Updata codes

README.md

 # GLM-4-9b
 ## 论文
-`GLM: General Language Model Pretraining with Autoregressive Blank Infilling`
-- https://arxiv.org/abs/2103.10360
 
 ## 模型结构
 
@@ -10,7 +8,7 @@
 </div>
 
 ## 算法原理
-
+GLM-4-9B是智谱AI推出的最新一代预训练模型GLM-4系列中的开源版本，在语义、数学、推理、代码和知识等多方面的数据集测评中，GLM-4-9B及其人类偏好对齐的版本GLM-4-9B-Chat均表现出超越Llama-3-8B的卓越性能
 <div align=center>
     <img src="./doc/xx.png" witdh=800 height=300/>
 </div>
@@ -62,13 +60,31 @@ pip install -r requirements.txt
 
 ## 数据集
 ### 准备数据集
-对于数据文件，样例采用如下格式。这里是一个不带有工具的例子:
+本仓库以[ADGEN](https://aclanthology.org/D19-1321.pdf) (广告生成) 数据集为例介绍代码的使用方法，可通过[Google Drive](https://drive.google.com/file/d/13_vf0xRTQsyneRKdD1bZIr93vBGOczrk/view?usp=sharing) 或者 [Tsinghua Cloud](https://cloud.tsinghua.edu.cn/f/b3f119a008264b1cabd1/?dl=1)下载处理好的 ADGEN 数据集。数据集下载完成后，将数据解压到[data](./data)目录下。
+
+数据按路径存放后，执行下面的数据转换代码，生成的`dev.jsonl``train.jsonl`默认保存在`AdvertiseGen/saves`目录下:
+```
+python gen_messages_data.py --data_path /path/to/AdvertiseGen
+```
+
+数据集目录结构如下：
+```
+├── data
+│   ├── AdvertiseGen
+│       ├── saves # 生成的
+│       ├── dev.json
+│       └── train.json
+```
+
+若想生成自己的数据文件，代码可参考[gen_messages_data.py](./gen_messages_data.py)进行修改，样例采用如下格式。
+
+- 这里是一个不带有工具的例子:
 
 ```
 {"messages": [{"role": "user", "content": "类型#裤*材质#牛仔布*风格#性感"}, {"role": "assistant", "content": "3x1的这款牛仔裤采用浅白的牛仔面料为裤身材质，其柔然的手感和细腻的质地，在穿着舒适的同时，透露着清纯甜美的个性气质。除此之外，流畅的裤身剪裁将性感的腿部曲线彰显的淋漓尽致，不失为一款随性出街的必备单品。"}]}
 ```
 
-这是一个带有工具调用的例子:
+- 这是一个带有工具调用的例子:
 
 ```
 {"messages": [{"role": "system", "content": "", "tools": [{"type": "function", "function": {"name": "get_recommended_books", "description": "Get recommended books based on user's interests", "parameters": {"type": "object", "properties": {"interests": {"type": "array", "items": {"type": "string"}, "description": "The interests to recommend books for"}}, "required": ["interests"]}}}]}, {"role": "user", "content": "Hi, I am looking for some book recommendations. I am interested in history and science fiction."}, {"role": "assistant", "content": "{\"name\": \"get_recommended_books\", \"arguments\": {\"interests\": [\"history\", \"science fiction\"]}}"}, {"role": "observation", "content": "{\"books\": [\"Sapiens: A Brief History of Humankind by Yuval Noah Harari\", \"A Brief History of Time by Stephen Hawking\", \"Dune by Frank Herbert\", \"The Martian by Andy Weir\"]}"}, {"role": "assistant", "content": "Based on your interests in history and science fiction, I would recommend the following books: \"Sapiens: A Brief History of Humankind\" by Yuval Noah Harari, \"A Brief History of Time\" by Stephen Hawking, \"Dune\" by Frank Herbert, and \"The Martian\" by Andy Weir."}]}
@@ -87,7 +103,7 @@ pip install -r requirements.txt
 ```
 
