diff --git a/python/sglang/srt/model_loader/weight_utils.py b/python/sglang/srt/model_loader/weight_utils.py
index 965123fcc..fb96ff210 100644
--- a/python/sglang/srt/model_loader/weight_utils.py
+++ b/python/sglang/srt/model_loader/weight_utils.py
@@ -1664,8 +1664,11 @@ def initialize_dummy_weights(
     is fixed, the random values generated by this function only depends on
     the parameter's number of elements and its data type.
     """
-    for param in model.state_dict().values():
+    for name, param in model.state_dict().items():
         if torch.is_floating_point(param):
+            if name.endswith("weight_scale_inv"):
+                param.fill_(1.0)
+                continue
             generator = torch.Generator(device=param.data.device)
             generator.manual_seed(seed)
             # Tensor subclasses such as MXFP8 wrappers expose a low-bit raw