参考资料：
- https://openbayes.com/console/bbruceyuan/containers/OPg9Oo99ET6
- https://www.bilibili.com/video/BV1NM1tY3Eu5
LoRA微调案例
首先还是要安装：
!pip install -q accelerate peft bitsandbytes transformers sentencepiece
!pip install trl==0.12.0
配置环境变量：
import os
os.environ["HF_HOME"] = "/openbayes/home/huggingface"
读取数据集（这是一个将文言文翻译成现代文的数据集）
from transformers import AutoTokenizer
from datasets import load_dataset
# 加载一个 tokenizer 来使用它的聊天模板
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")
def format_prompt(example):chat = [{"role": "system", "content": "你是一个非常棒的人工智能助手，是UP主 “用代码打点酱油的chaofa” 开发的"},{"role": "user", "content": example["input"]},{"role": "assistant", "content": example["target"]}]prompt = tokenizer.apply_chat_template(chat, tokenize=False)return {"text": prompt}
# Load and format the data using the template TinyLLama is using
dataset = load_dataset("YeungNLP/firefly-train-1.1M", split="train[:500]")
dataset = dataset.map(format_prompt)
这里要顺便用模型的tokenizer来apply一下prompt_template
获得的dataset形如：
Dataset({features: ['kind', 'input', 'target', 'text'],num_rows: 500
})
然后可以查看一个format好的prompt：
# Example of formatted prompt
print(dataset["text"][100])Output:
<|im_start|>system
你是一个非常棒的人工智能助手，是UP主 “用代码打点酱油的chaofa” 开发的<|im_end|>
<|im_start|>user
我当时在三司，访求太祖、仁宗的手书敕令没有见到，然而人人能传诵那些话，禁止私盐的建议也最终被搁置。
翻译成文言文：<|im_end|>
<|im_start|>assistant
余时在三司，求访两朝墨敕不获，然人人能诵其言，议亦竟寝。<|im_end|>
接下来开始加载模型，我们使用量化来节约显存：
# 量化，为了省显存
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfigmodel_name = "Qwen/Qwen2.5-0.5B-Instruct"# 4-bit quantization configuration - Q in QLoRA
bnb_config = BitsAndBytesConfig(load_in_4bit=True, # Use 4-bit precision model loadingbnb_4bit_quant_type="nf4", # Quantization typebnb_4bit_compute_dtype="float16", # Compute dtypebnb_4bit_use_double_quant=True, # Apply nested quantization
)
# Load the model to train on the GPU
model = AutoModelForCausalLM.from_pretrained(model_name,device_map="auto",# Leave this out for regular SFTquantization_config=bnb_config,
)
model.config.use_cache = False
model.config.pretraining_tp = 1 # 上面这两个配置，只有在 k-bit 量化的时候需要设置
# Load LLaMA tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
# tokenizer.pad_token = "<PAD>" # qwen2 的 pad token 不是 <pad>，所以用 <im_end>，因此需要注释掉
tokenizer.padding_side = "left" # 训练不重要，推理比较重要
然后配置LoraConfig：
from peft import LoraConfig, prepare_model_for_kbit_training, get_peft_model# Prepare LoRA Configuration
peft_config = LoraConfig(lora_alpha=32, # LoRA Scalinglora_dropout=0.1, # Dropout for LoRA Layersr=64, # Rank，可训练数据越多，设置越大bias="none",task_type="CAUSAL_LM",target_modules= ['k_proj', 'v_proj', 'q_proj']# Layers to target# ['k_proj', 'gate_proj', 'v_proj', 'up_proj', 'q_proj', 'o_proj', 'down_proj']
)# prepare model for training
model = prepare_model_for_kbit_training(model)
# 如果没有 prepare_model_for_kbit_training，
# 且 training args 中配置了 gradient_checkpointing=True （这个其实也是为了省显存，其实不重要）
# 那么需要设置 model.enable_input_require_grads()
# model.enable_input_require_grads()
model = get_peft_model(model, peft_config)
一些训练的配置
from transformers import TrainingArgumentsoutput_dir = "./results"# Training arguments
training_arguments = TrainingArguments(output_dir=output_dir,per_device_train_batch_size=2,gradient_accumulation_steps=4,optim="adamw_torch",learning_rate=2e-4,lr_scheduler_type="cosine",num_train_epochs=1,logging_steps=10,fp16=True,gradient_checkpointing=True
)
就绪之后就是直接跑trl.SFTTrainer即可：
from trl import SFTTrainer
# Set supervised fine-tuning parameters
trainer = SFTTrainer(model=model,train_dataset=dataset,dataset_text_field="text", # 注意 dataset 中的 text 字段tokenizer=tokenizer,args=training_arguments,max_seq_length=512,# Leave this out for regular SFTpeft_config=peft_config,
)
# Train model
trainer.train()
# Save QLoRA weights
trainer.model.save_pretrained("qwen2.5-0.5b-instruct-chaofa")
微调后需要做什么？
Merge Adapter (LoRA 和 base model 合并)
from peft import AutoPeftModelForCausalLM
model = AutoPeftModelForCausalLM.from_pretrained("qwen2.5-0.5b-instruct-chaofa",low_cpu_mem_usage=True,device_map="auto",
)
# Merge LoRA and base model
merged_model = model.merge_and_unload()
from transformers import pipeline
pipe = pipeline(task="text-generation", model=merged_model, tokenizer=tokenizer)
prompt_example = """<|im_start|>system
你是一个非常棒的人工智能助手，是UP主 “用代码打点酱油的chaofa” 开发的。<|im_end|>
<|im_start|>user
天气太热了，所以我今天没有学习一点。
翻译成文言文：<|im_end|>
<|im_start|>assistant
"""
print(pipe(prompt_example, max_new_tokens=50)[0]["generated_text"])
输出结果：
<|im_start|>system
你是一个非常棒的人工智能助手，是UP主 “用代码打点酱油的chaofa” 开发的。<|im_end|>
<|im_start|>user
天气太热了，所以我今天没有学习一点。
翻译成文言文：<|im_end|>
<|im_start|>assistant
天气甚热，故今日无学一息。
其他微调方法
