set -x

# ==================== DDP Configuration ====================
GPUS=${GPUS:-8}
NNODES=${NNODES:-1}
NODE_RANK=${NODE_RANK:-0}
MASTER_ADDR=${MASTER_ADDR:-"127.0.0.1"}
PORT=${PORT:-29501}
TOTAL_GPUS=$((GPUS * NNODES))

export PYTHONPATH="${PYTHONPATH}:$(pwd)"
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

OUTPUT_DIR="./shell/playground/train_states/Debug_SDLM_7B_D8"
BASE_MODEL="./shell/playground/ckpt/Qwen/Qwen2.5-7B"
# META_JSON="./shell/playground/data/meta/sft_opc436k_scale_math_1m_smoltalk_1m_tulu_1m.json"
META_JSON="./shell/playground/data/meta/sft_3500k_filter.json"

if [ ! -d "$OUTPUT_DIR" ]; then
  mkdir -p "$OUTPUT_DIR"
fi

# number of gpus: 1
# batch size per gpu: 2
# gradient accumulation steps: 1
# total batch size: 2
# num of examples:
# epoch: 1
torchrun \
    --nnodes=$NNODES \
    --node_rank=$NODE_RANK \
    --master_addr=$MASTER_ADDR \
    --nproc_per_node=$GPUS \
    --master_port=$PORT \
  sdlm/train/sdlm_train.py \
  --model_name_or_path ${BASE_MODEL} \
  --conv_style "Qwen-2-5" \
  --use_fast_tokenizer False \
  --output_dir ${OUTPUT_DIR} \
  --meta_path ${META_JSON} \
  --overwrite_output_dir True \
  --dataloader_num_workers 4 \
  --bf16 True \
  --num_train_epochs 1 \
  --per_device_train_batch_size 2 \
  --gradient_accumulation_steps 2 \
  --save_strategy "steps" \
  --save_steps 400 \
  --save_total_limit 100 \
  --learning_rate 5e-6 \
  --weight_decay 0.01 \
  --warmup_ratio 0.03 \
  --lr_scheduler_type "constant" \
  --logging_steps 1 \
  --max_seq_length 6144 \
  --do_train True \
  --grad_checkpoint True \
  --deepspeed "zero_stage1_config.json" \
  --report_to "tensorboard" \
  --block_size 8 \
  --causal_attn False \
  --attn_implementation sdpa \
  2>&1 | tee -a "${OUTPUT_DIR}/training_log_$(date +%Y%m%d_%H%M%S).txt"