[BUG] load checkpoint

Describe the bug

When training with DeepSpeed Zero-1, I use model_engine.save_checkpoint(save_dir, tag=tag) to save the model.
The save_dir like

bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt  bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt  bf16_zero_pp_rank_4_mp_rank_00_optim_states.pt  bf16_zero_pp_rank_6_mp_rank_00_optim_states.pt  final_scores.json
bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt  bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt  bf16_zero_pp_rank_5_mp_rank_00_optim_states.pt  bf16_zero_pp_rank_7_mp_rank_00_optim_states.pt  mp_rank_00_model_states.pt

During inference, I refer to thie tutorial,

ds_infer_config =  {
    "type": "ds_model",
    "version": 0.0,
    "checkpoints": "save_dir",
}


model_engine = deepspeed.init_inference(
    model,
    checkpoint=ds_infer_config,
)

model_engine.eval()

But the checkpoint is not loaded. Because for the same data, the output of all checkpoints I load is consistent. What's the problem here?

To Reproduce
deepspeed 0.14.4

Start with the DeepSpeed inference tutorial and the init_inference call using the ds_model checkpoint configuration, then compare it with the Zero-1 save_checkpoint output shown in the issue. Reproduce the load with DeepSpeed 0.14.4 and verify whether inference uses the saved checkpoint rather than producing identical outputs across checkpoints.