# Flush memorydel trainer, modelgc.collect()torch.cuda.empty_cache()# Reload tokenizer and modeltokenizer = AutoTokenizer.from_pretrained(base_model)model = AutoModelForCausalLM.from_pretrained(    base_model,    low_cpu_mem_usage=True,    return_dict=True,    torch_dtype=torch.float16,    device_map="auto",)model, tokenizer = setup_chat_format(model, tokenizer)# Merge adapter with base modelmodel = PeftModel.from_pretrained(model, new_model)model = model.merge_and_unload()model.push_to_hub(new_model, use_temp_dir=False)tokenizer.push_to_hub(new_model, use_temp_dir=False)