diff --git a/nanochat/optim.py b/nanochat/optim.py
index 190a1ed..ea623fa 100644
--- a/nanochat/optim.py
+++ b/nanochat/optim.py
@@ -217,7 +217,7 @@ class MuonAdamW(torch.optim.Optimizer):
 
             # Fused update: weight_decay -> momentum -> bias_correction -> param_update
             adamw_step_fused(
-                p, grad, exp_avg, exp_avg_sq,
+                p.data, grad, exp_avg, exp_avg_sq,
                 self._adamw_step_t, self._adamw_lr_t, self._adamw_beta1_t,
                 self._adamw_beta2_t, self._adamw_eps_t, self._adamw_wd_t,
             )