FORCE_COLOR=1 deepspeed --hostfile hostfile \
  --num_nodes 6 --num_gpus 4 \
  train_sft_lora.py \
    --model_name_or_path /home/test/Qwen3-32B \
    --data_glob "/home/test/datasets/my_corpus/train*.jsonl" \
    --eval_data_glob "/home/test/jd_train/datasets/test/*.jsonl" \
    --output_dir /home/test/checkpoints/q3-32b-lora \
    --seq_len 512 \
    --bf16 \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 1 \
    --learning_rate 2e-4 \
    --warmup_ratio 0.03 \
    --lora_r 32 --lora_alpha 64 --lora_dropout 0.05 \
    --lora_exclude lm_head \
    --max_steps 300 \
    --log_interval 10 \
    --eval_steps 50 \
    --save_steps 50 \
    --save_total_limit 4 \
    --gradient_checkpointing \
    --deepspeed /home/test/jd_train/ds_config_zero3_lora.json \
    --report_to wandb --wandb_project ds-qwen3-lora