 2. 配置文件位于configs目录下，包括以下文件：
-- `deepspeed配置文件`：[ds_zereo_2](./finetune_demo/configs/ds_zereo_2.json),[ds_zereo_3](./finetune_demo/configs/ds_zereo_3.json)
+- `deepspeed配置文件`：[ds_zereo_2](./finetune_demo/configs/ds_zereo_2.json)，[ds_zereo_3](./finetune_demo/configs/ds_zereo_3.json)
 - `lora.yaml/ ptuning_v2.yaml / sft.yaml`: 模型不同方式的配置文件，包括模型参数、优化器参数、训练参数等。部分重要参数解释如下：
     + data_config 部分
         + train_file: 训练数据集的文件路径。
@@ -125,54 +141,59 @@ pip install -r requirements.txt
         + num_attention_heads: 2: P-TuningV2 的注意力头数(不要改动)。
         + token_dim: 256: P-TuningV2 的 token 维度(不要改动)。
 
+3. `data/AdvertiseGen/saves/`是`.jsonl`数据地址，`THUDM/glm-4-9b-chat`为模型地址，`configs/lora.yaml`为配置文件地址，以上参数均可根据自身数据地址进行替换。
+
 ### 单机单卡
 ```shell
-python finetune_hf.py  data/AdvertiseGen/  THUDM/glm-4-9b-chat  configs/lora.yaml
+bash train.sh
 ```
 
 ### 单机多卡/多机多卡
-这是使用 `deepspeed` 作为加速方案的，您需要安装 `deepspeed`。
+这里使用`deepspeed`作为加速方案，请确认当前环境已经根据[环境配置章节](#环境配置)安装好了`deepspeed`库。
 ```shell
-OMP_NUM_THREADS=1 torchrun --standalone --nnodes=1 --nproc_per_node=8  finetune_hf.py  data/AdvertiseGen/  THUDM/glm-4-9b  configs/lora.yaml
+bash train_dp.sh
 ```
+
 ### 从保存点进行微调
 如果按照上述方式进行训练，每次微调都会从头开始，如果你想从训练一半的模型开始微调，你可以加入第四个参数，这个参数有两种传入方式:
 
-1. `yes`, 自动从最后一个保存的 Checkpoint开始训练
-2. `XX`, 断点号数字 例 `600` 则从序号600 Checkpoint开始训练
-例如，这就是一个从最后一个保存点继续微调的示例代码
+1. `yes`, 自动从**最后一个保存的Checkpoint**开始训练，例如:
+```shell
+python finetune.py ../data/AdvertiseGen/saves/ ../checkpoints/glm-4-9b-chat/ configs/lora.yaml yes
+```
 
+2. `XX`, 断点号数字，例`600`则从序号**600 Checkpoint**开始训练，例如：
 ```shell
-python finetune_hf.py  data/AdvertiseGen/  THUDM/glm-4-9b-chat  configs/lora.yaml yes
+python finetune.py ../data/AdvertiseGen/saves/ ../checkpoints/glm-4-9b-chat/ configs/lora.yaml 600
 ```
 
 ## 推理
 进入[basic_demo](./basic_demo/)目录下
 
 ### 快速调用
+**参数解释**
+- --model_name_or_path:待测模型名或模型地址，当前默认"THUDM/glm-4-9b-chat"
+- --device: 当前默认"cuda"
+- --query: 待测输入语句，当前默认"你好"
+
 ```
 python inference.py
 ```
 
 ### 使用命令行与 GLM-4-9B 模型进行对话
 ```
-python trans_cli_demo.py # GLM-4-9B-Chat
-python trans_cli_vision_demo.py # GLM-4V-9B
+python trans_cli_demo.py --model_name_or_path ../checkpoints/GLM-4-9B-Chat
+python trans_cli_vision_demo.py --model_name_or_path ../checkpoints/GLM-4V-9B
 ```
 
 ### 使用 Gradio 网页端与 GLM-4-9B-Chat 模型进行对话
 ```
-python trans_web_demo.py
-```
-
-### 使用 Batch 推理
-```
-python cli_batch_request_demo.py
+python trans_web_demo.py --model_name_or_path ../checkpoints/GLM-4-9B-Chat
 ```
 
 ## result
 <div align=center>
-    <img src="./doc/xx.png" width=1500 heigh=400/>
+    <img src="./doc/result.png" width=1500 heigh=400/>
 </div>
 
 ### 精度

basic_demo/inference.py

 import torch
+import argparse
+
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
-device = "cuda"
-tokenizer = AutoTokenizer.from_pretrained("THUDM/glm-4-9b-chat", trust_remote_code=True)
+parse = argparse.ArgumentParser()
+parse.add_argument('--model_name_or_path', default="THUDM/glm-4-9b-chat")
+parse.add_argument('--device', default="cuda")
+parse.add_argument('--query', type=str, default="你好")
+args = parse.parse_args()
+
+device = args.device
+model_name_or_path = args.model_name_or_path
+
+tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=True)
 
