oft训练好的模型export报错

[zhanghang-official]

Reminder

• I have read the above rules and searched the existing issues.

System Info

• llamafactory version: 0.9.4.dev0
• Platform: Linux-5.4.0-139-generic-x86_64-with-glibc2.31
• Python version: 3.10.18
• PyTorch version: 2.6.0+cu124 (GPU)
• Transformers version: 4.51.3
• Datasets version: 3.6.0
• Accelerate version: 1.7.0
• PEFT version: 0.17.0
• TRL version: 0.9.6
• GPU type: NVIDIA A800-SXM4-80GB
• GPU number: 8
• GPU memory: 79.32GB
• DeepSpeed version: 0.16.9
• vLLM version: 0.8.2
• Default data directory: detected

Reproduction

训练脚本：
### model
model_name_or_path: /raid/zhanghang02/weights/pretrained/Qwen2.5-VL-7B-Instruct
image_max_pixels: 262144
video_max_pixels: 262144
video_fps: 2
video_maxlen: 128
trust_remote_code: true
deepspeed: examples/deepspeed/ds_z2_config.json

### method
stage: sft
do_train: true
finetuning_type: oft
freeze_vision_tower: true
freeze_multi_modal_projector: true
freeze_language_model: false
flash_attn: fa2
oft_block_size: 128
oft_target: all

### dataset
dataset: caption_image16_train_250825
template: qwen2_vl
cutoff_len: 40960
max_samples: 1000000
# max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 32
dataloader_num_workers: 8
dataloader_pin_memory: true

### output
output_dir: /raid/zhanghang02/weights/checkpiont/caption/caption_qwen25vl_p512_image_250825_oft
logging_steps: 1
save_steps: 500
plot_loss: true
overwrite_output_dir: true
save_only_model: false
report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

### train
per_device_train_batch_size: 4
gradient_accumulation_steps: 1
learning_rate: 1.0e-4
num_train_epochs: 8.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
resume_from_checkpoint: null

## eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500

# export脚本
### model
model_name_or_path: /raid/zhanghang02/weights/pretrained/Qwen2.5-VL-7B-Instruct
adapter_name_or_path: /raid/zhanghang02/weights/checkpiont/caption/caption_qwen25vl_p512_image_250825_oft/epoch1
trust_remote_code: true

### method
finetuning_type: oft

### dataset
template: qwen2_vl

### output
export_dir: /raid/zhanghang02/weights/export/caption/caption_qwen25vl_p512_image_250825_oft/caption_qwen25vl_p512_image_250825_oft_epoch1
export_size: 5
export_device: cpu
export_legacy_format: False


报错如下：
(qwen25vl_oft) [~/factory/llamafactory]: DISABLE_VERSION_CHECK=1 CUDA_VISIBLE_DEVICES=0 llamafactory-cli export \
>     task/export/caption/caption_qwen25vl_p512_image_250825_oft/caption_qwen25vl_p512_image_250825_oft_epoch7.yaml
[2025-09-04 16:13:37,390] [INFO] [real_accelerator.py:254:get_accelerator] Setting ds_accelerator to cuda (auto detect)
Warning: The cache directory for DeepSpeed Triton autotune, /home/zhanghang02/.triton/autotune, appears to be on an NFS system. While thi
s is generally acceptable, if you experience slowdowns or hanging when DeepSpeed exits, it is recommended to set the TRITON_CACHE_DIR env
ironment variable to a non-NFS path.                                
[WARNING|2025-09-04 16:13:39] llamafactory.extras.misc:154 >> Version checking has been disabled, may lead to unexpected behaviors.
/home/zhanghang02/anaconda3/envs/qwen25vl_oft/lib/python3.10/site-packages/jieba/_compat.py:18: UserWarning: pkg_resources is deprecated 
as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-
11-30. Refrain from using this package or pin to Setuptools<81.
  import pkg_resources                                              
INFO 09-04 16:13:40 [__init__.py:239] Automatically detected platform cuda.                                                              
Traceback (most recent call last):                                  
  File "/home/zhanghang02/anaconda3/envs/qwen25vl_oft/bin/llamafactory-cli", line 7, in <module>
    sys.exit(main())                                                
  File "/home/zhanghang02/anaconda3/envs/qwen25vl_oft/lib/python3.10/site-packages/llamafactory/cli.py", line 151, in main
    COMMAND_MAP[command]()                                          
  File "/home/zhanghang02/anaconda3/envs/qwen25vl_oft/lib/python3.10/site-packages/llamafactory/train/tuner.py", line 114, in export_mode
l
    model_args, data_args, finetuning_args, _ = get_infer_args(args)                                                                     
  File "/home/zhanghang02/anaconda3/envs/qwen25vl_oft/lib/python3.10/site-packages/llamafactory/hparams/parser.py", line 446, in get_infe
r_args
    _verify_model_args(model_args, data_args, finetuning_args)                                                                           
  File "/home/zhanghang02/anaconda3/envs/qwen25vl_oft/lib/python3.10/site-packages/llamafactory/hparams/parser.py", line 112, in _verify_
model_args
    raise ValueError("Adapter is only valid for the LoRA method.")                                                                       
ValueError: Adapter is only valid for the LoRA method.                     

Others

No response

[zhanghang-official]

oft的checkpoint内容如下