#!/bin/bash deepspeed --num_nodes=1 --num_gpus=7 --master_port=25001 llava/train/train_mem.py \ --deepspeed ./scripts/zero2.json \ --model_name_or_path /mnt/bn/algo-masp-nas-2/xiangchen/model/gemma-7b-it \ --version plain \ --dataset_config /mnt/bn/algo-masp-nas-2/xiangchen/repo/LLaVA/llava/configs/pretrain_data.yaml \ --vision_tower eva-vit-g \ --vit_model_path /mnt/bn/data-tns-algo-masp/baiyi.by/masp/model/eva_vit_g.pth \ --qformer_model_path /mnt/bn/data-tns-algo-masp/baiyi.by/masp/model/blip2_pretrained_flant5xxl.pth \ --tune_mm_mlp_adapter True \ --mm_vision_select_layer -2 \ --mm_use_start_end True \ --mm_use_patch_token False \ --image_aspect_ratio pad \ --bf16 True \ --output_dir /mnt/bn/algo-masp-nas-2/xiangchen/model/masp_models/checkpoints/llava-gemma-pretrain-projector \ --num_train_epochs 1 \ --per_device_train_batch_size 32 \ --per_device_eval_batch_size 4 \ --gradient_accumulation_steps 1 \ --evaluation_strategy "no" \ --save_strategy "steps" \ --save_steps 2000 \ --save_total_limit 1 \ --learning_rate 1e-3 \ --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 1 \ --lazy_preprocess True \ --report_to none