Checklist / 检查清单
Bug Description / Bug 描述
设置--use_flash_ckpt true时，训练中途会报错FileNotFoundError: [Errno 2] No such file or directory: ''但是不影响训练
训练结束时会报错
[rank0]: TypeError: CheckpointEngine.wait_latest_checkpoint() takes from 1 to 2 positional arguments but 3 were given
[rank0]: TypeError: HfFlashCheckpointer.save_checkpoint_to_storage() takes 2 positional arguments but 3 were given
查了下dlrover里面的调用是：
save_checkpoint_to_storage(step)
wait_latest_checkpoint(timeout)
而ms-swift里面是trainers/mixin.py：
save_checkpoint_to_storage(step, True)
wait_latest_checkpoint(timeout, max_steps)
How to Reproduce / 如何复现
脚本：
NPROC_PER_NODE=2
CUDA_VISIBLE_DEVICES=0,1
swift sft
--model Qwen/Qwen3-0.6B
--dataset AI-ModelScope/alpaca-gpt4-data-zh#1000
--split_dataset_ratio 0.01
--torch_dtype bfloat16
--tuner_type lora
--target_modules all-linear
--lora_rank 8
--lora_alpha 32
--num_train_epochs 1
--per_device_train_batch_size 1
--per_device_eval_batch_size 1
--gradient_accumulation_steps 8
--learning_rate 1e-4
--max_length 2048
--save_steps 20
--eval_steps 20
--save_total_limit 2
--use_flash_ckpt true
--logging_steps 10
--output_dir output/npu-modelscope-qwen3-0_6b-lora_deepspeed
版本：
ms_swift 4.5.0.dev0
dlrover 0.6.1
Additional Information / 补充信息
No response
Checklist / 检查清单
Bug Description / Bug 描述
设置--use_flash_ckpt true时，训练中途会报错FileNotFoundError: [Errno 2] No such file or directory: ''但是不影响训练
训练结束时会报错
[rank0]: TypeError: CheckpointEngine.wait_latest_checkpoint() takes from 1 to 2 positional arguments but 3 were given
[rank0]: TypeError: HfFlashCheckpointer.save_checkpoint_to_storage() takes 2 positional arguments but 3 were given
查了下dlrover里面的调用是：
save_checkpoint_to_storage(step)
wait_latest_checkpoint(timeout)
而ms-swift里面是trainers/mixin.py：
save_checkpoint_to_storage(step, True)
wait_latest_checkpoint(timeout, max_steps)
How to Reproduce / 如何复现
脚本：
NPROC_PER_NODE=2
CUDA_VISIBLE_DEVICES=0,1
swift sft
--model Qwen/Qwen3-0.6B
--dataset AI-ModelScope/alpaca-gpt4-data-zh#1000
--split_dataset_ratio 0.01
--torch_dtype bfloat16
--tuner_type lora
--target_modules all-linear
--lora_rank 8
--lora_alpha 32
--num_train_epochs 1
--per_device_train_batch_size 1
--per_device_eval_batch_size 1
--gradient_accumulation_steps 8
--learning_rate 1e-4
--max_length 2048
--save_steps 20
--eval_steps 20
--save_total_limit 2
--use_flash_ckpt true
--logging_steps 10
--output_dir output/npu-modelscope-qwen3-0_6b-lora_deepspeed
版本：
ms_swift 4.5.0.dev0
dlrover 0.6.1
Additional Information / 补充信息
No response