CUDA_HOME=/usr/local/cuda
gpus=${gpus:-0,1,2,3,4,5,6,7}
dataset=${dataset:-sft_train_data}
output_dir=${output_dir:-experiments/$(date +"%Y%m%d_%H%M%S")}

port=$(shuf -i 10000-20000 -n 1)
deepspeed --include localhost:${gpus} --master_port=$port grpo_demo.py \
    --deepspeed "ds_zero2.json" \
    --model_name_or_path "path to Qwen2.5-1.5B-Instruct/" \
    --output_dir outputs/Qwen2.5-1.5B-GRPO-gsm8k \
    --run_name Qwen2.5-1.5B-GRPO-gsm8k \
    --learning_rate 1e-5 \
    --adam_beta1 0.9 \
    --adam_beta2 0.99 \
    --weight_decay 0.1 \
    --warmup_ratio 0.1 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --bf16 True \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 2 \
    --num_generations 16 \
    --max_prompt_length 512 \
    --max_completion_length 768 \
    --num_train_epochs 5 \
    --save_steps 100 \
    --max_grad_norm 0.1 \
    --report_to tensorboard \
    --log_on_each_node False