diff --git a/train.py b/train.py
index 2f49c43..894162f 100644
--- a/train.py
+++ b/train.py
@@ -169,7 +169,7 @@ def train(config_path):
         eval_steps=config.get("eval_steps", 100),
         bf16=True,
         gradient_checkpointing=False,  # FSDP activation_checkpointing handles it
-        fsdp=True,
+        fsdp="full_shard",  # Force FSDP1 (not FSDP2)
         fsdp_config={
             "sharding_strategy": "SHARD_GRAD_OP",
             "cpu_offload": False,