[Usage] Deepspeed Zero Stage 3 not able to shard the model

[Shubham Agarwal]

Hi @haotian-liu !

Interesting work around LLaVa!

Issue:

I am trying to finetune LLaVa using 8 X H100.

When I try to use DeepSpeed Zero Stage 3, it seems that the model gets replicated on all the GPUs, instead of being sharded. I get OOM issues when finetuning model. I am trying to use a context length of 2048 and ViT with 336 resolution.

Could you please suggest what I might be doing wrong here?

Command:

deepspeed llava/train/train_mem.py \
    --deepspeed ./scripts/zero3.json \
    --model_name_or_path ../$MODEL_VERSION \
    --version $PROMPT_VERSION \
    --data_path ./finetune_data/cleaned_finetune_data.json \
    --image_folder ./finetune_data/images \
    --vision_tower openai/clip-vit-large-patch14-336 \
    --pretrain_mm_mlp_adapter ./checkpoints/llava-$MODEL_VERSION-pretrain/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-$MODEL_VERSION-finetune \
    --num_train_epochs 1 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 16 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 500 \
    --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 4 \
    --lazy_preprocess True \

When I run the model using CUDA_VISIBLE_DEVICES=0 bash ./scripts/sample_stage3.sh, the memory usage before training is:

However, when I am using the stage 3 deepspeed, the GPU usage before training is

And the model gets OOM after this. Could you please suggest what flag we might need to change?