-flashinfer-autotune",
"--mem-fraction-static",
"0.85",
]
prefill_args += cls.transfer_backend + cls.rdma_devices
cls.process_prefill = popen_launch_pd_server(
cls.model,
cls.prefill_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=prefill_args,
)
@classmethod
def start_decode(cls):
decode_args = [
"--trust-remote-code",
"--disaggregation-mode",
"decode",
"--disaggregation-bootstrap-port",
cls.bootstrap_port,
"--tp",
str(cls.NUM_DECODE_GPUS),
extra_decode_args = [
"--dp",
str(cls.NUM_DECODE_GPUS),
str(NUM_DECODE_GPUS),
"--enable-dp-attention",
"--disable-flashinfer-autotune",
"--mem-fraction-static",
"0.85",
"--base-gpu-id",
str(cls.NUM_PREFILL_GPUS),
]
decode_args += cls.transfer_backend + cls.rdma_devices
cls.process_decode = popen_launch_pd_server(
cls.model,
cls.decode_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=decode_args,
)
@classmethod
def setUpClass(cls):
super().setUpClass()
cls.model = GPT_OSS_MODEL_PATH
cls.launch_all()
def test_gsm8k(self):
metrics = run_eval(