PPO / DPO（以Llama为例）
依然是先加载数据集
from datasets import load_dataset
def format_prompt(example):"""Format the prompt to using the <|user|> template TinyLLama is using"""# Format answerssystem = "<|system|>\n" + example['system'] + "</s>\n"prompt = "<|user|>\n" + example['input'] + "</s>\n<|assistant|>\n"chosen = example['chosen'] + "</s>\n"rejected = example['rejected'] + "</s>\n"return {"prompt": system + prompt,"chosen": chosen,"rejected": rejected,}# Apply formatting to the dataset and select relatively short answers
dpo_dataset = load_dataset("argilla/distilabel-intel-orca-dpo-pairs", split="train")
dpo_dataset = dpo_dataset.filter(lambda r:r["status"] != "tie" andr["chosen_score"] >= 8 andnot r["in_gsm8k_train"]
)
dpo_dataset = dpo_dataset.map(format_prompt, remove_columns=dpo_dataset.column_names)
dpo_dataset
"""
Dataset({features: ['chosen', 'rejected', 'prompt'],num_rows: 5922
})
"""
然后同样的加载模型（使用量化节约显存）
from peft import AutoPeftModelForCausalLM
from transformers import BitsAndBytesConfig, AutoTokenizer# 4-bit quantization configuration - Q in QLoRA
bnb_config = BitsAndBytesConfig(load_in_4bit=True, # Use 4-bit precision model loadingbnb_4bit_quant_type="nf4", # Quantization typebnb_4bit_compute_dtype="float16", # Compute dtypebnb_4bit_use_double_quant=True, # Apply nested quantization
)# Merge LoRA and base model
model = AutoPeftModelForCausalLM.from_pretrained("TinyLlama-1.1B-qlora",low_cpu_mem_usage=True,device_map="auto",quantization_config=bnb_config,
)
merged_model = model.merge_and_unload()# Load LLaMA tokenizer
model_name = "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
tokenizer.pad_token = "<PAD>"
tokenizer.padding_side = "left"
这里使用的是TinyLlama-1.1B-qlora
接着还是先配置LoRA的Config：
from peft import LoraConfig, prepare_model_for_kbit_training, get_peft_model# Prepare LoRA Configuration
peft_config = LoraConfig(lora_alpha=32, # LoRA Scalinglora_dropout=0.1, # Dropout for LoRA Layersr=64, # Rankbias="none",task_type="CAUSAL_LM",target_modules= # Layers to target['k_proj', 'gate_proj', 'v_proj', 'up_proj', 'q_proj', 'o_proj', 'down_proj']
)# prepare model for training
model = prepare_model_for_kbit_training(model)
model = get_peft_model(model, peft_config)
然后在训练参数设置时，可以使用DPO策略进行训练微调：
from trl import DPOConfigoutput_dir = "./results"# Training arguments
training_arguments = DPOConfig(output_dir=output_dir,per_device_train_batch_size=2,gradient_accumulation_steps=4,optim="paged_adamw_32bit",learning_rate=1e-5,lr_scheduler_type="cosine",max_steps=200,logging_steps=10,fp16=True,gradient_checkpointing=True,warmup_ratio=0.1
)
如法炮制，开干：
from trl import DPOTrainer# Create DPO trainer
dpo_trainer = DPOTrainer(model,args=training_arguments,train_dataset=dpo_dataset,tokenizer=tokenizer,peft_config=peft_config,beta=0.1,max_prompt_length=512,max_length=512,
)
同理最后需要合并Adapters
from peft import PeftModel# Merge LoRA and base model
model = AutoPeftModelForCausalLM.from_pretrained("TinyLlama-1.1B-qlora",low_cpu_mem_usage=True,device_map="auto",
)
sft_model = model.merge_and_unload()# Merge DPO LoRA and SFT model
dpo_model = PeftModel.from_pretrained(sft_model,"TinyLlama-1.1B-dpo-qlora",device_map="auto",
)
dpo_model = dpo_model.merge_and_unload()
微调好的模型，直接使用transformers的pipeline调用
from transformers import pipeline
# Use our predefined prompt template
prompt = """<|user|>
Tell me something about Large Language Models.</s>
<|assistant|>
"""
# Run our instruction-tuned model
pipe = pipeline(task="text-generation", model=dpo_model, tokenizer=tokenizer)
print(pipe(prompt)[0]["generated_text"])
输出结果：
<|user|>
Tell me something about Large Language Models.</s>
<|assistant|>
Large Language Models (LLMs) are a type of artificial intelligence (AI) that can generate human-like language. They are trained on large amounts of data, including text, audio, and video, and are capable of generating complex and nuanced language.LLMs are used in a variety of applications, including natural language processing (NLP), machine translation, and chatbots. They can be used to generate text, speech, or images, and can be trained to understand different languages and dialects.One of the most significant applications of LLMs is in the field of natural language generation (NLG). LLMs can be used to generate text in a variety of languages, including English, French, and German. They can also be used to generate speech, such as in chatbots or voice assistants.LLMs have the potential to revolutionize the way we communicate and interact with each other. They can help us create more engaging and personalized content, and they can also help us understand each other better.