diff --git a/train.py b/train.py
index e3fdbed..4f459ce 100644
--- a/train.py
+++ b/train.py
@@ -201,7 +201,7 @@ def train(config_path):
             auto_wrap_policy=get_auto_wrap_policy(model),
             device_id=None,  # Keep on CPU initially
             mixed_precision=None,
-            sync_module_states=True,
+            sync_module_states=False,  # Model is on CPU, no sync needed
             use_orig_params=True,
         )
         print("✓ Model wrapped with FSDP (will be sharded across GPUs during training)")