76 lines
2.1 KiB
Python
76 lines
2.1 KiB
Python
import unittest
|
|
from types import SimpleNamespace
|
|
|
|
from sglang.srt.utils import 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_URL_FOR_TEST,
|
|
CustomTestCase,
|
|
is_in_ci,
|
|
popen_launch_server,
|
|
write_github_step_summary,
|
|
)
|
|
|
|
register_cuda_ci(est_time=900, suite="nightly-4-gpu-b200", nightly=True)
|
|
|
|
FULL_DEEPSEEK_V3_FP4_MODEL_PATH = "nvidia/DeepSeek-V3-0324-FP4"
|
|
SERVER_LAUNCH_TIMEOUT = 1000
|
|
|
|
|
|
class TestDeepseekV3FP4CutlassMoE(CustomTestCase):
|
|
@classmethod
|
|
def setUpClass(cls):
|
|
cls.model = FULL_DEEPSEEK_V3_FP4_MODEL_PATH
|
|
cls.base_url = DEFAULT_URL_FOR_TEST
|
|
other_args = [
|
|
"--tp",
|
|
"4",
|
|
"--ep",
|
|
"4",
|
|
"--attention-backend",
|
|
"trtllm_mla",
|
|
"--moe-runner-backend",
|
|
"flashinfer_cutlass",
|
|
"--quantization",
|
|
"modelopt_fp4",
|
|
"--model-loader-extra-config",
|
|
'{"enable_multithread_load": true}',
|
|
]
|
|
cls.process = popen_launch_server(
|
|
cls.model,
|
|
cls.base_url,
|
|
timeout=SERVER_LAUNCH_TIMEOUT,
|
|
other_args=other_args,
|
|
)
|
|
|
|
@classmethod
|
|
def tearDownClass(cls):
|
|
kill_process_tree(cls.process.pid)
|
|
|
|
def test_a_gsm8k(
|
|
self,
|
|
): # Append an "a" to make this test run first (alphabetically) to warm up the server
|
|
args = SimpleNamespace(
|
|
num_shots=8,
|
|
data_path=None,
|
|
num_questions=1319,
|
|
parallel=1319,
|
|
max_new_tokens=512,
|
|
host="http://127.0.0.1",
|
|
port=int(self.base_url.split(":")[-1]),
|
|
)
|
|
metrics = run_eval_few_shot_gsm8k(args)
|
|
print(f"{metrics=}")
|
|
|
|
if is_in_ci():
|
|
write_github_step_summary(
|
|
f"### test_gsm8k (deepseek-v3-fp4-cutlass-moe)\n"
|
|
f'{metrics["accuracy"]=:.3f}\n'
|
|
)
|
|
self.assertGreater(metrics["accuracy"], 0.935)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|