[BugFix] Fix fsdp option on shard_grad_op. (#19131)

src/transformers/training_args.py

@@ -1197,7 +1197,7 @@ class TrainingArguments:
                 "`--fsdp offload` can't work on its own. It needs to be added to `--fsdp full_shard` or "
                 '`--fsdp shard_grad_op`. For example, `--fsdp "full_shard offload"`.'
             )
-        elif FSDPOption.FULL_SHARD in self.fsdp and FSDPOption.SHARD_GRAD_OP in self.sharded_ddp:
+        elif FSDPOption.FULL_SHARD in self.fsdp and FSDPOption.SHARD_GRAD_OP in self.fsdp:
             raise ValueError("`--fsdp full_shard` is not compatible with `--fsdp shard_grad_op`.")
 
         if len(self.fsdp) == 0 and self.fsdp_min_num_params > 0: