diff --git a/vllm/kernels/linear/nvfp4/cutedsl.py b/vllm/kernels/linear/nvfp4/cutedsl.py
index 367b60a5..55f58b61 100644
--- a/vllm/kernels/linear/nvfp4/cutedsl.py
+++ b/vllm/kernels/linear/nvfp4/cutedsl.py
@@ -84,7 +84,11 @@ class CuTeDSLNvFp4LinearKernel(NvFp4LinearKernel):
         if hasattr(layer, 'input_global_scale_inv') and layer.input_global_scale_inv is not None:
             inv = layer.input_global_scale_inv.data.item()
             if inv != 0:
-                activation_global_scale = 1.0 / inv
+                # input_global_scale_inv = 1.0 / input_global_scale
+                # input_global_scale = 1.0 / nvfp4_global_scale (the dequant scale)
+                # So input_global_scale_inv = nvfp4_global_scale = amax / (6.0 * 448.0)
+                # This is exactly what quantize_activation_nvfp4 expects.
+                activation_global_scale = inv
         runner._activation_global_scale = activation_global_scale
 
         # Register the runner and store the ID (not the runner itself)
diff --git a/vllm/patches/fused_moe/experts/cutedsl_moe.py b/vllm/patches/fused_moe/experts/cutedsl_moe.py
index d8dfa93e..a2b07ad1 100644
--- a/vllm/patches/fused_moe/experts/cutedsl_moe.py
+++ b/vllm/patches/fused_moe/experts/cutedsl_moe.py
@@ -183,20 +183,21 @@ class CuTeDSLMoEExperts(mk.FusedMoEExpertsModular):
             self._runner.set_swiglu_limit(float(swiglu_limit))
 
         # Set initial activation global scales from checkpoint input_scale.
-        # The CuTeDSL runner uses activation_gs = 1.0 / input_scale from the
-        # checkpoint as the starting value. The warmup step
-        # (compute_activation_global_scales) will override this with an
-        # empirically computed value before the first inference.
+        # After undoing the inversion from convert_to_nvfp4_moe_kernel_format,
+        # w13_input_scale_orig = amax / (6.0 * 448.0), which IS the activation
+        # global scale that quantize_activation_nvfp4 expects.
+        # The warmup step (compute_activation_global_scales) will override
+        # this with an empirically computed value before the first inference.
         if w13_input_scale_orig is not None:
-            # input_scale = 448.0 / amax → activation_gs = 1.0 / input_scale = amax / 448.0
+            # w13_input_scale_orig = amax / (6.0 * 448.0) = activation gs
             # Mean across experts (they should be similar)
             mean_l1_gs = float(w13_input_scale_orig.mean().item())
             if mean_l1_gs > 0:
-                self._runner._l1_activation_global_scale = 1.0 / mean_l1_gs
+                self._runner._l1_activation_global_scale = mean_l1_gs
         if w2_input_scale_orig is not None:
             mean_l2_gs = float(w2_input_scale_orig.mean().item())
             if mean_l2_gs > 0:
-                self._runner._l2_activation_global_scale = 1.0 / mean_l2_gs
+                self._runner._l2_activation_global_scale = mean_l2_gs
 
         # Note: activation global scale warmup must be done after
         # process_weights_after_loading, before the first inference.