fix: 在模块加载时进行vllm_npu_C的导入检查，以避免torch.compile/dynamo跟踪失败

vllm_npu/utils.py

@@ -47,7 +47,15 @@ REGISTERED_ASCEND_OPS = {}
 ACL_FORMAT_FRACTAL_ND = 2
 ACL_FORMAT_FRACTAL_NZ = 29
 
-_CUSTOM_OP_ENABLED = None
+# Pre-check for vllm_npu_C at module load time so that torch.compile/dynamo
+# never encounters a failing import during tracing.
+try:
+    import vllm_npu.vllm_npu_C  # type: ignore  # noqa: F401
+    import vllm_npu.meta_registration  # type: ignore  # noqa: F401
+    _CUSTOM_OP_ENABLED = True
+except (ImportError, ModuleNotFoundError):
+    _CUSTOM_OP_ENABLED = False
+
 _IS_310P = None
 _SLEEP_MODE_ENABLED = None
 _CURRENT_STREAM = None
@@ -219,25 +227,10 @@ def try_register_lib(lib_name: str, lib_info: str = ""):
 
 def enable_custom_op():
     """
-    Enable lazy init for vllm_npu_C to avoid early initialization of CANN's RTS component.
+    Check if vllm_npu_C custom ops are available.
-    Ensure that ASCEND_RT_VISIBLE_DEVICES can be dynamically modified before torch.npu.set_device().
+    The import check is done at module load time to avoid torch.compile/dynamo
+    tracing failures.
     """
-    global _CUSTOM_OP_ENABLED
-    if _CUSTOM_OP_ENABLED is not None:
-        return _CUSTOM_OP_ENABLED
-    try:
-        # isort: off
-        # register custom ops into torch_library here
-        import vllm_npu.vllm_npu_C  # type: ignore  # noqa: F401
-        # register the meta implementation for custom kernel if necessary
-        import vllm_npu.meta_registration  # type: ignore  # noqa: F401
-        # isort: on
-        _CUSTOM_OP_ENABLED = True
-    except ImportError:
-        _CUSTOM_OP_ENABLED = False
-        logger.warning(
-            "Warning: Failed to register custom ops, all custom ops will be disabled"
-        )
     return _CUSTOM_OP_ENABLED