fix(runtime): launch_server.py强制覆盖所有vllm路径后启动

根因：patch_ops.sh部署到VLLM_ROOT(lib64)，但python3 -m vllm走的是
/usr/local/corex/lib/python3/dist-packages/vllm/（未被覆盖的路径）
导致基础镜像原版api_server.py运行，不识别qwen3_coder/reasoning-parser

修复：launch_server.py在import前遍历sys.path所有vllm安装，
用shutil.copy2强制覆盖api_server/cli_args/serving_chat等
然后from vllm.entrypoints.openai.api_server import *启动

computility-run.yaml

@@ -1,8 +1,7 @@
 concurrency: 1
 command:
     - python3
-    - -m
-    - vllm.entrypoints.openai.api_server
+    - /workspace/qwen3_6_scripts/launch_server.py
     - --model
     - /model
     - --served-model-name

qwen3_6_scripts/launch_server.py

@@ -0,0 +1,49 @@
+#!/usr/bin/env python3
+"""launch_server.py — Ensure our patched api_server.py runs, not the base image's.
+
+Patches the RUNNING vllm install's api_server.py/cli_args.py in-place before
+importing, then delegates to the standard vllm api_server main().
+"""
+import os, sys, shutil
+
+def _force_patch():
+    """Copy our files over ALL vllm installs found on sys.path."""
+    src_dir = os.path.dirname(os.path.abspath(__file__))
+    patched = set()
+    
+    for p in sys.path:
+        api = os.path.join(p, "vllm", "entrypoints", "openai", "api_server.py")
+        if os.path.isfile(api) and api not in patched:
+            for f in ["api_server.py", "cli_args.py", "serving_chat.py", 
+                       "protocol.py", "serving_tokenization.py"]:
+                src = os.path.join(src_dir, f)
+                dst = os.path.join(p, "vllm", "entrypoints", "openai", f)
+                if os.path.isfile(src):
+                    shutil.copy2(src, dst)
+            # chat_utils
+            cu_src = os.path.join(src_dir, "chat_utils.py")
+            cu_dst = os.path.join(p, "vllm", "entrypoints", "chat_utils.py")
+            if os.path.isfile(cu_src):
+                shutil.copy2(cu_src, cu_dst)
+            # reasoning
+            reason_src = os.path.join(src_dir, "reasoning")
+            reason_dst = os.path.join(p, "vllm", "reasoning")
+            if os.path.isdir(reason_src):
+                shutil.copytree(reason_src, reason_dst, dirs_exist_ok=True)
+            # tool parser
+            tp_src = os.path.join(src_dir, "qwen3coder_tool_parser.py")
+            tp_dst = os.path.join(p, "vllm", "entrypoints", "openai", 
+                                  "tool_parsers", "qwen3coder_tool_parser.py")
+            if os.path.isfile(tp_src) and os.path.isdir(os.path.dirname(tp_dst)):
+                shutil.copy2(tp_src, tp_dst)
+            patched.add(api)
+    
+    if patched:
+        print(f"[launch] Force-patched {len(patched)} vllm installs", file=sys.stderr)
+    else:
+        print("[launch] WARNING: no vllm installs found to patch", file=sys.stderr)
+
+_force_patch()
+
+# Now run the standard vllm api_server with our patched version
+from vllm.entrypoints.openai.api_server import *