Fix by comments.

README.md

@@ -74,10 +74,10 @@ from the latest checkpoint when a failure happens. The actions of flash checkpoi
 <text> The Performance of DLRover Flash Checkpoint to Save/Load GPT2-1.5B.</text>
 </div>
 
-The figure illustrates that the I/O time overhead to read checkpoint files
-when resuming by restarting training processes. With DLRover Flash Checkpoint,
-recovery directly from shared memory takes essentially
-on the order of seconds wich is much faster than SSD and NAS.
+The figure illustrates that the I/O time to read checkpoint files
+when resuming training processes. With DLRover Flash Checkpoint,
+recovery could be completed in the order of seconds by loading checkpoints directly from shared memory,
+which is much faster compared to loading checkpoints from SSD and NAS.
 
 #### Fault Tolerance Improves the Stability of TensorFlow PS Training
 

docs/blogs/flash_checkpoint.md

@@ -307,10 +307,10 @@ Compared to NAS remote file systems, FCP reduces the blocking time by nearly a h
 <text>Figure 4: The Paused Training Time to Save Checkpoint.</text>
 </div>
 
-The figure illustrates that the I/O time overhead to read checkpoint files
-when resuming by restarting training processes. With DLRover Flash Checkpoint,
-recovery directly from shared memory takes essentially
-on the order of seconds wich is much faster than SSD and NAS.
+The figure illustrates that the I/O time to read checkpoint files
+when resuming training processes. With DLRover Flash Checkpoint,
+recovery could be completed in the order of seconds by loading checkpoints directly from shared memory,
+which is much faster compared to loading checkpoints from SSD and NAS.
 
 <div align="center">
 <img src="../figures/ft_llm_training/checkpoint_load_time.jpg" alt="Editor" width="600">