export TOKENIZERS_PARALLELISM=false
export PYTORCH_CUDA_ALLOC_CONF="max_split_size_mb:128"
export NCCL_DEBUG=INFO

torchrun --nproc_per_node 4 /home/test/jd_train/train_sft_ds.py \
  --model_name_or_path /home/test/Qwen3-1.7B \
  --data_glob "/home/test/datasets/my_corpus/train.jsonl" \
  --output_dir /home/test/checkpoints/q3-1_7b-ds4 \
  --seq_len 512 \
  --per_device_train_batch_size 1 \
  --gradient_accumulation_steps 1 \
  --learning_rate 2e-5 --weight_decay 0.1 --warmup_ratio 0.02 \
  --max_steps 375 --log_interval 1 \
  --bf16 \
  --deepspeed /home/test/jd_train/ds_config_zero3.json \
  --report_to none \
  --eval_data_glob "/home/test/datasets/my_corpus/test.jsonl"