Files
sglang/test/registered/attention/test_flash_attention_4.py
2026-01-11 15:53:38 +08:00

61 lines
1.8 KiB
Python

import unittest
from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import get_device_sm, kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
popen_launch_server,
)
# FlashAttention4 integration test (requires SM 100+ / Blackwell B200)
register_cuda_ci(est_time=200, suite="stage-b-test-4-gpu-b200")
@unittest.skipIf(get_device_sm() < 100, "Test requires CUDA SM 100 or higher")
class TestFlashAttention4(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.model = "Qwen/Qwen3-8B"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = [
"--trust-remote-code",
"--prefill-attention-backend",
"fa4",
"--decode-attention-backend",
"flashinfer",
]
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
parsed_url = urlparse(self.base_url)
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=1319,
max_new_tokens=512,
parallel=200,
host=f"{parsed_url.scheme}://{parsed_url.hostname}",
port=parsed_url.port,
)
metrics = run_eval_few_shot_gsm8k(args)
print(metrics)
self.assertGreater(metrics["accuracy"], 0.89)
if __name__ == "__main__":
unittest.main()