40 lines
1.2 KiB
Python
40 lines
1.2 KiB
Python
"""
|
|
Usage:
|
|
python3 -m unittest test_overlap_schedule.TestOverlapSchedule.test_radix_attention_chunked_prefill
|
|
python3 test_overlap_schedule.py
|
|
"""
|
|
|
|
import unittest
|
|
|
|
from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
|
|
from sglang.test.test_utils import CustomTestCase, run_mmlu_test
|
|
|
|
register_cuda_ci(est_time=245, suite="stage-b-test-large-1-gpu")
|
|
register_amd_ci(est_time=275, suite="stage-b-test-small-1-gpu-amd")
|
|
|
|
|
|
class TestOverlapSchedule(CustomTestCase):
|
|
def test_no_radix_attention_chunked_prefill(self):
|
|
run_mmlu_test(
|
|
disable_radix_cache=True, chunked_prefill_size=32, disable_overlap=True
|
|
)
|
|
|
|
def test_no_radix_attention_no_chunked_prefill(self):
|
|
run_mmlu_test(
|
|
disable_radix_cache=True, chunked_prefill_size=-1, disable_overlap=True
|
|
)
|
|
|
|
def test_radix_attention_chunked_prefill(self):
|
|
run_mmlu_test(
|
|
disable_radix_cache=False, chunked_prefill_size=32, disable_overlap=True
|
|
)
|
|
|
|
def test_radix_attention_no_chunked_prefill(self):
|
|
run_mmlu_test(
|
|
disable_radix_cache=False, chunked_prefill_size=-1, disable_overlap=True
|
|
)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|