Zero loss values during finetuning

[mmderakhshani]

Dear @xiaoachen98,

Thank you very much for releasing the code. I am running your fine-tuning script to replicate your LLaMA 3 results. However, I am getting zero loss after the first iteration. Have you encountered such an issue before? I am using 8 A6000 GPUs with the total batch size of 128.

#!/bin/bash
export WANDB_DIR="/nvmestore/mderakh/wandb_llava/"
export LLaVA_PATH="/nvmestore/mderakh/vlm_datasets/playground/"

export DATA_PATH=${LLaVA_PATH}/data/open-llava-next/open-llava-next_instruct_mix1M.json
export SAVE_PATH=llava-v1.6-8b_llama3-8b_clip-large-336_pretrain_lcs-558k_sft-mix1M_lr-mlp-2e-5-vit-2e-6-llm-2e-5


OUTPUT="/nvmestore/mderakh/LLaVA-Next/clip-vit-large-patch14-336-llama3-8b/"
OUT_RESULTS=$OUTPUT/results/
mkdir -p "$OUT_RESULTS"

export BASE_LR=2e-5
export VIT_LR=2e-6
DEVICE_BATCH_SIZE=2
GRADIENT_ACCU_STEPS=8


deepspeed llava/train/train_mem.py \
    --deepspeed ./scripts/zero2.json \
    --model_name_or_path Lin-Chen/open-llava-next-llama3-8b \
    --version llava_llama_3 \
    --data_path ${DATA_PATH} \
    --image_folder ${LLaVA_PATH}/data \
    --vision_tower openai/clip-vit-large-patch14-336 \
    --mm_projector_type mlp2x_gelu \
    --pretrain_mm_mlp_adapter ${OUTPUT}/checkpoints/llava-v1.6-8b_llama3-8b_pretrain_lcs-558k_ft-mlp-lr-1e-3/mm_projector.bin \
    --unfreeze_mm_vision_tower True \
    --mm_vision_tower_lr ${VIT_LR} \
    --image_aspect_ratio anyres \
    --group_by_modality_length True \
    --mm_vision_select_layer -2 \
    --mm_vision_select_feature patch \
    --mm_patch_merge_type spatial_unpad \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --bf16 True \
    --output_dir ${OUTPUT}/checkpoints/${SAVE_PATH} \
    --num_train_epochs 1 \
    --per_device_train_batch_size ${DEVICE_BATCH_SIZE} \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps ${GRADIENT_ACCU_STEPS} \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 7975 \
    --save_total_limit 1 \
    --learning_rate ${BASE_LR} \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 6144 \
    --gradient_checkpointing True \
    --dataloader_num_workers 4 \
    --lazy_preprocess True \
    --report_to wandb \
    --run_name ${SAVE_PATH}

[StevenSmith2000]

Hi, I have encountered the same problem. Have you solved it?

[mmderakhshani]

Hi @StevenSmith2000 I could not solve it.

[Lauch1ng]

Hi, have you solved it?