Unsloth problem with LoRA

[hienhayho]

Describe the bug
I ran SFT on Qwen2.5-VL-3B-Instruct using LoRA, but Unsloth still indicates that full fine-tuning is being performed.

My script:

swift sft \
    --model Qwen/Qwen2.5-VL-3B-Instruct \
    --tuner_backend unsloth \
    --train_type lora \
    --dataset "$dataset_path" \
    --split_dataset_ratio 0 \
    --torch_dtype bfloat16 \
    --num_train_epochs "$epochs" \
    --per_device_train_batch_size "$batch_size" \
    --learning_rate 1e-5 \
    --gradient_accumulation_steps "$gradient_accumulation_steps" \
    --save_steps 1000 \
    --report_to wandb \
    --save_total_limit 2 \
    --logging_steps 50 \
    --max_length "$max_length" \
    --warmup_ratio 0.05 \
    --dataloader_num_workers 4 \
    --dataset_num_proc 4 \
    --output_dir "$output_path" \
    --attn_impl flash_attn \
    --use_hf true

But in terminal, it shows that:

Your hardware and system info

CUDA: 8.6
CUDA Toolkit: 12.4
Triton: 3.2.0
Ubuntu 22.04
GPU: NVIDIA GeForce RTX 3090 Ti

Library
ms_swift: 3.5.0.dev0
transformers: 4.51.3
torch: 2.6.0+cu124
unsloth: 2025.5.8
vllm: 0.8.5.post1

[Qiuzg]

The same error.
If set 'quant_bits ' at the sametime, the train_type can be 'lora', else Unsloth will still indicates 'full fine-tuning'

[slin000111]

Describe the bug I ran SFT on Qwen2.5-VL-3B-Instruct using LoRA, but Unsloth still indicates that full fine-tuning is being performed.

My script:

swift sft
--model Qwen/Qwen2.5-VL-3B-Instruct
--tuner_backend unsloth
--train_type lora
--dataset "$dataset_path"
--split_dataset_ratio 0
--torch_dtype bfloat16
--num_train_epochs "$epochs"
--per_device_train_batch_size "$batch_size"
--learning_rate 1e-5
--gradient_accumulation_steps "$gradient_accumulation_steps"
--save_steps 1000
--report_to wandb
--save_total_limit 2
--logging_steps 50
--max_length "$max_length"
--warmup_ratio 0.05
--dataloader_num_workers 4
--dataset_num_proc 4
--output_dir "$output_path"
--attn_impl flash_attn
--use_hf true
But in terminal, it shows that:

Your hardware and system info

CUDA: 8.6 CUDA Toolkit: 12.4 Triton: 3.2.0 Ubuntu 22.04 GPU: NVIDIA GeForce RTX 3090 Ti

Library ms_swift: 3.5.0.dev0 transformers: 4.51.3 torch: 2.6.0+cu124 unsloth: 2025.5.8 vllm: 0.8.5.post1

I tested it and found that Lora fine-tuning is normal.

[INFO:swift] model_parameter_info: PeftModel: 3769.5898M Params (14.9668M Trainable [0.3970%]), 0.0024M Buffers.

unsloth                                  2025.8.5
ms_swift                                 3.7.0