[1/N] Introduce Mooncake Backend and Mooncake EP to Support Elastic EP (#10423)

Co-authored-by: Hank Han <hanhan7630@outlook.com>
Co-authored-by: Shangming Cai <csmthu@gmail.com>
This commit is contained in:
Xun Sun
2025-10-15 10:40:54 +08:00
committed by GitHub
parent 74737b2863
commit a40229f6f8
13 changed files with 798 additions and 32 deletions

View File

@@ -0,0 +1,286 @@
import unittest
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_utils import (
DEFAULT_MODEL_NAME_FOR_TEST_MLA,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
class TestPureDP(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",
"--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",
"512",
"--mem-fraction-static",
"0.5",
],
)
@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(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 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 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 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",
],
)
@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)
if __name__ == "__main__":
unittest.main()

View File

@@ -171,6 +171,7 @@ suites = {
],
"per-commit-4-gpu-deepep": [
TestFile("ep/test_deepep_small.py", 531),
TestFile("ep/test_mooncake_ep_small.py", 450),
],
"per-commit-8-gpu-deepep": [
TestFile("ep/test_deepep_large.py", 338),