scripts/v1_5/resampler/formal_task_pretrain_siglip_64.sh

#!/bin/bash

# cd /mnt/bn/algo-masp-nas-2/xiangchen/repo/LLaVA
cd /opt/tiger/masp_models
pip install --upgrade pip
pip install -e .

echo "$PWD"

ports=(`echo $METIS_WORKER_0_PORT | tr ',' ' '`)
port=${ports[0]}

echo "total workers: ${ARNOLD_WORKER_NUM}"
echo "cur worker id: ${ARNOLD_ID}"
echo "gpus per worker: ${ARNOLD_WORKER_GPU}"
echo "master ip: ${METIS_WORKER_0_HOST}"
echo "master port: ${port}"

#export OMP_NUM_THREADS=8
#export NCCL_IB_DISABLE=0
#export NCCL_IB_GID_INDEX=3
#export NCCL_IB_HCA=${ARNOLD_RDMA_DEVICE}
#export NCCL_SOCKET_IFNAME=eth0
export NCCL_DEBUG=INFO

env="$1"
cmd="$2"
echo $env
echo $cmd

deepspeed \
    --num_nodes=$ARNOLD_WORKER_NUM \
    --num_gpus=$ARNOLD_WORKER_GPU \
    --master_port=$port \
    --master_addr $METIS_WORKER_0_HOST \
    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 \
    --adapter_module_name naive_resampler \
    --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 64 \
    --num_query_token 64 \
    --bf16 True \
    --output_dir /mnt/bn/masp-nas/xiangchen/model/masp_models/checkpoints/llava-pretrain-siglip_resampler_64_projector \
    --group_by_modality_length True \
    --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