使用SFT和VLLM微调和部署Llama3-8b模型

本文主要是介绍使用SFT和VLLM微调和部署Llama3-8b模型，希望对大家解决编程问题提供一定的参考价值，需要的开发者们随着小编来一起学习吧！

1. 环境安装

pip install -q -U bitsandbytes
pip install -q -U git+https://github.com/huggingface/transformers.git
pip install -q -U git+https://github.com/huggingface/peft.git
pip install -q -U git+https://github.com/huggingface/accelerate.git
pip install trl

2. accelerator准备

import os
import torch
from datasets import load_dataset
from transformers import (AutoModelForCausalLM,AutoTokenizer,BitsAndBytesConfig,HfArgumentParser,TrainingArguments,pipeline,logging,
)
from peft import LoraConfig, PeftModel
from trl import SFTTrainer
from accelerate import FullyShardedDataParallelPlugin, Accelerator
from torch.distributed.fsdp.fully_sharded_data_parallel import FullOptimStateDictConfig, FullStateDictConfigfsdp_plugin = FullyShardedDataParallelPlugin(state_dict_config=FullStateDictConfig(offload_to_cpu=True, rank0_only=False),optim_state_dict_config=FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=False),
)accelerator = Accelerator(fsdp_plugin=fsdp_plugin)

3. 加载llama3和数据

因为使用的是base模型，所以没有一个严格的提示模板需要遵循。使用的数据集遵循LLama3的模板格式，因此对于使用Llama3聊天格式的下游任务来说应该没问题。如果你使用自己的数据，你可以自定义格式，在下游任务中也使用相同的格式即可。

base_model_id = "meta-llama/Meta-Llama-3-8B"
dataset_name = "scooterman/guanaco-llama3-1k"
new_model = "llama3-8b-SFT"from datasets import load_dataset
dataset = load_dataset(dataset_name, split="train")import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfigmodel = AutoModelForCausalLM.from_pretrained(base_model_id, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(base_model_id,add_eos_token=True,add_bos_token=True, 
)
tokenizer.pad_token = tokenizer.eos_token

4. 训练参数配置

许多教程只是简单地粘贴一个参数列表，让读者自己去弄清楚每个参数的作用。下面我添加了注释来解释每个参数的作用!

# Output directory where the results and checkpoint are stored
output_dir = "./results"# Number of training epochs - how many times does the model see the whole dataset
num_train_epochs = 1 #Increase this for a larger finetune# Enable fp16/bf16 training. This is the type of each weight. Since we are on an A100
# we can set bf16 to true because it can handle that type of computation
bf16 = True# Batch size is the number of training examples used to train a single forward and backward pass. 
per_device_train_batch_size = 4# Gradients are accumulated over multiple mini-batches before updating the model weights. 
# This allows for effectively training with a larger batch size on hardware with limited memory
gradient_accumulation_steps = 2# memory optimization technique that reduces RAM usage during training by intermittently storing 
# intermediate activations instead of retaining them throughout the entire forward pass, trading 
# computational time for lower memory consumption.
gradient_checkpointing = True# Maximum gradient normal (gradient clipping)
max_grad_norm = 0.3# Initial learning rate (AdamW optimizer)
learning_rate = 2e-4# Weight decay to apply to all layers except bias/LayerNorm weights
weight_decay = 0.001# Optimizer to use
optim = "paged_adamw_32bit"# Number of training steps (overrides num_train_epochs)
max_steps = 5# Ratio of steps for a linear warmup (from 0 to learning rate)
warmup_ratio = 0.03# Group sequences into batches with same length
# Saves memory and speeds up training considerably
group_by_length = True# Save checkpoint every X updates steps
save_steps = 100# Log every X updates steps
logging_steps = 5

5. 微调

建立一个wandb帐户来监控这次微调任务。

pip install wandb

import wandb
training_arguments = TrainingArguments(output_dir=output_dir,num_train_epochs=num_train_epochs,per_device_train_batch_size=per_device_train_batch_size,gradient_accumulation_steps=gradient_accumulation_steps,optim=optim,save_steps=save_steps,logging_steps=logging_steps,learning_rate=learning_rate,weight_decay=weight_decay,bf16=bf16,max_grad_norm=max_grad_norm,max_steps=max_steps,warmup_ratio=warmup_ratio,group_by_length=group_by_length,report_to="wandb"
)trainer = SFTTrainer(model=model,train_dataset=dataset,dataset_text_field="text",tokenizer=tokenizer,args=training_arguments,
)trainer.train()# Save trained model
trainer.model.save_pretrained(new_model)

6. vllm部署

为了部署这个模型以进行极快的推理，使用VLLM并托管一个OpenAI兼容端点。可能需要重新启动内核，然后运行下面的单元。

pip install vllm

python -O -u -m vllm.entrypoints.openai.api_server \--host=127.0.0.1 \--port=8000 \--model=brev-llama3-8b-SFT \--tokenizer=meta-llama/Meta-Llama-3-8B \--tensor-parallel-size=2

7. Llama-3-8b-instruct的使用

Instruct 版本对话prompt结构:

<|begin_of_text|><|start_header_id|>system<|end_header_id|>{{ system_prompt }}<|eot_id|><|start_header_id|>user<|end_header_id|>{{ user_msg_1 }}<|eot_id|><|start_header_id|>assistant<|end_header_id|>{{ model_answer_1 }}<|eot_id|>

16 GB 的 RAM，包括 3090 或 4090 等消费级 GPU

import transformers
import torchmodel_id = "meta-llama/Meta-Llama-3-8B-Instruct"pipeline = transformers.pipeline("text-generation",model=model_id,model_kwargs={"torch_dtype": torch.bfloat16},device="cuda",
)messages = [{"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},{"role": "user", "content": "Who are you?"},
]prompt = pipeline.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True
)terminators = [pipeline.tokenizer.eos_token_id,pipeline.tokenizer.convert_tokens_to_ids("<|eot_id|>")
]outputs = pipeline(prompt,max_new_tokens=256,eos_token_id=terminators,do_sample=True,temperature=0.6,top_p=0.9,
)
print(outputs[0]["generated_text"][len(prompt):])

量化版，4 bits加载需要大约 7 GB 的内存运行

pipeline = transformers.pipeline("text-generation",model=model_id,model_kwargs={"torch_dtype": torch.float16,"quantization_config": {"load_in_4bit": True},"low_cpu_mem_usage": True,},
)