ChatGLM2-6B微调实践-QLora方案-Toy模板网

这篇具有很好参考价值的文章主要介绍了ChatGLM2-6B微调实践-QLora方案。希望对大家有所帮助。如果存在错误或未考虑完全的地方，请大家不吝赐教，您也可以点击"举报违法"按钮提交疑问。

环境部署

申请阿里云GPU服务器：

CentOS 7.6 64
Anaconda3-2023.07-1-Linux-x86_64
Python 3.11.5
GPU NVIDIA A10（显存24 G/1 core）
CPU 8 vCore/30G

安装Anaconda、CUDA、PyTorch
参考：ChatGLM2-6B微调实践-P-Tuning方案

Lora微调

项目部署

git clone https://github.com/shuxueslpi/chatGLM-6B-QLoRA.git
cd chatGLM-6B-QLoRA
pip install -r requirements.txt

准备数据集

准备我们自己的数据集，分别生成训练文件和测试文件这两个文件，放在项目data文件夹，数据格式为：

{
   
  "instruction": "类型#裤*版型#宽松*风格#性感*图案#线条*裤型#阔腿裤", 
  "output": "宽松的阔腿裤这两年真的吸粉不少，明星时尚达人的心头爱。毕竟好穿时尚，谁都能穿出腿长2米的效果宽松的裤腿，当然是遮肉小能手啊。上身随性自然不拘束，面料亲肤舒适贴身体验感棒棒哒。系带部分增加设计看点，还让单品的设计感更强。腿部线条若隐若现的，性感撩人。颜色敲温柔的，与裤子本身所呈现的风格有点反差萌。"
}

ChatGLM2-6B微调实践-QLora方案,大模型,linux,chatglm

训练集文件： train.json
测试集文件： dev.json

修改训练脚本

创建train_qlora.sh，添加以下命令：

python3 train_qlora.py \
--train_args_json chatGLM_6B_QLoRA.json \
--model_name_or_path THUDM/chatglm-6b \
--train_data_path data/train.json \
--eval_data_path data/dev.json \
--lora_rank 4 \
--lora_dropout 0.05 \
--compute_dtype fp32

lora_rank： qlora矩阵的秩。一般设置为8、16、32、64等，在qlora论文中作者设为64。越大则参与训练的参数量越大，一般来说效果会更好，但需要更多显存。
lora_dropout： lora权重的dropout rate。
compute_dtype：量化精度。

修改model_name_or_path参数为本地真实的模型路径。
chatGLM_6B_QLoRA.json文件为所有transformers框架支持的TrainingArguments，参考：https://huggingface.co/docs/transformers/main_classes/trainer#transformers.TrainingArguments

chatGLM_6B_QLoRA.json默认配置如下，可根据实际情况自行修改：

{
   
    "output_dir": "saved_files/chatGLM_6B_QLoRA_t32",
    "per_device_train_batch_size": 4,
    "gradient_accumulation_steps": 8,
    "per_device_eval_batch_size": 4,
    "learning_rate": 1e-3,
    "num_train_epochs": 1.0,
    "lr_scheduler_type": "linear",
    "warmup_ratio": 0.1,
    "logging_steps": 100,
    "save_strategy": "steps",
    "save_steps": 500,
    "evaluation_strategy": "steps",
    "eval_steps": 500,
    "optim": "adamw_torch",
    "fp16": false,
    "remove_unused_columns": false,
    "ddp_find_unused_parameters": false,
    "seed": 42
}

各参数说明：

per_device_train_batch_size：每个训练设备上的批量大小。
per_device_eval_batch_size：每个评估设备上的批量大小。
gradient_accumulation_steps：梯度累积步骤，用于更大的批次训练。
learning_rate：初始学习率，一般为1e-4、2e-4。
num_train_epochs：训练的轮数（epochs）,如果数据量足够大，一般建议只训一个epoch。
lr_scheduler_type：选择什么类型的学习率调度器来更新模型的学习率
warmup_ratio：线性预热从0达到learning_rate时，每步学习率的增长率
logging_steps：定义多少个更新步骤打印一次训练日志。
save_strategy：训练过程中，checkpoint的保存策略，可选值有"steps"（每隔一定步骤保存）和"epoch"（每个epoch保存一次）。
save_steps：定义多少个更新步骤保存一次模型。
evaluation_strategy：评估策略，可选值有"steps"（每隔一定步骤评估）和"epoch"（每个epoch评估一次）。
save_steps：定义多少个更新步骤评估一次模型。
optim：可以使用的优化器
fp16：是否使用bf16 16位精度训练替代32位训练
remove_unused_columns：是否自动删除模型forward方法不使用的列
seed：训练开始时设置的随机种子

