Files
sglang/test/registered/radix_cache/test_radix_cache_hit.py
T

53 lines
1.4 KiB
Python

import unittest
from sglang.test.ci.ci_register import (
register_amd_ci,
register_cpu_ci,
register_cuda_ci,
)
from sglang.test.kits.cache_hit_kit import run_multiturn_cache_hit_test
from sglang.test.test_utils import (
DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
terminate_and_kill_process_tree,
)
register_cuda_ci(est_time=71, stage="base-b", runner_config="1-gpu-small")
register_amd_ci(est_time=55, suite="stage-b-test-1-gpu-small-amd")
register_cpu_ci(est_time=113, suite="stage-b-test-cpu-intel")
MODEL = DEFAULT_SMALL_MODEL_NAME_FOR_TEST
class TestRadixCacheHit(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
)
@classmethod
def tearDownClass(cls):
terminate_and_kill_process_tree(cls.process, wait_timeout=60)
def test_multiturn_cache_hit(self):
run_multiturn_cache_hit_test(
base_url=self.base_url,
model_path=self.model,
num_clients=8,
num_rounds=6,
request_length=289,
output_length=367,
)
if __name__ == "__main__":
unittest.main()