[Cherry-pick] [0.11.0] pd proxy support ipv6 and fix proxy (#4242)

### What this PR does / why we need it?
pd proxy support ipv6, mooncake connector check whether the IPv6 address
is used and notify the user.

---------

Signed-off-by: liziyu <[email protected]>

Author: liziyu179

examples/disaggregated_prefill_v1/load_balance_proxy_layerwise_server_example.py

@@ -88,6 +88,7 @@
 import asyncio
 import functools
 import heapq
+import ipaddress
 import os
 import sys
 import threading
@@ -116,6 +117,12 @@ def __init__(self, host, port):
         self.host = host
         self.port = port
         self.url = f'http://{host}:{port}/v1'
+        try:
+            ip = ipaddress.ip_address(self.host)
+            if isinstance(ip, ipaddress.IPv6Address):
+                self.url = f'http://[{host}]:{port}/v1'
+        except Exception:
+            pass
         self.client = httpx.AsyncClient(timeout=None,
                                         base_url=self.url,
                                         limits=httpx.Limits(
@@ -356,6 +363,9 @@ async def send_request_to_service(client: httpx.AsyncClient,
     req_data = req_data.copy()
     req_data["stream"] = False
     req_data["max_tokens"] = 1
+    req_data["min_tokens"] = 1
+    if "max_completion_tokens" in req_data:
+        req_data["max_completion_tokens"] = 1
     if "stream_options" in req_data:
         del req_data["stream_options"]
     headers = {

examples/disaggregated_prefill_v1/load_balance_proxy_server_example.py

@@ -88,6 +88,7 @@
 import asyncio
 import functools
 import heapq
+import ipaddress
 import json
 import os
 import sys
@@ -118,6 +119,12 @@ def __init__(self, host, port):
         self.host = host
         self.port = port
         self.url = f'http://{host}:{port}/v1'
+        try:
+            ip = ipaddress.ip_address(self.host)
+            if isinstance(ip, ipaddress.IPv6Address):
+                self.url = f'http://[{host}]:{port}/v1'
+        except Exception:
+            pass
         self.client = httpx.AsyncClient(timeout=None,
                                         base_url=self.url,
                                         limits=httpx.Limits(
@@ -366,6 +373,8 @@ async def send_request_to_service(client: httpx.AsyncClient,
     req_data["stream"] = False
     req_data["max_tokens"] = 1
     req_data["min_tokens"] = 1
+    if "max_completion_tokens" in req_data:
+        req_data["max_completion_tokens"] = 1
     if "stream_options" in req_data:
         del req_data["stream_options"]
     headers = {

vllm_ascend/distributed/mooncake/transfer_engine.py

@@ -1,3 +1,4 @@
+import ipaddress
 import threading
 from typing import Optional
 
@@ -8,6 +9,15 @@
 
 
 def get_global_te(hostname: str, device_name: Optional[str]):
+    try:
+        ip = ipaddress.ip_address(hostname)
+        if isinstance(ip, ipaddress.IPv6Address):
+            raise RuntimeError(
+                "The backend of mooncake's Ascend Direct Xfer Library currently does not support IPv6."
+            )
+    except ValueError:
+        pass
+
     global _global_te
     if _global_te is None:
         with _global_te_lock:

vllm_ascend/envs.py

@@ -162,11 +162,6 @@
     # Whether to enable msMonitor tool to monitor the performance of vllm-ascend.
     "MSMONITOR_USE_DAEMON":
     lambda: bool(int(os.getenv("MSMONITOR_USE_DAEMON", '0'))),
-    # Timeout (in seconds) for delayed KVCache block release. In the prefill
-    # node, if a request is marked for delayed KV block release and the blocks
-    # are not freed within this timeout, they will be forcibly released.
-    "VLLM_ASCEND_KVCACHE_DELAY_FREE_TIMEOUT":
-    lambda: int(os.getenv("VLLM_ASCEND_KVCACHE_DELAY_FREE_TIMEOUT", 250)),
     "VLLM_ASCEND_ENABLE_MLAPO":
     lambda: bool(int(os.getenv("VLLM_ASCEND_ENABLE_MLAPO", '0'))),
     # Whether to enable transpose weight and cast format to FRACTAL_NZ.