fix(CRITICAL): moe_topk_softmax用importlib加载+patch_ops编译部署

- _custom_ops.py: 改用importlib.util加载.so替代torch.ops.load_library
- patch_ops.sh: docker build时编译moe_topk_softmax_v3 CUDA kernel
- 真机验证: weights sum=1.0, GPU test passed

qwen3_6_scripts/_custom_ops.py

@@ -1054,9 +1054,11 @@ def _init_moe_topk():
     for pattern in so_patterns:
         for so_path in glob.glob(pattern):
             try:
-                torch.ops.load_library(so_path)
-                # After load_library, the pybind module should be importable
-                import moe_topk_softmax_v3 as ext
+                import importlib.util
+                spec = importlib.util.spec_from_file_location(
+                    "moe_topk_softmax_v3", so_path)
+                ext = importlib.util.module_from_spec(spec)
+                spec.loader.exec_module(ext)
                 _moe_topk_ext = ext
                 logger.info("topk_softmax: loaded CUDA kernel from %s", so_path)
                 return

qwen3_6_scripts/patch_ops.sh

@@ -179,6 +179,11 @@ fi
 build_stage "installing hash-pinned CoreX 3.2.3 extensions"
 bash ./install_prebuilt_corex.sh "${VLLM_ROOT}"
 
+build_stage "compiling moe_topk_softmax CUDA kernel"
+cd /workspace && bash ex_engine/build_moe_topk.sh 2>&1 || echo "[WARN] moe_topk build failed (non-fatal)"
+# Deploy to workspace search path (_custom_ops.py looks in /workspace/ex_engine/build/)
+cd "${OLDPWD}"
+
 build_stage "installing BI100 runtime modules"
 cp ./bi100_env.py "${VLLM_ROOT}/bi100_env.py"
 cp ./bi100_profile.py "${VLLM_ROOT}/bi100_profile.py"