#!/bin/bash
# C3 Phase 2 SFT: 10000 opt-steps from stage1_merged, stage_0=0/1=24000/2=48000/vqa=64000
set -euo pipefail
source /etc/network_turbo 2>/dev/null || true
source /root/autodl-tmp/covt_repro/scripts/_common_env.sh

STAGE1_MERGED=$BASE/covt_data/output/merged/stage1_merged
OUTPUT_DIR=$BASE/covt_data/output/lora_stage234
mkdir -p $OUTPUT_DIR

export MASTER_PORT=22813
export CUDA_VISIBLE_DEVICES=0

NUM_DEVICES=1
BATCH_PER_DEVICE=4
GLOBAL_BATCH_SIZE=8
GRAD_ACCUM_STEPS=$((GLOBAL_BATCH_SIZE / (BATCH_PER_DEVICE * NUM_DEVICES)))

STAGE_0_STEP=0
STAGE_1_STEP=6000
STAGE_2_STEP=12000
VQA_ONLY_STAGE=16000
MAX_STEPS=10000

cd $COVT_CODE/train

deepspeed \
    --num_gpus 1 \
    --master_port $MASTER_PORT \
    src/training/train.py \
    --use_liger True \
    --lora_enable True \
    --vision_lora True \
    --use_dora False \
    --lora_namespan_exclude '["embed_tokens", "lm_head", "dino", "sam", "depth", "SD", "internvit", "pidinet", "siglip", "metaclip"]' \
    --lora_rank 16 \
    --lora_alpha 32 \
    --lora_dropout 0.01 \
    --num_lora_modules -1 \
    --model_id $STAGE1_MERGED \
    --model_path $STAGE1_MERGED \
    --anchor_model_id '["sam", "depth", "dino"]' \
    --data_path $DATA_DIR/dataset/CoVT-Dataset \
    --image_folder $DATA_DIR/dataset/CoVT-Dataset \
    --remove_unused_columns False \
    --freeze_vision_tower True \
    --freeze_llm True \
    --tune_merger False \
    --bf16 True \
    --fp16 False \
    --disable_flash_attn2 False \
    --output_dir "$OUTPUT_DIR" \
    --resume_from_checkpoint True \
    --max_steps $MAX_STEPS \
    --per_device_train_batch_size $BATCH_PER_DEVICE \
    --gradient_accumulation_steps $GRAD_ACCUM_STEPS \
    --image_min_pixels $IMAGE_MIN_PIXELS \
    --image_max_pixels $IMAGE_MAX_PIXELS \
    --image_resized_width $IMAGE_RESIZE_W \
    --image_resized_height $IMAGE_RESIZE_H \
    --learning_rate 5e-5 \
    --projection_layer_lr 1e-5 \
    --weight_decay 0.1 \
    --warmup_ratio 0.05 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --gradient_checkpointing True \
    --lazy_preprocess True \
    --save_strategy "steps" \
    --save_steps 1000 \
    --dataloader_num_workers 0 \
    --deepspeed $SCRIPTS_DIR/zero2_pro6000.json \
    --report_to tensorboard \
    --run_name "sft_phase2" \
    --training_stage "full" \
    --stage_0_step $((STAGE_0_STEP * BATCH_PER_DEVICE)) \
    --stage_1_step $((STAGE_1_STEP * BATCH_PER_DEVICE)) \
    --stage_2_step $((STAGE_2_STEP * BATCH_PER_DEVICE)) \
    --vqa_only_stage $VQA_ONLY_STAGE \
    2>&1 | tee $OUTPUT_DIR/sft_phase2.log