FSDP2 doesn't save HF checkpoint and doesn't use parameters from fsdp_config

Checklist / 检查清单

• I have searched existing issues, and this is a new bug report. / 我已经搜索过现有的 issues，确认这是一个新的 bug report。

Bug Description / Bug 描述

when flag --fsdp fsdp2 used, there is no HF checkpoint saved during saving, only pytorch_model_fsdp_0 and optimizer_0 along with rng_states and scheduler.pt

Also I've tried to pass my own json with fsdp_config to change state_dict_type:

    "fsdp": "shard_grad_op auto_wrap",
    "fsdp_config": {
        "fsdp_version": 2,
        "reshard_after_forward": false,
        "auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
        "cpu_ram_efficient_loading": true,
        "state_dict_type": "FULL_STATE_DICT",
        "activation_checkpointing": true
    }

But nothing was changed in terms of speed, memory consumption and saving behavior, so I've checked source code and noticed that only parameter that used is fsdp_version

After that I've tried to change state_dict_type via env variable

os.environ["FSDP_STATE_DICT_TYPE"] = "FULL_STATE_DICT"

And it worked for common models like qwen, but when I used my custom model that I've registered via ms-swift, my model_modeling.py file was not saved. Such behavior only occurs when I use fsdp2, and persistent in both versions of ms-swift 4.0.0 and 3.12.6, when I use deepspeed everything works fine. Also in version 3.12.6 I could pass config for fsdp1 which looked like that

fsdp: "shard_grad_op auto_wrap"
fsdp_config:
  fsdp_version: 1
  backward_prefetch: "backward_pre"
  forward_prefetch: false
  limit_all_gathers: true
  use_orig_params: true
  sync_module_states: true
  cpu_ram_efficient_loading: true
  activation_checkpointing: true
  min_num_params: 0

And it also worked as intended and saved hf checkpoint along everything else

How to Reproduce / 如何复现

just launch example of train with fsdp2 and wait for saving

Additional Information / 补充信息

This bug persistent in newest 4.0.0 and previous 3.12.6 versions of ms-swift
In both cases I've used torch==2.9.1 and cuda==12.8

Start with the training example using --fsdp fsdp2, then trace how fsdp_config and the FSDP_STATE_DICT_TYPE environment variable are handled. Compare behavior with the provided FSDP1 configuration and a registered custom model's model_modeling.py; done means the configuration affects saving and an HF checkpoint is produced alongside the existing shard files.