diff --git a/python/sglang/srt/models/deepseek_v2.py b/python/sglang/srt/models/deepseek_v2.py index 4b5eeb951..064e91cf0 100644 --- a/python/sglang/srt/models/deepseek_v2.py +++ b/python/sglang/srt/models/deepseek_v2.py @@ -245,10 +245,14 @@ class DeepseekV2MLP(nn.Module): tp_rank=tp_rank, tp_size=tp_size, ) - if not hasattr(self.gate_up_proj, "weight"): - self.gate_up_proj.weight = getattr(self.gate_up_proj, "weight_packed") - if not hasattr(self.down_proj, "weight"): - self.down_proj.weight = getattr(self.down_proj, "weight_packed") + if not hasattr(self.gate_up_proj, "weight") and hasattr( + self.gate_up_proj, "weight_packed" + ): + self.gate_up_proj.weight = self.gate_up_proj.weight_packed + if not hasattr(self.down_proj, "weight") and hasattr( + self.down_proj, "weight_packed" + ): + self.down_proj.weight = self.down_proj.weight_packed if hidden_act != "silu": raise ValueError( f"Unsupported activation: {hidden_act}. "