-query = "你好"
+query = args.query
 inputs = tokenizer.apply_chat_template([{"role": "user", "content": query}],
                                        add_generation_prompt=True,
                                        tokenize=True,
@@ -14,7 +24,7 @@ inputs = tokenizer.apply_chat_template([{"role": "user", "content": query}],
 
 inputs = inputs.to(device)
 model = AutoModelForCausalLM.from_pretrained(
-    "THUDM/glm-4-9b-chat",
+    model_name_or_path,
     torch_dtype=torch.bfloat16,
     low_cpu_mem_usage=True,
     trust_remote_code=True
@@ -24,4 +34,4 @@ gen_kwargs = {"max_length": 2500, "do_sample": True, "top_k": 1}
 with torch.no_grad():
     outputs = model.generate(**inputs, **gen_kwargs)
     outputs = outputs[:, inputs['input_ids'].shape[1]:]
-    print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+    print('Result', tokenizer.decode(outputs[0], skip_special_tokens=True))

basic_demo/trans_cli_demo.py

@@ -11,6 +11,7 @@ ensuring that the CLI interface displays formatted text correctly.
 """
 
 import os
+import argparse
 import torch
 from threading import Thread
 from typing import Union
@@ -27,11 +28,16 @@ from transformers import (
     TextIteratorStreamer
 )
 
+# add model path
+parser = argparse.ArgumentParser()
+parser.add_argument('--model_name_or_path', default='THUDM/glm-4-9b-chat')
+args = parser.parse_args()
+
 ModelType = Union[PreTrainedModel, PeftModelForCausalLM]
 TokenizerType = Union[PreTrainedTokenizer, PreTrainedTokenizerFast]
 
-MODEL_PATH = os.environ.get('MODEL_PATH', 'THUDM/glm-4-9b-chat')
-
+# MODEL_PATH = os.environ.get('MODEL_PATH', 'THUDM/glm-4-9b-chat')
+MODEL_PATH = args.model_name_or_path
 
 def load_model_and_tokenizer(
         model_dir: Union[str, Path], trust_remote_code: bool = True

basic_demo/trans_cli_vision_demo.py

@@ -11,6 +11,8 @@ ensuring that the CLI interface displays formatted text correctly.
 """
 
 import os
+import argparse
+
 import torch
 from threading import Thread
 from transformers import (
@@ -22,13 +24,20 @@ from transformers import (
 
 from PIL import Image
 
-MODEL_PATH = os.environ.get('MODEL_PATH', 'THUDM/glm-4v-9b')
+# add model path
+parser = argparse.ArgumentParser()
+parser.add_argument('--model_name_or_path', default='THUDM/glm-4v-9b')
+args = parser.parse_args()
+
+# MODEL_PATH = os.environ.get('MODEL_PATH', 'THUDM/glm-4v-9b')
+MODEL_PATH = args.model_name_or_path
 
 tokenizer = AutoTokenizer.from_pretrained(
     MODEL_PATH,
     trust_remote_code=True,
     encode_special_tokens=True
 )
+
 model = AutoModel.from_pretrained(
     MODEL_PATH,
     trust_remote_code=True,

basic_demo/trans_web_demo.py

@@ -6,10 +6,12 @@ allowing users to interact with the model through a chat-like interface.
 """
 
 import os
-import gradio as gr
+import argparse
 import torch
-from threading import Thread
 
+import gradio as gr
+
+from threading import Thread
 from typing import Union
 from pathlib import Path
 from peft import AutoPeftModelForCausalLM, PeftModelForCausalLM
@@ -26,8 +28,14 @@ from transformers import (
 
 ModelType = Union[PreTrainedModel, PeftModelForCausalLM]
 TokenizerType = Union[PreTrainedTokenizer, PreTrainedTokenizerFast]
+# add model path
+parser = argparse.ArgumentParser()
+parser.add_argument('--model_name_or_path', default='THUDM/glm-4-9b-chat')
+args = parser.parse_args()
+
+# MODEL_PATH = os.environ.get('MODEL_PATH', 'THUDM/glm-4-9b-chat')
+MODEL_PATH = args.model_name_or_path
 
-MODEL_PATH = os.environ.get('MODEL_PATH', 'THUDM/glm-4-9b-chat')
 TOKENIZER_PATH = os.environ.get("TOKENIZER_PATH", MODEL_PATH)
 
