Using synthetic dataset_type to decrease the training time

Author: Charles Li

tests/integration/gradient_accumulation_test.py

@@ -65,6 +65,7 @@ def test_grad_accumulate_same_loss(self):
         get_test_config_path(),
         f"base_output_directory={self.base_output_directory}",
         f"dataset_path={self.dataset_path}",
+        "dataset_type=synthetic",
         "gradient_clipping_threshold=0",  # Ensures we are testing raw scales of gradients (clipping off)
         "enable_checkpointing=False",
         "enable_goodput_recording=False",