h = x @ (W_A @ W_B) * α

# step1 安装依赖!pip install trl transformers accelerate datasets bitsandbytes einops torch huggingface-hub git+https://github.com/huggingface/peft.git
from datasets import load_datasetfrom random import randrangeimport torchfrom transformers import AutoTokenizer, AutoModelForSeq2SeqLM,TrainingArguments,pipelinefrom peft import LoraConfig, prepare_model_for_kbit_training, get_peft_model, AutoPeftModelForCausalLMfrom trl import SFTTrainer
# 训练数据使用：https://huggingface.co/datasets/samsumdataset = load_dataset("samsum")
model_name = "google/flan-t5-small"model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
#Makes training faster but a little less accurate model.config.pretraining_tp = 1
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
#setting padding instructions for tokenizertokenizer.pad_token = tokenizer.eos_tokentokenizer.padding_side = "right"
def prompt_instruction_format(sample):  # 格式化promptreturn f"""### Instruction:Use the Task below and the Input given to write the Response:
### Task:Summarize the Input
### Input:{sample['dialogue']}
### Response:{sample['summary']}"""
# Create the trainertrainingArgs = TrainingArguments(output_dir='output',num_train_epochs=1,per_device_train_batch_size=4,save_strategy="epoch",learning_rate=2e-4)
peft_config = LoraConfig(lora_alpha=16,lora_dropout=0.1,r=64,bias="none",task_type="CAUSAL_LM",)
trainer = SFTTrainer(model=model,train_dataset=dataset['train'],eval_dataset = dataset['test'],peft_config=peft_config,tokenizer=tokenizer,packing=True,formatting_func=prompt_instruction_format,args=trainingArgs,)
trainer.train()