玩转 Gemma 2，模型的部署与 Fine-Tuning

Google 近期发布了最新开放模型 Gemma 2，目前与同等规模的开放模型相比，取得了明显的优势，同时在安全可控性上得到了显著的增强。

Fine-Tune 演示

 a. 微调数据集 

 b. 训练环境 

 c. 微调步骤 

from transformers import AutoTokenizer, AutoModelForCausalLM

import transformers

import torch

#加载模型

model_id = "google/gemma-2-27b"
dtype = torch.bfloat16
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="cuda",
    torch_dtype=dtype,
)

#准备prompt（对话格式）

prompt = """### Human: I can't cancel the subscription to your corporate newsletter### Assistant: """

#将prompt转换成token id并输入model生成内容

inputs = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt")

outputs = model.generate(input_ids=inputs.to(model.device), max_new_tokens=100)

#解码token id，打印输出

print(tokenizer.decode(outputs[0]))

!git clone https://github.com/huggingface/peft.git

!cd examples/sft/

Gemma 2 与开源的 Hugging Face 的 Transformer library 完全兼容，因此用于训练其它模型的脚本在更改模型名称后，便可以开始训练 Gemma 模型。但是需要注意的是，不同模型多轮对话支持的特殊 token 可能不一样，因此与之相对应的 tokenzier 里面 chat 模板需要提前进行处理。本次我们使用 chatml 格式的模板，只需要在示例脚本上设置对应参数，代码对会自动对 tokenizer 进行处理。

!echo 'export   HUGGINGFACE_TOKEN=YOUR_HF_API_TOKEN' >> ~/.bashrc

!echo 'export   WANDB_API_KEY=YOUR_WANDB_API_KEY' >> ~/.bashrc

!source   ~/.bashrc

!pip install accelerate==0.32.1 

!pip install deepspeed==0.14.4

!pip install transformers==4.42.4

!pip install peft==0.11.1

!pip install trl==0.9.6

!pip install bitsandbytes==0.43.1

!pip install wandb==0.17.4

# 任务训练任务启动命令

!accelerate launch --config_file   "configs/deepspeed_config.yaml"    train.py \
  --seed 100 \
  --model_name_or_path "google/gemma-2-27b" \
  --dataset_name "smangrul/ultrachat-10k-chatml" \
  --chat_template_format "chatml" \
  --add_special_tokens False \
  --append_concat_token False \
  --splits "train,test" \
  --max_seq_len 2048 \
  --num_train_epochs 1 \
  --logging_steps 5 \
  --log_level "info" \
  --logging_strategy "steps" \
  --evaluation_strategy "epoch" \
  --save_strategy "epoch" \
  --bf16 True \
  --learning_rate 1e-4 \
  --lr_scheduler_type "cosine" \
  --weight_decay 1e-4 \
  --warmup_ratio 0.0 \
  --max_grad_norm 1.0 \
  --output_dir "gemma2-27b-sft-deepspeed" \
  --per_device_train_batch_size 1 \
  --per_device_eval_batch_size 1 \
  --gradient_accumulation_steps 4 \
  --gradient_checkpointing True \
  --use_reentrant False \
  --dataset_text_field "content"

# Deespeed配置脚本

compute_environment: LOCAL_MACHINE                                                                                                                                            
  debug: false
  deepspeed_config:
    deepspeed_multinode_launcher: standard
    gradient_accumulation_steps: 4
    offload_optimizer_device: none
    offload_param_device: none
    zero3_init_flag: true
    zero3_save_16bit_model: true
    zero_stage: 3
  distributed_type: DEEPSPEED
  downcast_bf16: 'no'
  machine_rank: 0
  main_training_function: main
  mixed_precision: bf16
  num_machines: 1
  num_processes: 8
  rdzv_backend: static
  same_network: true
  tpu_env: []
  tpu_use_cluster: false
  tpu_use_sudo: false
  use_cpu: false

from transformers import AutoTokenizer, AutoModelForCausalLM
import transformers
import torch

model_id = "./gemma2-27b-sft-deepspeed"
dtype = torch.bfloat16

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="cuda",
    torch_dtype=dtype,
)

prompt = """### Human: I can't cancel the subscription to your corporate newsletter### Assistant: """

inputs = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt")
outputs = model.generate(input_ids=inputs.to(model.device), max_new_tokens=100)
print(tokenizer.decode(outputs[0]))