fix: minor fix for modelopt weight load compatibility (#7953)

python/sglang/srt/layers/moe/fused_moe_triton/layer.py

@@ -518,6 +518,7 @@ class FusedMoE(torch.nn.Module):
                 self.quant_method.enable_flashinfer_moe = self.enable_flashinfer_moe
         assert self.quant_method is not None
 
+        self.quant_config = quant_config
         self.quant_method.create_weights(
             layer=self,
             num_experts=self.local_num_experts,
@@ -661,7 +662,11 @@ class FusedMoE(torch.nn.Module):
         ):
             raise ValueError("expert_data and loaded_weight must be torch.Tensor")
 
-        if expert_data.dim() != 2 or loaded_weight.dim() != 2:
+        if (
+            self.quant_config is not None
+            and "modelopt" in self.quant_config.get_name()
+            and (expert_data.dim() != 2 or loaded_weight.dim() != 2)
+        ):
             raise ValueError(
                 f"Expected 2D tensors, got expert_data shape {expert_data.shape} and loaded_weight shape {loaded_weight.shape}"
             )