How to fine-tune llava-v1.6-mistral-7b on GQA dataset

[qhz991029]

Question

Thank you for your great work!

I am trying to fine-tune llava-v1.6-mistral-7b on the provided GQA dataset, using the script finetune_task_lora.sh. However, the loss dosen't decrease and test result on GQA is worse. I wonder how should I fine-tune llava-v1.6 models?

This is the modified script:

#!/bin/bash

deepspeed --include "node-0:0,1,2,3" llava/train/train_mem.py \
    --lora_enable True --lora_r 128 --lora_alpha 256 --mm_projector_lr 2e-5 \
    --deepspeed ./scripts/zero2.json \
    --model_name_or_path liuhaotian/llava-v1.5-7b \
    --version mistral_instruct \
    --data_path /SOME/PATH/TO/LLaVA/playground/gqa.json \
    --data_length 10000 \
    --image_folder ./playground/data \
    --vision_tower openai/clip-vit-large-patch14-336 \
    --mm_patch_merge_type spatial_unpad \
    --mm_projector_type mlp2x_gelu \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --image_aspect_ratio anyres \
    --group_by_modality_length True \
    --bf16 True \
    --output_dir ./checkpoints/llava-v1.5-7b-gqa-official-lora \
    --num_train_epochs 1 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 1 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 50000 \
    --save_total_limit 1 \
    --learning_rate 2e-4 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 4096 \
    --gradient_checkpointing True \
    --dataloader_num_workers 8 \
    --lazy_preprocess True \
    --report_to wandb

This the modified part in train.py:

model = LlavaMistralForCausalLM.from_pretrained(
                model_args.model_name_or_path,
                cache_dir=training_args.cache_dir,
                attn_implementation=attn_implementation,
                torch_dtype=(torch.bfloat16 if training_args.bf16 else None),
                **bnb_model_from_pretrained_args
            )
tokenizer = transformers.AutoTokenizer.from_pretrained(
            model_args.model_name_or_path,
            cache_dir=training_args.cache_dir,
            model_max_length=training_args.model_max_length,
            padding_side="left"
        )

Looking forward to reply!

[NicoZenith]

There is not training script for 1.6 unfortunately...

[YuyangYe]

There is not training script for 1.6 unfortunately...

Hello, I was wondering if there is a current training script available for version 1.6? Thank you!

[NicoZenith]

what do you guys think of this ?
https://github.com/NielsRogge/Transformers-Tutorials/blob/master/LLaVa/Fine_tune_LLaVa_on_a_custom_dataset_(with_PyTorch_Lightning).ipynb

By replacing llava by lava-next (processor and model)

[dacian7]

There is not training script for 1.6 unfortunately...

Hello, I was wondering if there is a current training script available for version 1.6? Thank you!

Hi @YuyangYe , have you found the scripts for that? Thanks!

[sweetning0809]

There is not training script for 1.6 unfortunately...

have you find any progress on this？

[sweetning0809]

there are some works on mac Blaizzy/mlx-vlm#43

[sweetning0809]

what do you guys think of this ? https://github.com/NielsRogge/Transformers-Tutorials/blob/master/LLaVa/Fine_tune_LLaVa_on_a_custom_dataset_(with_PyTorch_Lightning).ipynb

By replacing llava by lava-next (processor and model)

maybe is the right way

[NicoZenith]

There is not training script for 1.6 unfortunately...

have you find any progress on this？

Yes, the owner of this repo answered on another issue:
https://github.com/chuangchuangtan/LLaVA-NeXT-Image-Llama3-Lora

The fine-tuning does work with llama3! I suggest we move to his repo if we have new issues.