[Usage] The issue encountered when fine-tuning llava_mistral1.6 using LoRA

[yuwang4321]

Describe the issue

Issue:
"I used llava_mistral 1.6 and LoRA for fine-tuning. The model loads and works fine when epoch=1, but there is no output when epoch=10. Has anyone encountered the same issue? How can I troubleshoot and resolve it?"
Command:

deepspeed 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 /root/autodl-tmp/checkpoints/llava-v1.6-mistral-7b-1114 \
    --version v1 \
    --data_path /root/autodl-tmp/code_wy/dataset/fine_tuned_llava/MMQA_finetuned_data.jsonl \
    --image_folder /root/autodl-tmp/final_dataset_images \
    --vision_tower /root/autodl-tmp/checkpoints/clip-vit-large-patch14-336 \
    --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 pad \
    --group_by_modality_length True \
    --bf16 True \
    --output_dir /root/autodl-tmp/checkpoints/llava-v1.6-mistral-7b-hf-task-lora \
    --num_train_epochs 10 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 8 \
    --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 2048 \
    --gradient_checkpointing True \
    --dataloader_num_workers 4 \
    --lazy_preprocess True \
    --report_to wandb

python scripts/merge_lora_weights.py --model-path "/root/autodl-tmp/checkpoints/llava-v1.6-mistral-7b-hf-task-lora" \
       --model-base "/root/autodl-tmp/checkpoints/llava-v1.6-mistral-7b-1114" \
       --save-model-path "/root/autodl-tmp/checkpoints/llava-v1.6-mistral-7b-hf-merged"

Log:

/root/miniconda3/envs/llava/lib/python3.10/site-packages/transformers/generation/configuration_utils.py:392: UserWarning: `do_sample` is set to `False`. However, `temperature` is set to `0` -- this flag is only used in sample-based generation modes. You should set `do_sample=True` or unset `temperature`.
  warnings.warn(
/root/miniconda3/envs/llava/lib/python3.10/site-packages/transformers/generation/configuration_utils.py:397: UserWarning: `do_sample` is set to `False`. However, `top_p` is set to `None` -- this flag is only used in sample-based generation modes. You should set `do_sample=True` or unset `top_p`.
  warnings.warn(
The attention mask and the pad token id were not set. As a consequence, you may observe unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results.
Setting `pad_token_id` to `eos_token_id`:2 for open-end generation.
![screenshot-20241116-110518](https://github.com/user-attachments/assets/865ab513-e54d-4f11-bfff-deed11a028ef)
![screenshot-20241116-110552](https://github.com/user-attachments/assets/73a805aa-1108-49fc-876c-92dbdbb97853)

Screenshots:
epoch=10

epoch=1