[Usage] Training speed for visual instruction tuning

[DietDietDiet]

Describe the issue

Issue: I am finetuning llava1.5-7B on 8 * A100 40G, and modified bs & accumulation steps accordingly. The estimated training time is approx. 24h. What could go wrong?

Env: cuda11.7 torch2.0.1 flash-attn 2.3.2

Command:

PASTE THE COMMANDS HERE.

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 deepspeed llava/train/train_mem.py \ --deepspeed ./scripts/zero3.json \ --model_name_or_path lmsys/vicuna-7b-v1.5 \ --version v1 \ --data_path data/llava_v1_5_mix665k.json.bak \ --image_folder ./playground/data \ --vision_tower openai/clip-vit-large-patch14-336 \ --pretrain_mm_mlp_adapter projectors/llava-v1.5-mlp2x-336px-pretrain-vicuna-7b-v1.5/mm_projector.bin \ --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 ./checkpoints/llava-v1.5-7b \ --num_train_epochs 1 \ --per_device_train_batch_size 4 \ --per_device_eval_batch_size 4 \ --gradient_accumulation_steps 4 \ --evaluation_strategy "no" \ --save_strategy "steps" \ --save_steps 50000 \ --save_total_limit 1 \ --learning_rate 2e-5 \ --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 8 \ --lazy_preprocess True \ --report_to wandb Log:

PASTE THE LOGS HERE.

Screenshots: You may attach screenshots if it better explains the issue.

[haotian-liu]

For 7B model, you probably do not need to set the batch size to be so small, maybe 8x2 is sufficient.

Also, are your GPUs NVLinked?