chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[bugfix] fix TBO crashes when attn_tp_size > 1 (#13730)

Browse Source 
Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com>
This commit is contained in:
Yuhao Yao
2025-12-11 18:18:40 -08:00
committed by GitHub
co-authored by Cheng Wan
parent 0aa3dec5c7
commit e9e7f15eb5
20 changed files with 285 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/srt/ep/test_deepep_small.py
+178
View File
@@ -251,6 +251,108 @@ class TestTBO(CustomTestCase):
self.assertGreater(metrics["accuracy"], 0.60)
class TestTBOWithTPAttn(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = DEFAULT_MODEL_NAME_FOR_TEST_MLA
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--trust-remote-code",
"--tp",
"4",
"--moe-a2a-backend",
"deepep",
"--enable-two-batch-overlap",
"--cuda-graph-max-bs",
"128",
"--max-running-requests",
"512",
"--mem-fraction-static", # temp fix as DeepEP buffer is too large.
"0.7",
],
env={
**os.environ,
"SGLANG_TBO_DEBUG": "1",
},
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
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(metrics)
self.assertGreater(metrics["accuracy"], 0.60)
# There exists bug when using MTP + TBO + attn_tp_size > 1, currently skip that case.
# @unittest.skip("covered in TestMTPWithTPAttnAndTBO")
class TestTBOWithTPAttnAndDenseDP(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = DEFAULT_MODEL_NAME_FOR_TEST_MLA
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--trust-remote-code",
"--tp",
"4",
"--moe-dense-tp-size",
"1",
"--moe-a2a-backend",
"deepep",
"--enable-two-batch-overlap",
"--cuda-graph-max-bs",
"128",
"--max-running-requests",
"512",
"--mem-fraction-static", # temp fix as DeepEP buffer is too large.
"0.7",
],
env={
**os.environ,
"SGLANG_TBO_DEBUG": "1",
},
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
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(metrics)
self.assertGreater(metrics["accuracy"], 0.60)
@unittest.skip("covered in TestMTPWithTBO")
class TestMTP(CustomTestCase):
@classmethod
@@ -393,5 +495,81 @@ class TestMTPWithTBO(CustomTestCase):
self.assertGreater(avg_spec_accept_length, 2.1)
@unittest.skip("skipped due to bug when using MTP & TBO & attn_tp_size > 1")
class TestMTPWithTPAttnAndTBO(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = DEFAULT_MODEL_NAME_FOR_TEST_MLA
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--tp-size",
"4",
"--moe-dense-tp-size",
"1",
"--enable-two-batch-overlap",
"--moe-a2a-backend",
"deepep",
"--trust-remote-code",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"2",
"--speculative-eagle-topk",
"3",
"--speculative-num-draft-tokens",
"3",
"--speculative-draft-model-path",
DEFAULT_MODEL_NAME_FOR_TEST_MLA_NEXTN,
"--chunked-prefill-size",
"256",
"--cuda-graph-max-bs",
"32",
"--max-running-requests",
"128",
"--mem-fraction-static", # temp fix as DeepEP buffer is too large.
"0.7",
],
env={
**os.environ,
"SGLANG_TBO_DEBUG": "1",
},
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
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(metrics)
self.assertGreater(metrics["accuracy"], 0.60)
server_info = requests.get(self.base_url + "/get_server_info")
avg_spec_accept_length = server_info.json()["internal_states"][0][
"avg_spec_accept_length"
]
print(
f"###test_gsm8k (deepseek-v3 mtp + dp + tbo):\n"
f"accuracy={metrics['accuracy']=:.3f}\n"
f"{avg_spec_accept_length=:.3f}\n"
)
self.assertGreater(avg_spec_accept_length, 2.1)
if __name__ == "__main__":
unittest.main()