EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

ci: unify the model launch method of nightly ci (#11230)

Browse Source
This commit is contained in:
Mick
2025-10-08 09:13:14 +08:00
committed by GitHub
parent f3764c26a3
commit 64d1505c0a
5 changed files with 192 additions and 153 deletions
Download Patch File Download Diff File Expand all files Collapse all files

47
test/srt/test_nightly_vlms_perf.py
View File

@@ -8,6 +8,7 @@ from sglang.srt.utils import kill_process_tree
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
ModelLaunchSettings,
_parse_int_list_env,
is_in_ci,
parse_models,
@@ -19,8 +20,13 @@ PROFILE_DIR = "performance_profiles_vlms"
MODEL_DEFAULTS = [
# Keep conservative defaults. Can be overridden by env NIGHTLY_VLM_MODELS
"Qwen/Qwen2.5-VL-7B-Instruct",
"google/gemma-3-27b-it",
ModelLaunchSettings(
"Qwen/Qwen2.5-VL-7B-Instruct",
extra_args=["--mem-fraction-static=0.7"],
),
ModelLaunchSettings(
"google/gemma-3-27b-it",
),
# "OpenGVLab/InternVL2_5-2B",
# buggy in official transformers impl
# "openbmb/MiniCPM-V-2_6",
@@ -33,9 +39,18 @@ class TestNightlyVLMModelsPerformance(unittest.TestCase):
warnings.filterwarnings(
"ignore", category=ResourceWarning, message="unclosed.*socket"
)
cls.models = parse_models(
os.environ.get("NIGHTLY_VLM_MODELS", ",".join(MODEL_DEFAULTS))
)
nightly_vlm_models_str = os.environ.get("NIGHTLY_VLM_MODELS")
if nightly_vlm_models_str:
cls.models = []
model_paths = parse_models(nightly_vlm_models_str)
for model_path in model_paths:
cls.models.append(
ModelLaunchSettings(model_path, extra_args=VLM_EXTRA_ARGS)
)
else:
cls.models = MODEL_DEFAULTS
cls.base_url = DEFAULT_URL_FOR_TEST
cls.batch_sizes = _parse_int_list_env("NIGHTLY_VLM_BATCH_SIZES", "1,1,2,8,16")
@@ -46,29 +61,31 @@ class TestNightlyVLMModelsPerformance(unittest.TestCase):
def test_bench_one_batch(self):
all_benchmark_results = []
for model in self.models:
for model_setup in self.models:
benchmark_results = []
with self.subTest(model=model):
with self.subTest(model=model_setup.model_path):
process = popen_launch_server(
model=model,
model=model_setup.model_path,
base_url=self.base_url,
other_args=["--mem-fraction-static=0.7"],
other_args=model_setup.extra_args,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
)
try:
# Run bench_one_batch_server against the launched server
profile_filename = f"{model.replace('/', '_')}"
profile_filename = f"{model_setup.model_path.replace('/', '_')}"
# path for this run
profile_path_prefix = os.path.join(PROFILE_DIR, profile_filename)
# JSON output file for this model
json_output_file = f"results_{model.replace('/', '_')}.json"
json_output_file = (
f"results_{model_setup.model_path.replace('/', '_')}.json"
)
command = [
"python3",
"-m",
"sglang.bench_one_batch_server",
f"--model={model}",
f"--model={model_setup.model_path}",
"--base-url",
self.base_url,
"--batch-size",
@@ -91,12 +108,14 @@ class TestNightlyVLMModelsPerformance(unittest.TestCase):
result = subprocess.run(command, capture_output=True, text=True)
if result.returncode != 0:
print(f"Error running benchmark for {model} with batch size:")
print(
f"Error running benchmark for {model_setup.model_path} with batch size:"
)
print(result.stderr)
# Continue to next batch size even if one fails
continue
print(f"Output for {model} with batch size:")
print(f"Output for {model_setup.model_path} with batch size:")
print(result.stdout)
# Load and deserialize JSON results