diff --git a/vllm/patches/fused_moe/experts/cutedsl_moe.py b/vllm/patches/fused_moe/experts/cutedsl_moe.py
index 0f56b3d4..5a8398ad 100644
--- a/vllm/patches/fused_moe/experts/cutedsl_moe.py
+++ b/vllm/patches/fused_moe/experts/cutedsl_moe.py
@@ -148,12 +148,12 @@ class CuTeDSLMoEExperts(mk.FusedMoEExpertsModular):
         # apply() before delegating to our expert impl, so we can't set the
         # weights to None. Instead, replace with a shape-preserving dummy on CPU
         # to free GPU memory while keeping the shape metadata accessible.
-        layer.w13_weight = torch.empty(
+        layer.w13_weight = torch.nn.Parameter(torch.empty(
             num_experts, 2 * intermediate_size, hidden_size // 2,
-            device='cpu', dtype=torch.uint8)
-        layer.w2_weight = torch.empty(
+            device='cpu', dtype=torch.uint8), requires_grad=False)
+        layer.w2_weight = torch.nn.Parameter(torch.empty(
             num_experts, hidden_size, intermediate_size // 2,
-            device='cpu', dtype=torch.uint8)
+            device='cpu', dtype=torch.uint8), requires_grad=False)
         layer.w13_weight_scale = None
         layer.w2_weight_scale = None
         layer.w13_weight_scale_2 = None