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

[CI] Trim the base-c 4-gpu-h100 stage from 5 shards to 4 (#35407)

Browse Source
This commit is contained in:
Liangsheng Yin
2026-08-19 00:48:07 -07:00
committed by GitHub
parent e614121866
commit ccbe380028
12 changed files with 502 additions and 1099 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/ep/test_deepep_small.py
+1 -265
View File
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_cuda_ci(est_time=478, stage="base-c", runner_config="4-gpu-h100")
register_cuda_ci(est_time=270, stage="base-c", runner_config="4-gpu-h100")
class TestPureDP(CustomTestCase):
@@ -66,51 +66,6 @@ class TestPureDP(CustomTestCase):
self.assertGreater(metrics["score"], 0.60)
class TestHybridDPTP(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",
"--enable-dp-attention",
"--dp",
"2",
"--moe-a2a-backend",
"deepep",
"--cuda-graph-max-bs-decode",
"128",
"--max-running-requests",
"256",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval(args)
print(metrics)
self.assertGreater(metrics["score"], 0.60)
class TestTP(CustomTestCase):
@classmethod
def setUpClass(cls):
@@ -153,55 +108,6 @@ class TestTP(CustomTestCase):
self.assertGreater(metrics["score"], 0.60)
@unittest.skip("covered in test_deepep_large.py")
class TestNoGatherdBuffer(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",
"--enable-dp-attention",
"--dp",
"4",
"--moe-dense-tp-size",
"1",
"--enable-dp-lm-head",
"--moe-a2a-backend",
"deepep",
"--cuda-graph-max-bs-decode",
"32",
"--max-running-requests",
"512",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval(args)
print(metrics)
self.assertGreater(metrics["score"], 0.60)
class TestTBO(CustomTestCase):
@classmethod
def setUpClass(cls):
@@ -254,176 +160,6 @@ class TestTBO(CustomTestCase):
self.assertGreater(metrics["score"], 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-decode",
"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(
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval(args)
print(metrics)
self.assertGreater(metrics["score"], 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-decode",
"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(
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval(args)
print(metrics)
self.assertGreater(metrics["score"], 0.60)
@unittest.skip("covered in TestMTPWithTBO")
class TestMTP(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",
"--enable-dp-attention",
"--dp",
"2",
"--enable-dp-lm-head",
"--moe-a2a-backend",
"deepep",
"--speculative-algo",
"EAGLE",
"--speculative-draft-model-path",
DEFAULT_MODEL_NAME_FOR_TEST_MLA_NEXTN,
"--speculative-num-steps",
"2",
"--speculative-eagle-topk",
"3",
"--speculative-num-draft-tokens",
"3",
"--cuda-graph-max-bs-decode",
"32",
"--max-running-requests",
"64",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval(args)
print(metrics)
self.assertGreater(metrics["score"], 0.60)
server_info = requests.get(self.base_url + "/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['score']=:.3f}\n"
f"{avg_spec_accept_length=:.3f}\n"
)
self.assertGreater(avg_spec_accept_length, 2.1)
class TestMTPWithTBO(CustomTestCase):
@classmethod
def setUpClass(cls):
test/registered/ep/test_deepep_small_extra.py
+167
View File
@@ -0,0 +1,167 @@
import os
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.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST_MLA,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
register_cuda_ci(est_time=210, stage="extra-b", runner_config="4-gpu-h100")
class TestHybridDPTP(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",
"--enable-dp-attention",
"--dp",
"2",
"--moe-a2a-backend",
"deepep",
"--cuda-graph-max-bs-decode",
"128",
"--max-running-requests",
"256",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval(args)
print(metrics)
self.assertGreater(metrics["score"], 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-decode",
"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(
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval(args)
print(metrics)
self.assertGreater(metrics["score"], 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-decode",
"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(
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval(args)
print(metrics)
self.assertGreater(metrics["score"], 0.60)
if __name__ == "__main__":
unittest.main()