88 lines
2.8 KiB
Python
88 lines
2.8 KiB
Python
import os
|
|
import unittest
|
|
import warnings
|
|
|
|
from sglang.test.nightly_utils import NightlyBenchmarkRunner
|
|
from sglang.test.test_utils import (
|
|
DEFAULT_URL_FOR_TEST,
|
|
ModelLaunchSettings,
|
|
_parse_int_list_env,
|
|
parse_models,
|
|
)
|
|
|
|
PROFILE_DIR = "performance_profiles_vlms"
|
|
|
|
MODEL_DEFAULTS = [
|
|
# Keep conservative defaults. Can be overridden by env NIGHTLY_VLM_MODELS
|
|
ModelLaunchSettings(
|
|
"Qwen/Qwen2.5-VL-7B-Instruct",
|
|
extra_args=["--mem-fraction-static=0.7"],
|
|
),
|
|
ModelLaunchSettings(
|
|
"google/gemma-3-27b-it",
|
|
),
|
|
ModelLaunchSettings("Qwen/Qwen3-VL-30B-A3B-Instruct", extra_args=["--tp=2"]),
|
|
# "OpenGVLab/InternVL2_5-2B",
|
|
# buggy in official transformers impl
|
|
# "openbmb/MiniCPM-V-2_6",
|
|
]
|
|
|
|
|
|
class TestNightlyVLMModelsPerformance(unittest.TestCase):
|
|
@classmethod
|
|
def setUpClass(cls):
|
|
warnings.filterwarnings(
|
|
"ignore", category=ResourceWarning, message="unclosed.*socket"
|
|
)
|
|
|
|
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))
|
|
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")
|
|
cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_VLM_INPUT_LENS", "4096"))
|
|
cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_VLM_OUTPUT_LENS", "512"))
|
|
cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
|
|
cls.runner.setup_profile_directory()
|
|
|
|
def test_bench_one_batch(self):
|
|
all_model_succeed = True
|
|
|
|
for model_setup in self.models:
|
|
with self.subTest(model=model_setup.model_path):
|
|
# VLMs need additional benchmark args for dataset and trust-remote-code
|
|
extra_bench_args = [
|
|
"--trust-remote-code",
|
|
"--dataset-name=mmmu",
|
|
]
|
|
|
|
results, success = self.runner.run_benchmark_for_model(
|
|
model_path=model_setup.model_path,
|
|
batch_sizes=self.batch_sizes,
|
|
input_lens=self.input_lens,
|
|
output_lens=self.output_lens,
|
|
other_args=model_setup.extra_args,
|
|
extra_bench_args=extra_bench_args,
|
|
)
|
|
|
|
if not success:
|
|
all_model_succeed = False
|
|
|
|
self.runner.add_report(results)
|
|
|
|
self.runner.write_final_report()
|
|
|
|
if not all_model_succeed:
|
|
raise AssertionError("Some models failed the perf tests.")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|