modifying the training script to be more memory efficient: Unsloth, also reduced the number of steps to 15 instead of 40

writing the GRPO script to train qwen1.7 on A100 GRPO