 

basic_demo/vllm_cli_demo.py

@@ -9,14 +9,22 @@ Usage:
 Note: The script includes a modification to handle markdown to plain text conversion,
 ensuring that the CLI interface displays formatted text correctly.
 """
+
 import time
 import asyncio
+import argparse
+
 from transformers import AutoTokenizer
 from vllm import SamplingParams, AsyncEngineArgs, AsyncLLMEngine
 from typing import List, Dict
 
-MODEL_PATH = 'THUDM/glm-4-9b'
+# add model path
+parser = argparse.ArgumentParser()
+parser.add_argument('--model_name_or_path', default='THUDM/glm-4-9b')
+args = parser.parse_args()
 
+# MODEL_PATH = 'THUDM/glm-4-9b'
+MODEL_PATH = args.model_name_or_path
 
 def load_model_and_tokenizer(model_dir: str):
     engine_args = AsyncEngineArgs(

finetune_demo/requirements.txt

 jieba>=0.42.1
 datasets>=2.19.1
 peft>=0.11.0
-nltk==3.8.1
+nltk==3.8.1
\ No newline at end of file
+ruamel.yaml==0.18.6
+rouge_chinese==1.0.3
\ No newline at end of file

finetune_demo/train.sh

+#!/bin/bash
+export HIP_VISIBLE_DEVICES=1 # 可自行修改为指定显卡号
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export USE_MIOPEN_BATCHNORM=1
+
+python finetune.py ../data/AdvertiseGen/saves/ ../checkpoints/glm-4-9b-chat/ configs/lora.yaml
\ No newline at end of file

finetune_demo/train_dp.sh

+#!/bin/bash
+export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 # 可自行修改为指定显卡号
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export USE_MIOPEN_BATCHNORM=1
+
+OMP_NUM_THREADS=1 torchrun --standalone --nnodes=1 --nproc_per_node=8  finetune.py ../data/AdvertiseGen/saves/ ../checkpoints/glm-4-9b-chat/ configs/lora.yaml
\ No newline at end of file

gen_messages_data.py

+import os
+import json
+import argparse
+
+# 配置数据
+parse = argparse.ArgumentParser()
+parse.add_argument('--data_path', default='./data/AdvertiseGen')
+args = parse.parse_args()
+
+# 默认保存路径
+save_root_path = os.path.join(args.data_path, 'saves')
+if not os.path.exists(save_root_path):
+    os.mkdir(save_root_path)
+
+
+def save_to_jsonl(train_infos, save_path):
+    '''将json数据保存到.jsonl文件中'''
+
+    with open(save_path, 'w', encoding='utf-8') as file:
+        for info in train_infos:
+            file.write(json.dumps(info, ensure_ascii=False)+'\n')
+    file.close()
+
+def load_json_infos(file_path):
+    '''读取json数据'''
+    all_data = []
+    with open(file_path, 'r', encoding='utf-8') as ofile:
+        for info in ofile.readlines():
+            json_info = json.loads(info)
+            output = {"messages": []}
+            content = {"role": "user", "content": json_info.get("content")}
+            summary = {"role": "assistant", "content": json_info.get("summary")}
+            output["messages"].extend([content, summary])
+            all_data.append(output)
+
+    save_file_path = os.path.join(save_root_path, os.path.basename(file_path)+'l')
+    save_to_jsonl(all_data, save_file_path)
+
+
+if __name__ == "__main__":
+
+    files = ['train.json', 'dev.json']
+    for file in files:
+        file_path = os.path.join(args.data_path, file)
+        output = load_json_infos(file_path)

model.properties

@@ -3,8 +3,8 @@ modelCode=684
 # 模型名称
 modelName=glm4-9b_pytorch
 # 模型描述
-modelDescription=GLM-4-9B 是智谱 AI 推出的最新一代预训练模型 GLM-4 系列中的开源版本
+modelDescription=GLM-4-9B是智谱AI推出的最新一代预训练模型GLM-4系列中的开源版本，在语义、数学、推理、代码和知识等多方面的数据集测评中，GLM-4-9B及其人类偏好对齐的版本GLM-4-9B-Chat均表现出超越Llama-3-8B的卓越性能。
 # 应用场景
-appScenario=推理,多轮对话,家居,教育,科研
+appScenario=推理,训练,多轮对话,家居,教育,科研
 # 框架类型
 frameType=pytorch