diff --git a/src/nvfp4_megamoe_kernel/nvfp4_mega_moe.py b/src/nvfp4_megamoe_kernel/nvfp4_mega_moe.py
index 8bfa6e3a..2a622026 100644
--- a/src/nvfp4_megamoe_kernel/nvfp4_mega_moe.py
+++ b/src/nvfp4_megamoe_kernel/nvfp4_mega_moe.py
@@ -426,18 +426,18 @@ def nvfp4_mega_moe_full(
             w_deq = w_signs * w_mags  # (K, N) = (7168, 6144)
             w_sf_exp = l1_sf[e0].to(torch.float32).repeat_interleave(16, dim=0)  # (K, N)
             # Full dequant: x = e2m1 * block_sf * igs, w = e2m1 * block_sf * gs
-            gs = l1_global_sf[e0]
+            gs = l1_global_sf[e0]  # shape (2,) or scalar
             igs = l1_alpha  # already the input global scale
             x_full = (x_deq * sf_exp * igs).to(torch.bfloat16)  # (K,)
             w_full = (w_deq * w_sf_exp).to(torch.bfloat16)  # (K, N) without gs
             ref_out = torch.nn.functional.linear(x_full.unsqueeze(0), w_full.T).squeeze(0)  # (N,)
             # Apply per-half global scale (gate_gs for first half, up_gs for second half)
             gn = ref_out.shape[0] // 2
-            if gs.dim() == 0:
-                ref_out = ref_out * gs.item()
+            if gs.numel() == 1:
+                ref_out = ref_out * float(gs)
             else:
-                ref_out[:gn] = ref_out[:gn] * gs[0].item()
-                ref_out[gn:] = ref_out[gn:] * gs[1].item()
+                ref_out[:gn] = ref_out[:gn] * float(gs[0])
+                ref_out[gn:] = ref_out[gn:] * float(gs[1])
             nvfp4_mega_moe_full._ref_l1 = (s0, e0, ref_out)
             print(f"[BF16-REF-L1] expert={e0} amax={ref_out.abs().max():.4e} mean={ref_out.mean():.4e}")
         except Exception as ex: