Add support for AutoRound quantized models (#10153)

This commit is contained in:
Weiwei
2025-10-27 18:17:29 +08:00
committed by GitHub
parent a88b006ecf
commit caa4819bfc
8 changed files with 541 additions and 1 deletions

View File

@@ -0,0 +1,62 @@
"""
Usage:
python3 -m unittest test_autoround.TestAutoRound.test_mmlu
"""
import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_AUTOROUND_MODEL_NAME_FOR_TEST,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
class TestAutoRound(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.base_url = DEFAULT_URL_FOR_TEST
@classmethod
def tearDownClass(cls):
pass
def test_mmlu(self):
device = "auto"
for model in DEFAULT_AUTOROUND_MODEL_NAME_FOR_TEST:
with self.subTest(model=model):
print(f"\n[INFO] Launching server for model: {model}")
process = popen_launch_server(
model,
self.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=["--trust-remote-code", "--quantization", "auto-round"],
device=device,
)
try:
args = SimpleNamespace(
base_url=self.base_url,
model=model,
eval_name="mmlu",
num_examples=32,
num_threads=32,
device=device,
)
metrics = run_eval(args)
if "Llama" in model:
self.assertGreaterEqual(metrics["score"], 0.6)
else:
self.assertGreaterEqual(metrics["score"], 0.26)
finally:
kill_process_tree(process.pid)
print(f"[INFO] Server for {model} stopped.")
if __name__ == "__main__":
unittest.main()

View File

@@ -67,6 +67,7 @@ suites = {
TestFile("quant/test_int8_kernel.py", 8),
TestFile("quant/test_triton_scaled_mm.py", 8),
TestFile("quant/test_w8a8_quantization.py", 46),
TestFile("quant/test_autoround.py", 60),
TestFile("rl/test_fp32_lm_head.py", 30),
TestFile("rl/test_update_weights_from_disk.py", 114),
TestFile("rl/test_update_weights_from_tensor.py", 48),