EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[2/N] Added the core structure of elastic EP and the eplb algorithm with faulty rank (#10606)

Browse Source 
Co-authored-by: Xun Sun <UNIDY2002@outlook.com>
Co-authored-by: Shangming Cai <csmthu@gmail.com>
This commit is contained in:
Hank Han
2025-10-22 16:13:31 +08:00
committed by GitHub
parent e028af6998
commit 904655c5fd
7 changed files with 297 additions and 234 deletions
Download Patch File Download Diff File Expand all files Collapse all files

277
test/srt/ep/test_mooncake_ep_small.py
View File

@@ -3,6 +3,7 @@ from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_disaggregation_utils import get_rdma_devices_args
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST_MLA,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -11,8 +12,12 @@ from sglang.test.test_utils import (
popen_launch_server,
)
ib_devices = get_rdma_devices_args()
class TestTP(CustomTestCase):
extra_args = []
class TestPureDP(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = DEFAULT_MODEL_NAME_FOR_TEST_MLA
@@ -25,13 +30,10 @@ class TestPureDP(CustomTestCase):
"--trust-remote-code",
"--tp",
"4",
"--enable-dp-attention",
"--dp",
"4",
"--elastic-ep-backend",
"mooncake",
"--mooncake-ib-device",
"mlx5_roce0,mlx5_roce1,mlx5_roce2,mlx5_roce3,mlx5_roce4,mlx5_roce5,mlx5_roce6,mlx5_roce7",
ib_devices,
"--moe-a2a-backend",
"deepep",
"--deepep-mode",
@@ -44,6 +46,7 @@ class TestPureDP(CustomTestCase):
"512",
"--mem-fraction-static",
"0.5",
*cls.extra_args,
],
)
@@ -67,219 +70,73 @@ class TestPureDP(CustomTestCase):
self.assertGreater(metrics["accuracy"], 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",
"--elastic-ep-backend",
"mooncake",
"--mooncake-ib-device",
"mlx5_roce0,mlx5_roce1,mlx5_roce2,mlx5_roce3,mlx5_roce4,mlx5_roce5,mlx5_roce6,mlx5_roce7",
"--moe-a2a-backend",
"deepep",
"--deepep-mode",
"low_latency",
"--chunked-prefill-size",
"512",
"--cuda-graph-max-bs",
"128",
"--max-running-requests",
"256",
],
)
@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)
class TestPureDP(TestTP):
extra_args = [
"--tp",
"4",
"--enable-dp-attention",
"--dp",
"4",
]
class TestTP(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",
"--elastic-ep-backend",
"mooncake",
"--mooncake-ib-device",
"mlx5_roce0,mlx5_roce1,mlx5_roce2,mlx5_roce3,mlx5_roce4,mlx5_roce5,mlx5_roce6,mlx5_roce7",
"--moe-a2a-backend",
"deepep",
"--deepep-mode",
"low_latency",
"--chunked-prefill-size",
"512",
"--cuda-graph-max-bs",
"128",
"--max-running-requests",
"128",
],
)
@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)
class TestHybridDPTP(TestTP):
extra_args = [
"--tp",
"4",
"--enable-dp-attention",
"--dp",
"2",
]
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",
"--elastic-ep-backend",
"mooncake",
"--mooncake-ib-device",
"mlx5_roce0,mlx5_roce1,mlx5_roce2,mlx5_roce3,mlx5_roce4,mlx5_roce5,mlx5_roce6,mlx5_roce7",
"--moe-a2a-backend",
"deepep",
"--deepep-mode",
"low_latency",
"--chunked-prefill-size",
"512",
"--cuda-graph-max-bs",
"32",
"--max-running-requests",
"512",
],
)
@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)
class TestNoGatherdBuffer(TestTP):
extra_args = [
"--tp",
"4",
"--enable-dp-attention",
"--dp",
"4",
"--moe-dense-tp-size",
"1",
]
class TestTBO(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",
"--elastic-ep-backend",
"mooncake",
"--mooncake-ib-device",
"mlx5_roce0,mlx5_roce1,mlx5_roce2,mlx5_roce3,mlx5_roce4,mlx5_roce5,mlx5_roce6,mlx5_roce7",
"--moe-a2a-backend",
"deepep",
"--deepep-mode",
"low_latency",
"--chunked-prefill-size",
"512",
"--enable-two-batch-overlap",
"--cuda-graph-max-bs",
"128",
"--max-running-requests",
"512",
],
)
class TestTBO(TestTP):
extra_args = [
"--tp",
"4",
"--enable-dp-attention",
"--dp",
"4",
"--moe-dense-tp-size",
"1",
"--enable-two-batch-overlap",
]
@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)
class TestMooncakeWitchEPLB(TestTP):
extra_args = [
"--tp",
"4",
"--enable-dp-attention",
"--dp",
"4",
"--moe-dense-tp-size",
"1",
"--enable-two-batch-overlap",
"--enable-eplb",
"--ep-num-redundant-experts",
"4",
"--eplb-rebalance-num-iterations",
"50",
"--expert-distribution-recorder-buffer-size",
"50",
"--expert-distribution-recorder-mode",
"stat",
"--ep-dispatch-algorithm",
"static",
]
if __name__ == "__main__":