diff --git a/init2winit/utils.py b/init2winit/utils.py
index 911718b2..66f5181e 100644
--- a/init2winit/utils.py
+++ b/init2winit/utils.py
@@ -229,6 +229,14 @@ def append_scalar_metrics(self, metrics):
     # TODO(gdahl,gilmer): Should this be an atomic file?
     with gfile.GFile(self._csv_path, 'w') as csv_file:
       measurements.to_csv(csv_file, index=False)
+
+    if 'global_step' in metrics:
+      if isinstance(metrics['global_step'], np.ndarray):
+        if metrics['global_step'].shape == (1,):
+          metrics['global_step'] = int(metrics['global_step'][0])
+        else:
+          metrics['global_step'] = int(metrics['global_step'])
+
     if self._xm_work_unit:
       for name, value in metrics.items():
         if name not in self._measurements:
@@ -239,7 +247,7 @@ def append_scalar_metrics(self, metrics):
 
     if self._tb_metric_writer:
       self._tb_metric_writer.write_scalars(
-          step=int(metrics['global_step']), scalars=metrics)
+          step=metrics['global_step'], scalars=metrics)
       # This gives a 1-2% slowdown in steps_per_sec on cifar-10 with batch
       # size 512. We could only flush at the end of training to optimize this.
       self._tb_metric_writer.flush()