diff --git a/python/sglang/srt/models/bailing_moe_linear.py b/python/sglang/srt/models/bailing_moe_linear.py index 0c141da70e..8a2173e6dc 100644 --- a/python/sglang/srt/models/bailing_moe_linear.py +++ b/python/sglang/srt/models/bailing_moe_linear.py @@ -237,7 +237,9 @@ class BailingMoE(nn.Module): self.tp_size = get_parallel().tp_size self.top_k = config.num_experts_per_tok - self.norm_expert_prob = getattr(config, "norm_topk_prob", False) + # The reference modeling always renormalizes the top-k weights, and the + # Ling-2.5/2.6 configs carry no norm_topk_prob key. + self.norm_expert_prob = getattr(config, "norm_topk_prob", True) self.hidden_size = config.hidden_size self.intermediate_size = config.moe_intermediate_size self.num_shared_experts = getattr(config, "num_shared_experts", 0)