执行脚本训练，训练过程如下：文章来源地址https://www.toymoban.com/news/detail-727297.html

(base) [root@iZbp178u8rw9n9ko94ubbyZ chatGLM-6B-QLoRA-main]# sh train_qlora.sh 

===================================BUG REPORT===================================
Welcome to bitsandbytes. For bug reports, please run

python -m bitsandbytes

 and submit this information together with your error trace to: https://github.com/TimDettmers/bitsandbytes/issues
================================================================================
bin /root/anaconda3/lib/python3.11/site-packages/bitsandbytes/libbitsandbytes_cuda118.so
CUDA SETUP: CUDA runtime path found: /root/anaconda3/lib/libcudart.so.11.0
CUDA SETUP: Highest compute capability among GPUs detected: 8.6
CUDA SETUP: Detected CUDA version 118
CUDA SETUP: Loading binary /root/anaconda3/lib/python3.11/site-packages/bitsandbytes/libbitsandbytes_cuda118.so...
The model weights are not tied. Please use the `tie_weights` method before using the `infer_auto_device` function.
Loading checkpoint shards: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 7/7 [00:07<00:00,  1.13s/it]
trainable params: 974,848 || all params: 3,389,286,400 || trainable%: 0.0287626327477076
Found cached dataset json (/root/.cache/huggingface/datasets/json/default-a26442a3257bd6e5/0.0.0/e347ab1c932092252e717ff3f949105a4dd28b27e842dd53157d2f72e276c2e4)
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 920.81it/s]
Loading cached processed dataset at /root/.cache/huggingface/datasets/json/default-a26442a3257bd6e5/0.0.0/e347ab1c932092252e717ff3f949105a4dd28b27e842dd53157d2f72e276c2e4/cache-a6ee5e96ac795161.arrow
Loading cached shuffled indices for dataset at /root/.cache/huggingface/datasets/json/default-a26442a3257bd6e5/0.0.0/e347ab1c932092252e717ff3f949105a4dd28b27e842dd53157d2f72e276c2e4/cache-20df68b061e7d292.arrow
Loading cached processed dataset at /root/.cache/huggingface/datasets/json/default-a26442a3257bd6e5/0.0.0/e347ab1c932092252e717ff3f949105a4dd28b27e842dd53157d2f72e276c2e4/cache-e9ff6a88c507a91d.arrow
Found cached dataset json (/root/.cache/huggingface/datasets/json/default-f82d1afe86c1e9ec/0.0.0/e347ab1c932092252e717ff3f949105a4dd28b27e842dd53157d2f72e276c2e4)
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 1191.56it/s]
Loading cached processed dataset at /root/.cache/huggingface/datasets/json/default-f82d1afe86c1e9ec/0.0.0/e347ab1c932092252e717ff3f949105a4dd28b27e842dd53157d2f72e276c2e4/cache-baa6cdf34a027bbb.arrow
Loading cached shuffled indices for dataset at /root/.cache/huggingface/datasets/json/default-f82d1afe86c1e9ec/0.0.0/e347ab1c932092252e717ff3f949105a4dd28b27e842dd53157d2f72e276c2e4/cache-8aa40269a670f4fd.arrow
Loading cached processed dataset at /root/.cache/huggingface/datasets/json/default-f82d1afe86c1e9ec/0.0.0/e347ab1c932092252e717ff3f949105a4dd28b27e842dd53157d2f72e276c2e4/cache-dd26c6462b17896e.arrow
wandb: Tracking run with wandb version 0.15.3
wandb: W&B syncing is set to `offline` in this directory.  
wandb: Run `wandb online` or set WANDB_MODE=online to enable cloud syncing.
  0%|                                                                                                                                                                                                       | 0/160 [00:00<?, ?it/s]`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...
/root/anaconda3/lib/python3.11/site-packages/torch/utils/checkpoint.py:429: UserWarning: torch.utils.checkpoint: please pass in use_reentrant=True or use_reentrant=False explicitly. The default value of use_reentrant will be updated to be False in the future. To maintain current behavior, pass use_reentrant=True. It is recommended that you use use_reentrant=False. Refer to docs for more details on the differences between the two variants.
  warnings.warn(
{
   'loss': 1.9799, 'learning_rate': 6.25e-05, 'epoch': 0.12}                                                                                                                                                                          
{
   'loss': 2.8439, 'learning_rate': 0.000125, 'epoch': 0.24}                                                                                                                                                                          
{
   'loss': 2.6293, 'learning_rate': 0.0001875, 'epoch': 0.35}                                                                                                                                                                         
{
   'loss': 2.6095, 'learning_rate': 0.00025, 'epoch': 0.47}                                                                                                                                                                           
{
   'loss': 2.2325, 'learning_rate': 0.0003125, 'epoch': 0.59}                                                                                                                                                                         
{
   'eval_loss': 2.7306337356567383, 'eval_runtime': 0.1659, 'eval_samples_per_second': 12.057, 'eval_steps_per_second': 12.057, 'epoch': 0.59}                                                                                        
  3%|█████▉                                                                                                                                                                                         | 5/160 [00:04<01:36,  1.60it/s/root/anaconda3/lib/python3.11/site-packages/torch/utils/checkpoint.py:429: UserWarning: torch.utils.checkpoint: please pass in use_reentrant=True or use_reentrant=False explicitly. The default value of use_reentrant will be updated to be False in the future. To maintain current behavior, pass use_reentrant=True. It is recommended that you use use_reentrant=False. Refer to docs for more details on the differences between the two variants.
  warnings.warn(
{
   'loss': 2.4916, 'learning_rate': 0.000375, 'epoch': 0.71}                                                                                                                                                                          
{
   'loss': 2.4591, 'learning_rate': 0.0004375, 'epoch': 0.82}                                                                                                                                                                         
{
