resume and load_from usage

[Ruining0916]

Hi, I am trying to use the resume and load_from features to resume training from my last checkpoint. Here is the issue:

From mmengine saving logic, the checkpoints will be saved as:
/work_dir
-log_dir//log

• last_checkpoint
  -iter_x.pth
  -iter_y.pth

However, the self.resume() or self.load_from() will need get_ckpt_name from deepseed engine: here. From their function, we can see "mp_rank_" + mp_rank_str + "_model_states.pt or f"{filename}mp_rank{mp_rank_str}_model_states.pt" are required to load model states. In addition, the engine load_module_state_dict instead of directly load_state_dict, which requires "module" key in checkpoint, but is missing from current checkpoint file under work_dir.

As only iter_x.pth/iter_y.pth/lats_checkpoint exist under work_dir, get_ckpt_list can not find *_model_states.pt from work_dir, so the current error message looks like below,

I wonder could you help me clarify the correct usage of load_from and resume, I really appreciate your insights!