[diffusion] security: pin multimodal-gen ZMQ scheduler ingress to loopback on wildcard --host (#36854)

Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
This commit is contained in:
Jinbao Chen
2026-10-06 16:05:17 +08:00
committed by GitHub
co-authored by Mick Qian
parent c7f3ae1dd5
commit d0494af08b
2 changed files with 154 additions and 34 deletions
@@ -3144,16 +3144,26 @@ class ServerArgs(DisaggServerArgsMixin):
def scheduler_endpoint(self):
"""
Internal endpoint for scheduler.
Normalizes localhost and IPv6 hosts to IPv4 loopback for internal ZMQ.
Wildcard, localhost, and IPv6 hosts use IPv4 loopback for internal ZMQ.
"""
return self.scheduler_endpoint_for(0)
def scheduler_endpoint_for(self, replica: int) -> str:
"""Ingress endpoint of one DP replica's driver rank."""
"""Ingress endpoint of one DP replica's driver rank.
The scheduler ingress is an unauthenticated pickle-RPC endpoint
(``managers/scheduler.py`` ``recv_reqs`` deserializes client bytes
with ``pickle.loads``), so it must never be derived from the public
``--host``: binding it to ``0.0.0.0`` would expose unsafe
deserialization to the network (CVE-2026-3059 family). Wildcard hosts
are pinned to loopback; IPv6 hosts also use IPv4 loopback for internal
ZMQ compatibility. Explicit non-wildcard IPv4 hosts and hostnames
(used for intentional cross-machine deployments) are honored.
"""
scheduler_host = self.host
if (
scheduler_host is None
or scheduler_host == "localhost"
or scheduler_host in ("localhost", "0.0.0.0")
or is_valid_ipv6_address(scheduler_host)
):
scheduler_host = "127.0.0.1"
@@ -3,7 +3,7 @@ import os
import sys
import tempfile
import unittest
from contextlib import contextmanager
from contextlib import ExitStack, contextmanager
from unittest.mock import Mock, patch
import zmq
@@ -160,6 +160,15 @@ class TestPlatformLifecycleHooks(unittest.TestCase):
apply_defaults.assert_called_once_with(server_args)
class _CudaPlatformTestCase(unittest.TestCase):
def setUp(self):
super().setUp()
# Residency queries inspect the platform after argument construction too.
stack = ExitStack()
self.addCleanup(stack.close)
stack.enter_context(_mock_cuda_platform())
class TestSchedulerEndpoints(unittest.TestCase):
def test_host_normalization_preserves_replica_ports(self):
args = _from_dict_without_model_resolution({"model_path": "test/model"})
@@ -171,7 +180,7 @@ class TestSchedulerEndpoints(unittest.TestCase):
("::", "127.0.0.1"),
("::1", "127.0.0.1"),
("2001:db8::1", "127.0.0.1"),
("0.0.0.0", "0.0.0.0"),
("0.0.0.0", "127.0.0.1"),
("127.0.0.1", "127.0.0.1"),
("192.0.2.1", "192.0.2.1"),
("scheduler.example", "scheduler.example"),
@@ -189,30 +198,39 @@ class TestSchedulerEndpoints(unittest.TestCase):
for replica, endpoint in enumerate(expected):
self.assertEqual(args.scheduler_endpoint_for(replica), endpoint)
def test_ipv6_http_host_allows_internal_zmq_round_trip(self):
def test_wildcard_and_ipv6_http_hosts_allow_internal_zmq_round_trip(self):
args = _from_dict_without_model_resolution({"model_path": "test/model"})
args.host = "::1"
args.scheduler_port = 0
args.scheduler_ports = None
self.assertEqual(args.url(), f"http://[::1]:{args.port}")
with zmq.Context() as context:
with context.socket(zmq.REP) as receiver, context.socket(zmq.REQ) as sender:
for socket in (receiver, sender):
socket.setsockopt(zmq.LINGER, 0)
socket.setsockopt(zmq.RCVTIMEO, 2000)
socket.setsockopt(zmq.SNDTIMEO, 2000)
receiver.bind(args.scheduler_endpoint)
bound_endpoint = receiver.getsockopt_string(zmq.LAST_ENDPOINT)
args.scheduler_port = int(bound_endpoint.rsplit(":", 1)[1])
sender.connect(args.scheduler_endpoint)
sender.send(b"ping")
self.assertEqual(receiver.recv(), b"ping")
receiver.send(b"pong")
self.assertEqual(sender.recv(), b"pong")
for host, http_host in (
("0.0.0.0", "127.0.0.1"),
("::", "[::1]"),
("::1", "[::1]"),
):
with self.subTest(host=host), zmq.Context() as context:
args.host = host
args.scheduler_port = 0
self.assertEqual(args.url(), f"http://{http_host}:{args.port}")
with (
context.socket(zmq.REP) as receiver,
context.socket(zmq.REQ) as sender,
):
for socket in (receiver, sender):
socket.setsockopt(zmq.LINGER, 0)
socket.setsockopt(zmq.RCVTIMEO, 2000)
socket.setsockopt(zmq.SNDTIMEO, 2000)
receiver.bind(args.scheduler_endpoint)
bound_endpoint = receiver.getsockopt_string(zmq.LAST_ENDPOINT)
self.assertTrue(bound_endpoint.startswith("tcp://127.0.0.1:"))
args.scheduler_port = int(bound_endpoint.rsplit(":", 1)[1])
sender.connect(args.scheduler_endpoint)
sender.send(b"ping")
self.assertEqual(receiver.recv(), b"ping")
receiver.send(b"pong")
self.assertEqual(sender.recv(), b"pong")
