fix(CRITICAL): launch_server.py覆盖所有vllm文件+运行xformers patch

varlen_fwd()崩溃是因为xformers patch只打到了corex-3.2.3/lib64路径,
但runtime加载的是corex/lib/python3路径(PYTHONPATH优先级更高)。

launch_server.py现在覆盖所有关键文件到每个vllm安装:
- qwen3_5.py, paged_attn.py, model_runner.py等
- 12个prebuilt .so
- 运行patch_xformers_sdpa_seq.py等源码patch脚本

qwen3_6_scripts/launch_server.py

@@ -12,31 +12,88 @@ def _force_patch():
     patched = set()
     
     for p in sys.path:
-        api = os.path.join(p, "vllm", "entrypoints", "openai", "api_server.py")
+        vllm_root = os.path.join(p, "vllm")
-        if os.path.isfile(api) and api not in patched:
+        api = os.path.join(vllm_root, "entrypoints", "openai", "api_server.py")
-            for f in ["api_server.py", "cli_args.py", "serving_chat.py", 
+        if not os.path.isfile(api) or api in patched:
-                       "protocol.py", "serving_tokenization.py"]:
+            continue
-                src = os.path.join(src_dir, f)
+        
-                dst = os.path.join(p, "vllm", "entrypoints", "openai", f)
+        # --- Entrypoints ---
-                if os.path.isfile(src):
+        for f in ["api_server.py", "cli_args.py", "serving_chat.py", 
-                    shutil.copy2(src, dst)
+                   "protocol.py", "serving_tokenization.py"]:
-            # chat_utils
+            src = os.path.join(src_dir, f)
-            cu_src = os.path.join(src_dir, "chat_utils.py")
+            dst = os.path.join(vllm_root, "entrypoints", "openai", f)
-            cu_dst = os.path.join(p, "vllm", "entrypoints", "chat_utils.py")
+            if os.path.isfile(src):
-            if os.path.isfile(cu_src):
+                shutil.copy2(src, dst)
-                shutil.copy2(cu_src, cu_dst)
+        # chat_utils
-            # reasoning
+        cu_src = os.path.join(src_dir, "chat_utils.py")
-            reason_src = os.path.join(src_dir, "reasoning")
+        cu_dst = os.path.join(vllm_root, "entrypoints", "chat_utils.py")
-            reason_dst = os.path.join(p, "vllm", "reasoning")
+        if os.path.isfile(cu_src):
-            if os.path.isdir(reason_src):
+            shutil.copy2(cu_src, cu_dst)
-                shutil.copytree(reason_src, reason_dst, dirs_exist_ok=True)
+        # reasoning
-            # tool parser
+        reason_src = os.path.join(src_dir, "reasoning")
-            tp_src = os.path.join(src_dir, "qwen3coder_tool_parser.py")
+        reason_dst = os.path.join(vllm_root, "reasoning")
-            tp_dst = os.path.join(p, "vllm", "entrypoints", "openai", 
+        if os.path.isdir(reason_src):
-                                  "tool_parsers", "qwen3coder_tool_parser.py")
+            shutil.copytree(reason_src, reason_dst, dirs_exist_ok=True)
-            if os.path.isfile(tp_src) and os.path.isdir(os.path.dirname(tp_dst)):
+        # tool parser
-                shutil.copy2(tp_src, tp_dst)
+        tp_src = os.path.join(src_dir, "qwen3coder_tool_parser.py")
-            patched.add(api)
+        tp_dst = os.path.join(vllm_root, "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)
+        
+        # --- Model, attention, engine files (critical for runtime) ---
+        model_dir = os.path.join(vllm_root, "model_executor", "models")
+        attn_dir = os.path.join(vllm_root, "attention", "ops")
+        core_dir = os.path.join(vllm_root, "core")
+        for fname, dst_dir in [
+            ("qwen3_5.py", model_dir),
+            ("mamba_cache.py", model_dir),
+            ("registry.py", model_dir),
+            ("_custom_ops.py", vllm_root),
+            ("paged_attn.py", attn_dir),
+            ("sequence.py", vllm_root),
+            ("scheduler.py", core_dir),
+            ("model_runner.py", os.path.join(vllm_root, "worker")),
+            ("bi100_env.py", vllm_root),
+            ("bi100_profile.py", vllm_root),
+            ("block_major_kv_cache.py", vllm_root),
+            ("gdn_prefix.py", vllm_root),
+            ("logits_processor.py", os.path.join(vllm_root, "model_executor", "layers")),
+            ("sampler.py", os.path.join(vllm_root, "model_executor", "layers")),
+        ]:
+            src = os.path.join(src_dir, fname)
+            if os.path.isfile(src) and os.path.isdir(dst_dir):
+                shutil.copy2(src, os.path.join(dst_dir, fname))
+        
+        # --- Prebuilt .so files ---
+        prebuilt_dir = os.path.join(src_dir, "prebuilt", "corex-3.2.3-ivcore10")
+        if os.path.isdir(prebuilt_dir):
+            for so_file in os.listdir(prebuilt_dir):
+                if so_file.endswith(".so"):
+                    src_so = os.path.join(prebuilt_dir, so_file)
+                    dst_so = os.path.join(vllm_root, so_file)
+                    if not os.path.isfile(dst_so):
+                        shutil.copy2(src_so, dst_so)
+        
+        # --- Run Python source patches on this vllm install ---
+        for patch_script in [
+            "patch_xformers_sdpa_seq.py",
+            "patch_xformers_profile.py",
+            "patch_model_runner.py",
+            "patch_vllm_qwen3_5.py",
+            "patch_corex_swap_blocks.py",
+        ]:
+            script_path = os.path.join(src_dir, patch_script)
+            if os.path.isfile(script_path):
+                try:
+                    import subprocess
+                    subprocess.run([sys.executable, script_path],
+                                   cwd=src_dir, timeout=30,
+                                   capture_output=True)
+                except Exception:
+                    pass
+        
+        patched.add(api)
     
     if patched:
         print(f"[launch] Force-patched {len(patched)} vllm installs", file=sys.stderr)