141 lines
3.4 KiB
Python
141 lines
3.4 KiB
Python
import os
|
|
import subprocess
|
|
import time
|
|
import unittest
|
|
from types import SimpleNamespace
|
|
|
|
import requests
|
|
|
|
from sglang.srt.utils import kill_process_tree
|
|
from sglang.test.run_eval import run_eval
|
|
from sglang.test.server_fixtures.disaggregation_fixture import get_rdma_devices_args
|
|
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,
|
|
)
|
|
|
|
TEST_MODEL = os.environ.get("NIXL_EP_TEST_MODEL", DEFAULT_MODEL_NAME_FOR_TEST_MLA)
|
|
os.environ.setdefault("SGLANG_NIXL_EP_NUM_MAX_DISPATCH_TOKENS_PER_RANK", "1024")
|
|
|
|
ib_devices = get_rdma_devices_args()
|
|
|
|
NIXL_COMMON = [
|
|
"--trust-remote-code",
|
|
"--moe-a2a-backend",
|
|
"nixl",
|
|
"--deepep-mode",
|
|
"low_latency",
|
|
"--tp",
|
|
"8",
|
|
"--mem-fraction-static",
|
|
"0.78",
|
|
]
|
|
DP_ATTN = ["--dp", "8", "--enable-dp-attention"]
|
|
ELASTIC_NIXL = [
|
|
"--elastic-ep-backend",
|
|
"nixl",
|
|
"--enable-eplb",
|
|
"--ep-num-redundant-experts",
|
|
"24",
|
|
]
|
|
ELASTIC_MOONCAKE = [
|
|
"--elastic-ep-backend",
|
|
"mooncake",
|
|
"--mooncake-ib-device",
|
|
ib_devices,
|
|
"--enable-eplb",
|
|
"--ep-num-redundant-experts",
|
|
"24",
|
|
]
|
|
|
|
|
|
class _EPTestBase(CustomTestCase):
|
|
server_args: list[str] = []
|
|
|
|
@classmethod
|
|
def setUpClass(cls):
|
|
cls.model = TEST_MODEL
|
|
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=cls.server_args,
|
|
)
|
|
|
|
@classmethod
|
|
def tearDownClass(cls):
|
|
kill_process_tree(cls.process.pid)
|
|
cls.process.wait(timeout=15)
|
|
time.sleep(2)
|
|
|
|
def _run_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)
|
|
return metrics
|
|
|
|
def test_gsm8k(self):
|
|
metrics = self._run_gsm8k()
|
|
self.assertGreater(metrics["score"], 0.60)
|
|
|
|
|
|
class TestNixlEPTP(_EPTestBase):
|
|
server_args = [*NIXL_COMMON]
|
|
|
|
|
|
class TestNixlEPDPAttn(_EPTestBase):
|
|
server_args = [*NIXL_COMMON, *DP_ATTN]
|
|
|
|
|
|
class TestNixlEPElasticEP(_EPTestBase):
|
|
server_args = [*NIXL_COMMON, *DP_ATTN, *ELASTIC_NIXL]
|
|
|
|
|
|
class TestNixlMoeMooncakeElasticEP(_EPTestBase):
|
|
server_args = [
|
|
*NIXL_COMMON,
|
|
*DP_ATTN,
|
|
*ELASTIC_MOONCAKE,
|
|
"--moe-dense-tp-size",
|
|
"1",
|
|
"--enable-dp-lm-head",
|
|
]
|
|
|
|
pkill_process_1 = "sglang::scheduler_DP1_TP1_EP1"
|
|
|
|
def test_gsm8k_fault_1(self):
|
|
subprocess.run(
|
|
["pkill", "-f", f"^{self.pkill_process_1}$"],
|
|
check=True,
|
|
)
|
|
# Bootstrap one forward on a survivor so the controller learns the
|
|
# post-fault active-rank mask before dispatching concurrent requests.
|
|
response = requests.post(
|
|
f"{self.base_url}/generate",
|
|
json={
|
|
"text": "Hello",
|
|
"sampling_params": {"max_new_tokens": 1},
|
|
"routed_dp_rank": 0,
|
|
},
|
|
timeout=120,
|
|
)
|
|
self.assertEqual(response.status_code, 200, response.text)
|
|
metrics = self._run_gsm8k()
|
|
self.assertGreater(metrics["score"], 0.60)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|