diff --git a/vllm/nvfp4_cutedsl.py b/vllm/nvfp4_cutedsl.py
index c9542e3a..31b72dc8 100644
--- a/vllm/nvfp4_cutedsl.py
+++ b/vllm/nvfp4_cutedsl.py
@@ -60,8 +60,10 @@ class CuTeDSLMoERunner:
         self._l1_gsb = None
         self._l2_gsb = None
         
-        self._l1_activation_global_scale = None  # set from checkpoint input_scale
-        self._l2_activation_global_scale = None
+        # Default: 1/2688 ≈ 0.000372 (amax=1 → gs=1/2688)
+        # Overridden in finalize_weights with checkpoint input_scale or warmup value
+        self._l1_activation_global_scale = 1.0 / (6.0 * 448.0)
+        self._l2_activation_global_scale = 1.0 / (6.0 * 448.0)
         
         # Pre-allocated cudagraph buffers (set in _allocate_buffers)
         self._token_indices = None