86 lines
2.5 KiB
Python
86 lines
2.5 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,
|
|
try_cached_model,
|
|
)
|
|
|
|
register_cuda_ci(est_time=360, suite="stage-c-test-4-gpu-b200")
|
|
|
|
MODEL_PATH = "nvidia/Llama-3.1-8B-Instruct-NVFP4"
|
|
|
|
|
|
class FP4GemmBase:
|
|
backend = None
|
|
|
|
@classmethod
|
|
def setUpClass(cls):
|
|
if cls.backend is None:
|
|
raise NotImplementedError("Subclass must set 'backend' attribute")
|
|
cls.model = try_cached_model(MODEL_PATH)
|
|
cls.base_url = DEFAULT_URL_FOR_TEST
|
|
other_args = [
|
|
"--trust-remote-code",
|
|
"--quantization",
|
|
"modelopt_fp4",
|
|
"--fp4-gemm-backend",
|
|
cls.backend,
|
|
]
|
|
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"], 0.64)
|
|
|
|
|
|
@unittest.skipIf(get_device_sm() < 100, "Test requires CUDA SM 100 or higher")
|
|
class TestFP4GemmAuto(FP4GemmBase, unittest.TestCase):
|
|
backend = "auto"
|
|
|
|
|
|
@unittest.skipIf(get_device_sm() < 100, "Test requires CUDA SM 100 or higher")
|
|
class TestFP4GemmFlashinferCutlass(FP4GemmBase, unittest.TestCase):
|
|
backend = "flashinfer_cutlass"
|
|
|
|
|
|
@unittest.skipIf(get_device_sm() < 100, "Test requires CUDA SM 100 or higher")
|
|
class TestFP4GemmFlashinferCudnn(FP4GemmBase, unittest.TestCase):
|
|
backend = "flashinfer_cudnn"
|
|
|
|
|
|
@unittest.skipIf(get_device_sm() < 100, "Test requires CUDA SM 100 or higher")
|
|
class TestFP4GemmFlashinferTrtllm(FP4GemmBase, unittest.TestCase):
|
|
backend = "flashinfer_trtllm"
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|