Add MoE verification for multi-modal models

Signed-off-by: shen-shanshan <[email protected]>

Author: shen-shanshan

vllm_ascend/utils.py

@@ -672,14 +672,24 @@ def prefill_context_parallel_enable() -> bool:
 
 
 def is_moe_model(vllm_config: VllmConfig):
+    """Checks if the model is a MoE model by config"""
     global _IS_MOE_MODEL
     if _IS_MOE_MODEL is None:
-        config = vllm_config.model_config.hf_config
-        _IS_MOE_MODEL = any('experts' in key.lower()
-                            for key in config.to_dict())
+        model_configs = vllm_config.model_config.hf_config.to_dict()
+        _IS_MOE_MODEL = _is_contain_expert(model_configs)
     return _IS_MOE_MODEL
 
 
+def _is_contain_expert(config: dict | str):
+    if isinstance(config, dict):
+        for k, v in config.items():
+            if "expert" in str(k):
+                return True
+            if _is_contain_expert(v):
+                return True
+    return False
+
+
 def weak_ref_tensor(tensor: Any) -> Any:
     """
     Create a weak reference to a tensor.