[PD] Fix prefill bootstrap registration failure with --host 0.0.0.0 (#27608)

This commit is contained in:
Jianhong Zhang
2026-06-10 09:15:26 +08:00
committed by GitHub
parent 2947781ce6
commit 77c4d53f19
2 changed files with 53 additions and 0 deletions
@@ -392,6 +392,11 @@ class CommonKVManager(BaseKVManager):
else:
# Single-node case: bootstrap server's host is the same as http server's host
host = self.bootstrap_host
# If the server was bound to the wildcard address (0.0.0.0 / ::), use the
# actual local IP instead — a PUT to http://0.0.0.0:<port>/route is rejected
# with 403 by aiohttp ≥3.9 because 0.0.0.0 is not a valid HTTP Host value.
if host in ("0.0.0.0", "::"):
host = self.local_ip
bootstrap_na = NetworkAddress(host, self.bootstrap_port)
url = f"{bootstrap_na.to_url()}/route"
@@ -183,6 +183,54 @@ class TestRegisterToBootstrap(CustomTestCase):
url_used = mock_put.call_args[0][0]
self.assertIn("10.0.0.1", url_used)
@patch("sglang.srt.disaggregation.common.conn.time")
@patch("sglang.srt.disaggregation.common.conn.requests.put")
def test_wildcard_host_0000_uses_local_ip(self, mock_put, mock_time):
"""When --host 0.0.0.0 is used, the PUT must target local_ip not 0.0.0.0.
Scenario: cross-node P/D disagg where each role runs on a single node
(tp=1). Each machine runs its own SGLang instance with --host 0.0.0.0
to accept remote connections. dist_init_addr is None because tp=1
needs no multi-node rendezvous, so register_to_bootstrap takes the
else-branch and would use bootstrap_host="0.0.0.0" as the PUT target.
aiohttp >=3.9 rejects that with HTTP 403 because 0.0.0.0 is not a
valid Host header value.
Fix: substitute self.local_ip when bootstrap_host is a wildcard.
"""
mock_time.monotonic.return_value = 0.0
success_resp = MagicMock()
success_resp.status_code = 200
mock_put.return_value = success_resp
mgr = self._make_manager()
mgr.bootstrap_host = "0.0.0.0"
mgr.local_ip = "192.168.1.10"
mgr.register_to_bootstrap()
url_used = mock_put.call_args[0][0]
self.assertNotIn("0.0.0.0", url_used)
self.assertIn("192.168.1.10", url_used)
@patch("sglang.srt.disaggregation.common.conn.time")
@patch("sglang.srt.disaggregation.common.conn.requests.put")
def test_wildcard_host_ipv6_uses_local_ip(self, mock_put, mock_time):
"""Same fix for the IPv6 wildcard \"::\": must use local_ip instead."""
mock_time.monotonic.return_value = 0.0
success_resp = MagicMock()
success_resp.status_code = 200
mock_put.return_value = success_resp
mgr = self._make_manager()
mgr.bootstrap_host = "::"
mgr.local_ip = "fd00::1"
mgr.register_to_bootstrap()
url_used = mock_put.call_args[0][0]
# "::" bracketed as "[::]:port" should not appear; local_ip should
self.assertNotIn("[::]", url_used)
self.assertIn("fd00", url_used)
def _make_manager(self, dist_init_addr=None):
"""Create a lightweight mock manager that has the attributes needed
by register_to_bootstrap, without going through CommonKVManager.__init__