[Fix] [H200] ci: Split GLM-5.3-Flash serving recipes into separate tests (#44892)
This commit is contained in:
@@ -0,0 +1,57 @@
|
||||
"""Shared server setup for the GLM-5.3-Flash H200 serving recipes."""
|
||||
|
||||
from sglang.srt.utils import kill_process_tree
|
||||
from sglang.test.test_utils import (
|
||||
DEFAULT_URL_FOR_TEST,
|
||||
CustomTestCase,
|
||||
_wait_for_gpu_idle_in_ci,
|
||||
popen_launch_server,
|
||||
try_cached_model,
|
||||
)
|
||||
|
||||
MODEL_PATH = "zai-org/GLM-5.3-Flash"
|
||||
SERVER_LAUNCH_TIMEOUT = 3600
|
||||
GPU_IDLE_TIMEOUT = 120
|
||||
|
||||
COMMON_SERVER_ARGS = [
|
||||
"--tp-size",
|
||||
"8",
|
||||
"--dsa-prefill-backend",
|
||||
"tilelang",
|
||||
"--dsa-decode-backend",
|
||||
"tilelang",
|
||||
"--kv-cache-dtype",
|
||||
"bf16",
|
||||
"--moe-runner-backend",
|
||||
"deep_gemm",
|
||||
"--reasoning-parser",
|
||||
"auto",
|
||||
"--tool-call-parser",
|
||||
"auto",
|
||||
]
|
||||
|
||||
|
||||
def _stop_server(process):
|
||||
if process:
|
||||
kill_process_tree(process.pid)
|
||||
_wait_for_gpu_idle_in_ci(timeout=GPU_IDLE_TIMEOUT)
|
||||
|
||||
|
||||
class _GLM53FlashH200Base(CustomTestCase):
|
||||
server_args: list[str]
|
||||
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
cls.model = try_cached_model(MODEL_PATH)
|
||||
cls.base_url = DEFAULT_URL_FOR_TEST
|
||||
cls.process = None
|
||||
cls.process = popen_launch_server(
|
||||
cls.model,
|
||||
cls.base_url,
|
||||
timeout=SERVER_LAUNCH_TIMEOUT,
|
||||
other_args=cls.server_args,
|
||||
)
|
||||
|
||||
@classmethod
|
||||
def tearDownClass(cls):
|
||||
_stop_server(getattr(cls, "process", None))
|
||||
Reference in New Issue
Block a user