chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[PP] Add pp support for Qwen3-VL (#12333)

Browse Source 
Signed-off-by: Xuchun Shang <xuchun.shang@gmail.com>
Signed-off-by: Kun(llfl) <i@imux.top>
Signed-off-by: Kun(llfl) <llfl@linux.alibaba.com>
Co-authored-by: kun-llfl <i@imux.top>
Co-authored-by: Kun(llfl) <llfl@linux.alibaba.com>
This commit is contained in:
Xuchun Shang
2025-12-17 16:03:58 +08:00
committed by GitHub
co-authored by kun-llfl Kun
parent cdce516331
commit 45a959d3e9
5 changed files with 119 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/srt/test_pp_single_node.py
+57
View File
@@ -3,6 +3,7 @@ Usage:
python3 -m unittest test_pp_single_node.TestPPAccuracy.test_gsm8k
python3 -m unittest test_pp_single_node.TestQwenPPAccuracy.test_pp_consistency
python3 -m unittest test_pp_single_node.TestFixedBugs.test_chunked_prefill_with_small_bs
python3 -m unittest test_pp_single_node.TestQwenVLPPAccuracy.test_mmmu
"""
import time
@@ -20,6 +21,7 @@ from sglang.test.test_utils import (
DEFAULT_MLA_MODEL_NAME_FOR_TEST,
DEFAULT_MODEL_NAME_FOR_TEST,
DEFAULT_MODEL_NAME_FOR_TEST_GLM_41V_PP,
DEFAULT_MODEL_NAME_FOR_TEST_VL_PP,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
@@ -131,6 +133,61 @@ class TestDPAttentionDP2PP2(CustomTestCase):
self.assertGreater(metrics["score"], 0.8)
class TestQwenVLPPAccuracy(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.model = DEFAULT_MODEL_NAME_FOR_TEST_VL_PP
cls.base_url = "http://127.0.0.1:23333"
cls.process = popen_launch_server(
DEFAULT_MODEL_NAME_FOR_TEST_VL_PP,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--tp-size",
1,
"--pp-size",
4,
"--chunked-prefill-size",
8192,
"--enable-multimodal",
],
)
def test_gsm8k(self):
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=200,
max_new_tokens=512,
parallel=128,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval_few_shot_gsm8k(args)
print(f"{metrics=}")
self.assertGreater(metrics["accuracy"], 0.65)
# Wait a little bit so that the memory check happens.
time.sleep(4)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
@unittest.skipIf(is_in_ci(), "To reduce the CI execution time.")
def test_mmmu(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="mmmu",
num_examples=None,
num_threads=32,
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreater(metrics["score"], 0.26)
class TestQwenPPAccuracy(unittest.TestCase):
@classmethod
def setUpClass(cls):