Fix IPv6 wildcard bootstrap address resolution in disagg (#29459)

This commit is contained in:
Lianmin Zheng
2026-06-26 16:05:36 -07:00
committed by GitHub
parent 5eebb4b2e8
commit 9c3227b689
2 changed files with 14 additions and 13 deletions
@@ -394,11 +394,12 @@ 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
# A wildcard bind address (0.0.0.0 / ::) is not a valid HTTP Host
# and can't be connected to; rewrite it to the same-family loopback,
# which the wildcard listener also binds. (self.local_ip is wrong
# here — it can resolve to a different family than the listener,
# e.g. IPv6 while the server is bound to 0.0.0.0.)
host = {"0.0.0.0": "127.0.0.1", "::": "::1"}.get(host, host)
bootstrap_na = NetworkAddress(host, self.bootstrap_port)
url = f"{bootstrap_na.to_url()}/route"
@@ -185,8 +185,8 @@ class TestRegisterToBootstrap(CustomTestCase):
@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.
def test_wildcard_host_0000_uses_ipv4_loopback(self, mock_put, mock_time):
"""When --host 0.0.0.0 is used, the PUT must target IPv4 loopback.
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
@@ -196,7 +196,7 @@ class TestRegisterToBootstrap(CustomTestCase):
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.
Fix: substitute same-family loopback when bootstrap_host is a wildcard.
"""
mock_time.monotonic.return_value = 0.0
success_resp = MagicMock()
@@ -210,12 +210,12 @@ class TestRegisterToBootstrap(CustomTestCase):
url_used = mock_put.call_args[0][0]
self.assertNotIn("0.0.0.0", url_used)
self.assertIn("192.168.1.10", url_used)
self.assertIn("127.0.0.1", 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."""
def test_wildcard_host_ipv6_uses_ipv6_loopback(self, mock_put, mock_time):
"""Same fix for the IPv6 wildcard \"::\": must use IPv6 loopback."""
mock_time.monotonic.return_value = 0.0
success_resp = MagicMock()
success_resp.status_code = 200
@@ -227,9 +227,9 @@ class TestRegisterToBootstrap(CustomTestCase):
mgr.register_to_bootstrap()
url_used = mock_put.call_args[0][0]
# "::" bracketed as "[::]:port" should not appear; local_ip should
# "::" bracketed as "[::]:port" should not appear; loopback should.
self.assertNotIn("[::]", url_used)
self.assertIn("fd00", url_used)
self.assertIn("[::1]", url_used)
def _make_manager(self, dist_init_addr=None):
"""Create a lightweight mock manager that has the attributes needed