diff --git a/vllm/patches/deepseek_v4.py b/vllm/patches/deepseek_v4.py
index 36e7ef8e..4c5096cc 100644
--- a/vllm/patches/deepseek_v4.py
+++ b/vllm/patches/deepseek_v4.py
@@ -515,11 +515,12 @@ class DeepseekV4MegaMoEExperts(nn.Module):
         # Set activation global scales from checkpoint input_scale
         # The input_scale is the pre-computed activation normalization factor
         # for each projection. Using hardcoded 1/2688 gives wrong scale for most layers.
+        # Read from the nn.Parameters BEFORE they're freed below.
         # input_scale shape: (num_experts, 2) for w13 (gate, up), (num_experts, 1) for w2
-        l1_igs = self._w13_input_scale[:, 0] if self._w13_input_scale is not None else None  # gate input_scale
-        l2_igs = self._w2_input_scale[:, 0] if self._w2_input_scale is not None else None  # down input_scale
-        self._cutedsl_runner.l1_activation_global_scale = l1_igs.mean().item() if l1_igs is not None else 1.0 / 2688.0
-        self._cutedsl_runner.l2_activation_global_scale = l2_igs.mean().item() if l2_igs is not None else 1.0 / 2688.0
+        l1_igs = self.w13_input_scale.data[:, 0]  # gate input_scale
+        l2_igs = self.w2_input_scale.data[:, 0]   # down input_scale
+        self._cutedsl_runner.l1_activation_global_scale = l1_igs.mean().item()
+        self._cutedsl_runner.l2_activation_global_scale = l2_igs.mean().item()
 
         # Drop the original loader-side parameters
         self._w13_input_scale = self.w13_input_scale.data.clone()