[BUG] Why in the checkpoints of later stage of training, the global_step number is inconsistent with checkpoint step number?

Describe the bug
I'm training a LLM on a 8-gpu node, using ZeRO3. In the checkpoints of later stage of training, for example after 300 steps when total steps equal to 500, the number of global_step folder inside the checkpoint folder is inconsistent with checkpoint step number. If I continue my training with this checkpoint, later the checkpoints of the new training will amplify this mismatch. The new checkpoints could be "checkpoint-1200" with "global_step1192". And I'm not sure if this mismatch will cause a spike of training loss.

To Reproduce
Steps to reproduce the behavior:

1. Go to later stage checkpoint, e.g. "checkpoint-800"
2. See error, there's a folder: e.g. "global_step799", the step doesn't match.

Expected behavior
In the early stage, this should be "checkpoint-200" with "global_step200".

ds_report output
The screenshot of my file system will be shown beow:

Screenshots

System info (please complete the following information):

• OS: [Red Hat Enterprise Linux 9.5]
• GPU count and types [1 machine with x8 A100s]
• Python version 3.11
• ZeRO3

Launcher context
Are you launching your experiment with the deepspeed launcher, MPI, or something else?
compute_environment: LOCAL_MACHINE
debug: false
deepspeed_config:
deepspeed_multinode_launcher: standard
offload_optimizer_device: none
offload_param_device: none
zero3_init_flag: true
zero3_save_16bit_model: true
zero_stage: 3
gradient_checkpointing: true
gradient_checkpointing_kwargs:
use_reentrant: true
bf16:
enabled: true
distributed_type: DEEPSPEED
downcast_bf16: 'no'
machine_rank: 0
main_training_function: main
mixed_precision: bf16
num_machines: 1
num_processes: 8
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false

Reproduce the report with the stated 8-GPU ZeRO3 configuration and compare checkpoint directory names with their global_step folders, including checkpoint-800/global_step799 and resumed-training examples. Trace the checkpoint save and resume entry points responsible for these names; done means later-stage and resumed checkpoints use consistent step values without altering training behavior.