class TestServerArgsPathExpansion(unittest.TestCase):
class TestServerArgsPathExpansion(_CudaPlatformTestCase):
def _from_dict_without_model_resolution(self, kwargs):
return _from_dict_without_model_resolution(kwargs)
@@ -1008,7 +1026,7 @@ class TestMiniMaxH3Routing(unittest.TestCase):
)
class TestOffloadDefaults(unittest.TestCase):
class TestOffloadDefaults(_CudaPlatformTestCase):
def test_wan_decode_precision_defaults(self):
for pipeline_config in (
WanT2V480PConfig(),
@@ -2532,14 +2550,18 @@ class TestOffloadDefaults(unittest.TestCase):
)
def test_explicit_multi_gpu_dit_layerwise_only_selects_dit_group(self):
args = self._from_dict_with_pipeline_config(
MOVAPipelineConfig(),
kwargs={
"model_path": "OpenMOSS-Team/MOVA-360p",
"num_gpus": 2,
"dit_layerwise_offload": True,
},
)
with patch(
"sglang.multimodal_gen.registry.maybe_download_model_index",
return_value={"_class_name": "MOVAPipeline"},
):
args = self._from_dict_with_pipeline_config(
MOVAPipelineConfig(),
kwargs={
"model_path": "OpenMOSS-Team/MOVA-360p",
"num_gpus": 2,
"dit_layerwise_offload": True,
},
)
self.assertFalse(args.use_fsdp_inference)
self.assertTrue(args.dit_cpu_offload)
@@ -3390,6 +3412,13 @@ class TestPerRoleParallelism(unittest.TestCase):
class TestPipelineResolutionCliOverride(unittest.TestCase):
def setUp(self):
_get_config_info.cache_clear()
model_index = patch(
"sglang.multimodal_gen.registry.maybe_download_model_index",
return_value={"_class_name": "QwenImageLayeredPipeline"},
)
model_index.start()
self.addCleanup(model_index.stop)
self.addCleanup(_get_config_info.cache_clear)
def test_resolution_flag_overrides_qwen_image_layered_pipeline_config(self):
parser = FlexibleArgumentParser()
@@ -3549,7 +3578,7 @@ class TestNcclNvlsArgs(unittest.TestCase):
self.assertFalse(disabled_args.enable_nccl_nvls)
class TestDirectGpuWeightLoading(unittest.TestCase):
class TestDirectGpuWeightLoading(_CudaPlatformTestCase):
def _args(self) -> ServerArgs:
args = ServerArgs.__new__(ServerArgs)
args.direct_gpu_weight_loading = True
@@ -3620,6 +3649,87 @@ class TestDirectGpuWeightLoading(unittest.TestCase):
tp_args._validate_direct_gpu_weight_loading()
class TestSchedulerEndpointBinding(unittest.TestCase):
"""The scheduler ingress is an unauthenticated pickle-RPC endpoint; it must
never follow a wildcard --host (see ServerArgs.scheduler_endpoint_for)."""
def _args(self, kwargs):
# Endpoint formatting does not need model resolution or port allocation.
args = ServerArgs.__new__(ServerArgs)
for key, value in kwargs.items():
setattr(args, key, value)
return args
def test_wildcard_host_pinned_to_loopback(self):
args = self._args(
{"model_path": "/fake/model", "host": "0.0.0.0", "scheduler_port": 5555}
)
self.assertEqual(args.scheduler_endpoint_for(0), "tcp://127.0.0.1:5555")
def test_ipv6_wildcard_host_pinned_to_loopback(self):
args = self._args(
{"model_path": "/fake/model", "host": "::", "scheduler_port": 5555}
)
self.assertEqual(args.scheduler_endpoint_for(0), "tcp://127.0.0.1:5555")
def test_default_host_stays_loopback(self):
args = self._args({"model_path": "/fake/model", "scheduler_port": 5555})
self.assertEqual(args.scheduler_endpoint_for(0), "tcp://127.0.0.1:5555")
def test_explicit_host_is_honored(self):
args = self._args(
{
"model_path": "/fake/model",
"host": "10.1.2.3",
"scheduler_port": 5555,
}
)
self.assertEqual(args.scheduler_endpoint_for(0), "tcp://10.1.2.3:5555")
def test_explicit_ipv6_host_uses_ipv4_loopback(self):
args = self._args(
{
"model_path": "/fake/model",
"host": "::1",
"scheduler_port": 5555,
}
)
self.assertEqual(args.scheduler_endpoint_for(0), "tcp://127.0.0.1:5555")
def test_replica_ports_increment_on_loopback(self):
args = self._args(
{"model_path": "/fake/model", "host": "0.0.0.0", "scheduler_port": 5555}
)
self.assertEqual(args.scheduler_endpoint_for(1), "tcp://127.0.0.1:5556")
def test_explicit_scheduler_ports_on_loopback(self):
args = self._args(
{
"model_path": "/fake/model",
"host": "0.0.0.0",
"scheduler_port": 5555,
"dp_size": 2,
}
)
args.scheduler_ports = [6100, 6200]
self.assertEqual(args.scheduler_endpoint, "tcp://127.0.0.1:6100")
self.assertEqual(
args.scheduler_endpoints,
["tcp://127.0.0.1:6100", "tcp://127.0.0.1:6200"],
)
def test_explicit_scheduler_ports_with_ipv6(self):
args = self._args(
{
"model_path": "/fake/model",
"host": "::1",
"scheduler_port": 5555,
}
)
args.scheduler_ports = [6100, 6200]
self.assertEqual(args.scheduler_endpoint_for(1), "tcp://127.0.0.1:6200")
class TestLayerwiseResidencyLifetime(unittest.TestCase):
def _help_by_option(self):
parser = FlexibleArgumentParser()