printing scores based on fsdp usage or single gpu

utils/train_utils.py

@@ -153,7 +153,11 @@ def train(model, train_dataloader,eval_dataloader, tokenizer, optimizer, lr_sche
             
             if eval_epoch_loss < best_val_loss:
                 best_val_loss = eval_epoch_loss
-                print(f"best eval loss on epoch {epoch} is {best_val_loss}")
+                if train_config.enable_fsdp:
+                    if rank==0:
+                        print(f"best eval loss on epoch {epoch} is {best_val_loss}")
+                else:
+                    print(f"best eval loss on epoch {epoch} is {best_val_loss}")
             val_loss.append(best_val_loss)
             val_prep.append(eval_ppl)