diff --git a/train.py b/train.py
index 6994fb9b..61308681 100644
--- a/train.py
+++ b/train.py
@@ -614,7 +614,7 @@ def get_weight_decay(progress):
 # Final summary
 t_end = time.time()
 startup_time = t_start_training - t_start
-steady_state_mfu = 100 * num_flops_per_token * TOTAL_BATCH_SIZE * (step - 10) / total_training_time / H100_BF16_PEAK_FLOPS if total_training_time > 0 else 0
+steady_state_mfu = 100 * num_flops_per_token * TOTAL_BATCH_SIZE * (step - 11) / total_training_time / H100_BF16_PEAK_FLOPS if total_training_time > 0 else 0
 peak_vram_mb = torch.cuda.max_memory_allocated() / 1024 / 1024
 
 print("---")