diff --git a/train.py b/train.py
index 524c7b5..2f49c43 100644
--- a/train.py
+++ b/train.py
@@ -168,7 +168,7 @@ def train(config_path):
         eval_strategy=config.get("eval_strategy", "steps"),
         eval_steps=config.get("eval_steps", 100),
         bf16=True,
-        gradient_checkpointing=True,
+        gradient_checkpointing=False,  # FSDP activation_checkpointing handles it
         fsdp=True,
         fsdp_config={
             "sharding_strategy": "SHARD_GRAD_OP",