diff --git a/vllm/transformers_utils/model_arch_config_convertor.py b/vllm/transformers_utils/model_arch_config_convertor.py index 861633b06c..2619bd4f30 100644 --- a/vllm/transformers_utils/model_arch_config_convertor.py +++ b/vllm/transformers_utils/model_arch_config_convertor.py @@ -172,7 +172,10 @@ class ModelArchConfigConvertorBase: "moe_topk", "moe_top_k", ] - return getattr_iter(self.hf_text_config, names, 0) or 0 + num_experts_per_token = getattr_iter(self.hf_text_config, names, 0) + if isinstance(num_experts_per_token, list): + return max(num_experts_per_token, default=0) + return num_experts_per_token or 0 @final @classmethod