2024-09-12 21:36:41 -07:00
|
|
|
import unittest
|
|
|
|
|
|
|
|
|
|
from sglang.test.test_utils import (
|
|
|
|
|
DEFAULT_MODEL_NAME_FOR_TEST,
|
|
|
|
|
DEFAULT_MOE_MODEL_NAME_FOR_TEST,
|
2024-09-14 15:38:37 -07:00
|
|
|
is_in_ci,
|
2024-11-21 20:07:48 -08:00
|
|
|
run_bench_one_batch,
|
2024-09-12 21:36:41 -07:00
|
|
|
)
|
|
|
|
|
|
|
|
|
|
|
2024-11-21 20:07:48 -08:00
|
|
|
class TestBenchOneBatch(unittest.TestCase):
|
2024-09-12 21:36:41 -07:00
|
|
|
def test_default(self):
|
2024-11-21 20:07:48 -08:00
|
|
|
output_throughput = run_bench_one_batch(DEFAULT_MODEL_NAME_FOR_TEST, [])
|
2024-09-12 21:36:41 -07:00
|
|
|
|
2024-09-14 15:38:37 -07:00
|
|
|
if is_in_ci():
|
2024-11-17 15:48:12 -08:00
|
|
|
self.assertGreater(output_throughput, 135)
|
2024-09-12 21:36:41 -07:00
|
|
|
|
|
|
|
|
def test_moe_default(self):
|
2024-11-21 20:07:48 -08:00
|
|
|
output_throughput = run_bench_one_batch(
|
2024-09-14 15:38:37 -07:00
|
|
|
DEFAULT_MOE_MODEL_NAME_FOR_TEST, ["--tp", "2"]
|
2024-09-12 21:36:41 -07:00
|
|
|
)
|
|
|
|
|
|
2024-09-14 15:38:37 -07:00
|
|
|
if is_in_ci():
|
2024-11-17 15:48:12 -08:00
|
|
|
self.assertGreater(output_throughput, 125)
|
2024-09-12 21:36:41 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
if __name__ == "__main__":
|
|
|
|
|
unittest.main()
|