Question about freezing parameters of VL model

[Pay20Y]

Hi, when I SFT a Qwen2.5-VL-32B model, I just want to fine-tune the aligner, so I set

--freeze_vit true
--freeze_llm true
--freeze_aligner false

But it does not work, part of log as follows:

2026-01-07T20:23:30.137320296+08:00 freeze_aligner=True,
2026-01-07T20:23:30.137322007+08:00 freeze_llm=False,
2026-01-07T20:23:30.137324016+08:00 freeze_parameters=['model.visual', 'model.visual.merger'],
2026-01-07T20:23:30.137325741+08:00 freeze_parameters_ratio=0.0,
2026-01-07T20:23:30.137327538+08:00 freeze_parameters_regex=None,
2026-01-07T20:23:30.137329196+08:00 freeze_vit=True,

I wonder the reason, and the whole setting of my training are as follows:

PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
NNODES=$nnodes \
NODE_RANK=$node_rank \
MASTER_ADDR=$master_addr \
MASTER_PORT=$master_port \
NPROC_PER_NODE=8 \
swift sft \
    --model_type qwen2_5_vl \
    --model XXX \
    --train_type full \
    --dataset XXX \
    --load_from_cache_file true \
    --bf16 true \
    --micro_batch_size 1 \
    --global_batch_size 256 \
    --sequence_parallel_size 1 \
    --recompute_granularity full \
    --recompute_method uniform \
    --recompute_num_layers 1 \
    --split_dataset_ratio 0.0 \
    --max_epochs 1 \
    --cross_entropy_loss_fusion true \
    --lr 1e-7 \
    --lr_warmup_fraction 0.05 \
    --min_lr 1e-7 \
    --save XXX \
    --save_interval 1000 \
    --eval_interval 500 \
    --max_length 16384 \
    --num_workers 8 \
    --dataset_num_proc 512 \
    --no_save_optim true \
    --no_save_rng true \
    --attention_backend flash \
    --deepspeed zero3 \
    --freeze_llm true \
    --freeze_vit true \
    --freeze_aligner false

Thanks!

[Jintao-Huang]

hello

This is usually caused by missing '\' in the English command line or having spaces after \

You can refer to the scripts in the examples/ directory