[Bug Fix] get_distributed_init_method should get the ip from get_ip i… (#20889)

Signed-off-by: Chen Li <lcpingping@gmail.com>
Co-authored-by: Russell Bryant <rbryant@redhat.com>
Signed-off-by: Russell Bryant <rbryant@redhat.com>

Author: Relics

vllm/envs.py

@@ -139,6 +139,7 @@
     VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16: bool = True
     VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB: Optional[int] = None
     VLLM_NIXL_ABORT_REQUEST_TIMEOUT: int = 120
+    VLLM_LOOPBACK_IP: str = ""
 
 
 def get_default_cache_root():
@@ -964,6 +965,10 @@ def get_vllm_port() -> Optional[int]:
     # If set to 1, use the TRTLLM Decode Attention backend in flashinfer.
     "VLLM_USE_TRTLLM_DECODE_ATTENTION":
     lambda: os.getenv("VLLM_USE_TRTLLM_DECODE_ATTENTION", None),
+
+    # Used to force set up loopback IP
+    "VLLM_LOOPBACK_IP":
+    lambda: os.getenv("VLLM_LOOPBACK_IP", ""),
 }
 
 # --8<-- [end:env-vars-definition]

vllm/utils/__init__.py

@@ -813,6 +813,33 @@ def get_ip() -> str:
     return "0.0.0.0"
 
 
+def test_loopback_bind(address, family):
+    try:
+        s = socket.socket(family, socket.SOCK_DGRAM)
+        s.bind((address, 0))  # Port 0 = auto assign
+        s.close()
+        return True
+    except OSError:
+        return False
+
+
+def get_loopback_ip() -> str:
+    loopback_ip = envs.VLLM_LOOPBACK_IP
+    if loopback_ip:
+        return loopback_ip
+
+    # VLLM_LOOPBACK_IP is not set, try to get it based on network interface
+
+    if test_loopback_bind("127.0.0.1", socket.AF_INET):
+        return "127.0.0.1"
+    elif test_loopback_bind("::1", socket.AF_INET6):
+        return "::1"
+    else:
+        raise RuntimeError(
+            "Neither 127.0.0.1 nor ::1 are bound to a local interface. "
+            "Set the VLLM_LOOPBACK_IP environment variable explicitly.")
+
+
 def is_valid_ipv6_address(address: str) -> bool:
     try:
         ipaddress.IPv6Address(address)

vllm/v1/executor/multiproc_executor.py

@@ -30,8 +30,8 @@
 from vllm.executor.multiproc_worker_utils import (
     _add_prefix, set_multiprocessing_worker_envs)
 from vllm.logger import init_logger
-from vllm.utils import (get_distributed_init_method, get_mp_context,
-                        get_open_port)
+from vllm.utils import (get_distributed_init_method, get_loopback_ip,
+                        get_mp_context, get_open_port)
 from vllm.v1.executor.abstract import Executor, FailureCallback
 from vllm.v1.outputs import ModelRunnerOutput
 from vllm.worker.worker_base import WorkerWrapperBase
@@ -63,9 +63,9 @@ def _init_executor(self) -> None:
 
         # Multiprocessing-based executor does not support multi-node setting.
         # Since it only works for single node, we can use the loopback address
-        # 127.0.0.1 for communication.
+        # get_loopback_ip() for communication.
         distributed_init_method = get_distributed_init_method(
-            "127.0.0.1", get_open_port())
+            get_loopback_ip(), get_open_port())
 
         # Initialize worker and set up message queues for SchedulerOutputs
         # and ModelRunnerOutputs