在checkpoint上继续训练，没有保存训练后的checkpint

[cuisws]

Reminder

• I have read the README and searched the existing issues.

System Info

Python version: 3.10.8
PyTorch version: 2.3.0+cu121 (GPU)

Reproduction

model

model_name_or_path: model/MiniCPM3-4B
resume_from_checkpoint: saves/checkpoint-5748

method

stage: sft
do_train: true
finetuning_type: lora

ddp

ddp_timeout: 180000000
deepspeed: examples/deepspeed/ds_z2_config.json

dataset

dataset: sft_zh_demo
template: cpm3
cutoff_len: 1800
max_samples: 500000
overwrite_cache: true
preprocessing_num_workers: 16

output

output_dir: saves
logging_steps: 20
save_strategy: epoch
plot_loss: true
overwrite_output_dir: true

train

训练时的batch_size

per_device_train_batch_size: 8

梯度累计次数

gradient_accumulation_steps: 8
learning_rate: 0.0001
num_train_epochs: 1.0
lr_scheduler_type: cosine

预热步数

warmup_steps: 10
bf16: true

eval

val_size: 0.1

测试时的batch_size

per_device_eval_batch_size: 8
eval_strategy: steps

500个step测试一次

eval_steps: 100

Expected behavior

我使用resume_from_checkpoint从断点处开始训练，训练1轮，
训练过程中没有输出logging_step日志，训练完后没有保存新的checkpoint，这是怎么回事？

Others

No response

[cuisws]

[cuisws]

数据集需要是之前的，或者大于之前的