#!/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 mistralai/Mistral-7B-Instruct-v0.1 \
    --version plain \
    --dataset_config /mnt/bn/algo-masp-nas-2/xiangchen/repo/LLaVA/llava/configs/pretrain_data.yaml \
    --vision_tower HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit \
    --qformer_model_path from_scratch \
    --freeze_qformer False \
    --tune_mm_mlp_adapter True \
    --mm_vision_select_layer -2 \
    --mm_use_start_end True \
    --mm_use_patch_token False \
    --image_aspect_ratio pad \
    --num_token_per_image 128 \
    --num_query_token 128 \
    --bf16 True \
    --output_dir /mnt/bn/algo-masp-nas-2/xiangchen/model/masp_models/checkpoints/llava-pretrain-siglip-qformer_scratch_128_projector \
    --group_by_modality_length True \
    --num_train_epochs 1 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 2 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 2000 \
    --save_total_limit 1 \
    --learning_rate 5e-4 \
    --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