   'loss': 2.0441, 'learning_rate': 0.0005, 'epoch': 0.94}                                                                                                                                                                            
{
   'loss': 1.8674, 'learning_rate': 0.0005625000000000001, 'epoch': 1.06}                                                                                                                                                             
{
   'loss': 1.5093, 'learning_rate': 0.000625, 'epoch': 1.18}                                                                                                                                                                          
{
   'eval_loss': 1.626299262046814, 'eval_runtime': 0.1665, 'eval_samples_per_second': 12.013, 'eval_steps_per_second': 12.013, 'epoch': 1.18}                                                                                         
  6%|███████████▉                                                                                                                                                                                  | 10/160 [00:06<01:23,  1.80it/s/root/anaconda3/lib/python3.11/site-packages/torch/utils/checkpoint.py:429: UserWarning: torch.utils.checkpoint: please pass in use_reentrant=True or use_reentrant=False explicitly. The default value of use_reentrant will be updated to be False in the future. To maintain current behavior, pass use_reentrant=True. It is recommended that you use use_reentrant=False. Refer to docs for more details on the differences between the two variants.
  warnings.warn(
{
   'loss': 1.7075, 'learning_rate': 0.0006875, 'epoch': 1.29}                                                                                                                                                                         
{
   'loss': 1.6792, 'learning_rate': 0.00075, 'epoch': 1.41}                                                                                                                                                                           
{
   'loss': 1.4942, 'learning_rate': 0.0008125000000000001, 'epoch': 1.53}                                                                                                                                                             
{
   'loss': 1.8202, 'learning_rate': 0.000875, 'epoch': 1.65}                                                                                                                                                                          
{
   'loss': 0.9729, 'learning_rate': 0.0009375, 'epoch': 1.76}                                                                                                                                                                         
{
   'eval_loss': 0.7719208002090454, 'eval_runtime': 0.1673, 'eval_samples_per_second': 11.953, 'eval_steps_per_second': 11.953, 'epoch': 1.76}                                                                                        
  9%|█████████████████▊                                                                                                                                                                            | 15/160 [00:09<01:20,  1.81it/s/root/anaconda3/lib/python3.11/site-packages/torch/utils/checkpoint.py:429: UserWarning: torch.utils.checkpoint: please pass in use_reentrant=True or use_reentrant=False explicitly. The default value of use_reentrant will be updated to be False in the future. To maintain current behavior, pass use_reentrant=True. It is recommended that you use use_reentrant=False. Refer to docs for more details on the differences between the two variants.
  warnings.warn(
{
   'loss': 1.3478, 'learning_rate': 0.001, 'epoch': 1.88}                                                                                                                                                                             
{
   'loss': 1.3449, 'learning_rate': 0.0009930555555555556, 'epoch': 2.0}                                                                                                                                                              
{
   'loss': 0.6173, 'learning_rate': 0.0009861111111111112, 'epoch': 2.12}                                                                                                                                                             
{
   'loss': 0.5325, 'learning_rate': 0.0009791666666666666, 'epoch': 2.24}                                                                                                                                                             
{
   'loss': 1.1995, 'learning_rate': 0.0009722222222222222, 'epoch': 2.35}                                                                                                                                                             
{
   'eval_loss': 0.06268511712551117, 'eval_runtime': 0.1694, 'eval_samples_per_second': 11.804, 'eval_steps_per_second': 11.804, 'epoch': 2.35}                                                                                       
 12%|███████████████████████▊                                                                                                                                                                      | 20/160 [00:12<01:17,  1.81it/s/root/anaconda3/lib/python3.11/site-packages/torch/utils/checkpoint.py:429: UserWarning: torch.utils.checkpoint: please pass in use_reentrant=True or use_reentrant=False explicitly. The default value of use_reentrant will be updated to be False in the future. To maintain current behavior, pass use_reentrant=True. It is recommended that you use use_reentrant=False. Refer to docs for more details on the differences between the two variants.
  warnings.warn(
{
   'loss': 1.0089, 'learning_rate': 0.0009652777777777778, 'epoch': 2.47}                                                                                                                                                             
{
   'loss': 0.9793, 'learning_rate': 0.0009583333333333334, 'epoch': 2.59}                                                                                                                                                             
{
   'loss': 0.814, 'learning_rate': 0.0009513888888888889, 'epoch': 2.71}                                                                                                                                                              
{
   'loss': 1.1905, 'learning_rate': 0.0009444444444444445, 'epoch'