Merge pull request mst272#13 from aJupyter/main

update readme and run_example.sh

README.md

@@ -61,7 +61,8 @@ RLHF训练框架，支持并持续更新Reward训练、PPO、DPO、RLOO、SimPO
 - [x] [deepseek-coder](https://github.com/deepseek-ai/DeepSeek-Coder)
 - [x] [哔哩哔哩 Index-1.9B](https://github.com/bilibili/Index-1.9B)
 - [x] [baichuan系列](https://github.com/baichuan-inc/Baichuan2)
-- 待更新GLM
+- [x] [GLM系列](https://github.com/THUDM/GLM-4)
+- 待更新Mistral系列
 ### 已更新tricks讲解
  所有相关的trciks及讲解都在llm_tricks文件夹下
 - [Dora代码讲解（llm_tricks/dora/READEME.md）](./llm_tricks/dora/READEME.md)

main_train.py

@@ -6,6 +6,7 @@
 from transformers import AutoConfig, AutoTokenizer, AutoModelForCausalLM, Trainer, \
     BitsAndBytesConfig, HfArgumentParser, set_seed
 from peft import LoraConfig, TaskType, get_peft_model, prepare_model_for_kbit_training, cast_mixed_precision_params
+from train_args import dpo_TrainArgument, sft_TrainArgument
 import bitsandbytes as bnb
 from utils.template import template_dict
 from utils.data_process import MultiRoundDataProcess, DpoDataset
@@ -31,8 +32,6 @@ def load_config(train_args_path):
 
 
 def initial_args():
-    # parser = HfArgumentParser((CommonArgs, TrainArgument))
-    # reward_args, train_args = parser.parse_args_into_dataclasses()
     parser = HfArgumentParser((CommonArgs,))
     args = parser.parse_args_into_dataclasses()[0]
     # 根据CommonArgs中的config_option动态加载配置

run_example.sh

@@ -0,0 +1,10 @@
+deepspeed --include localhost:0,1 main_train.py\
+    --train_data_path 数据集路径\
+    --model_name_or_path  模型路径\
+    --task_type sft\
+    --train_mode qlora\
+    --output_dir 输出路径
+
+# task_type:[pretrain, sft, dpo_multi, dpo_single]
+
+# python main_train.py --train_data_path 数据集路径 --model_name_or_path 模型路径

train_args/__init__.py

@@ -0,0 +1,7 @@
+from .dpo.dpo_config import TrainArgument as dpo_TrainArgument
+from .sft.lora_qlora.base import TrainArgument as sft_TrainArgument
+
+__all__ = [
+    "dpo_TrainArgument",
+    "sft_TrainArgument",
+]

utils/args.py

@@ -20,7 +20,7 @@ class CommonArgs:
     一些常用的自定义参数
     """
     # Deepspeed相关参数
-    local_rank: int = field(default=1, metadata={"help": "deepspeed所需参数,单机无需修改"})
+    local_rank: int = field(default=1, metadata={"help": "deepspeed所需参数,单机无需修改，如出现报错可注释掉"})
 
     train_args_path: TrainArgPath = field(default=TrainArgPath.SFT_LORA_QLORA_BASE.value,
                                           metadata={"help": "当前模式的训练参数,分为sft和dpo参数"})