[megatron] fix eval_iters 1

swift/megatron/model/gpt_bridge.py

@@ -1469,6 +1469,7 @@ def save_weights(self, mg_models, output_dir: str, is_peft_format: bool = False)
                     model_dirs=[args.model_dir],
                     additional_saved_files=self.hf_model.model_meta.additional_saved_files)
             logger.info_if(f'Successfully saved `safetensors` model weights in `{output_dir}`.', cond=is_last_rank())
+        dist.barrier()  # Ensure all weights are saved completely
 
 
 class MultimodalGPTBridge(GPTBridge):

swift/megatron/train/sft.py

@@ -81,7 +81,6 @@ def run(self):
                 images_dir = os.path.join(args.save, 'images')
                 logger.info(f'images_dir: {images_dir}')
                 plot_images(images_dir, args.tensorboard_dir)
-            dist.barrier()  # Ensure all weights are saved completely
 
 
 def megatron_sft_main(args: Optional[Union[List[str], MegatronTrainArguments]] = None):

swift/megatron/trainers/utils.py

@@ -34,6 +34,7 @@ def swift_datasets_provider(train_val_test_num_samples):
         step_batch_size = args.micro_batch_size * data_parallel_size
         # To avoid errors caused by the validation set being insufficient to complete a single step.
         if val_dataset is not None and hasattr(val_dataset, '__len__') and len(val_dataset) < step_batch_size:
+            args.eval_iters = 0
             val_dataset = None
         return train_dataset, val_dataset, None