LoRA Finetuning Tool for Qwen2-VL-7B in Web UI (DPO updated)

#2
by hiyouga - opened
Qwen org

LLaMA Factory has integrated Qwen2-VL models for SFT and DPO, try our training recipes and webUI🚀

LoRA repices: https://github.com/hiyouga/LLaMA-Factory/blob/main/examples/train_lora/qwen2vl_lora_sft.yaml

6efc2c1a72492a9f9ecd35c71b0ab9da.png

761896e0ef3bf6d958181d70d44f6766.png

hiyouga changed discussion title from LoRA Finetuning Qwen2-VL-7B in Web UI to LoRA Finetuning Tool for Qwen2-VL-7B in Web UI

There are some issues in the https://github.com/hiyouga/LLaMA-Factory ; the full fine-tuning YAML file and related modifications have not been updated to the latest version.

Qwen org

Hi @YangJiassh , we would be appreciated if you can provide some details

examples/train_full/qwen2vl_full_sft.yaml

@YangJiassh Thanks! You can try this recipe for full tuning:

### model
model_name_or_path: Qwen/Qwen2-VL-7B-Instruct

### method
stage: sft
do_train: true
finetuning_type: full
deepspeed: examples/deepspeed/ds_z3_config.json

### dataset
dataset: mllm_demo
template: qwen2_vl
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16

### output
output_dir: saves/qwen2_vl-7b/full/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true

### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000

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

OK thanks I'll try it now

LLaMA Factory has integrated Qwen2-VL models for SFT and DPO, try our training recipes and webUI🚀

LoRA repices: https://github.com/hiyouga/LLaMA-Factory/blob/main/examples/train_lora/qwen2vl_lora_sft.yaml

6efc2c1a72492a9f9ecd35c71b0ab9da.png

761896e0ef3bf6d958181d70d44f6766.png

Damn you were fast XD, is there an example for DPO tho ? I don't see any

Both this tool and the model are amazing! SFT with multimedia has never been so easy and it's interesting to see how well the image training translates to the video mode. 加油!

Qwen org

@nicolollo Yeah! We have just supported Qwen2-VL DPO training, try this example: https://github.com/hiyouga/LLaMA-Factory/blob/main/examples/train_lora/qwen2vl_lora_dpo.yaml

hiyouga changed discussion title from LoRA Finetuning Tool for Qwen2-VL-7B in Web UI to LoRA Finetuning Tool for Qwen2-VL-7B in Web UI (DPO updated)

@hiyouga
Amazing stuff.
Can you kindly add a config for Qwen Audio 2 as well? both fft and lora. this model is essential for those of us in the Audio realm

Can i fine tune it on an L4 GPU?

Sign up or log in to comment