featmodel): 添加保存模式选择功能

在模型管理页面中新增保存模式选择功能，用户可以通过下拉菜单选择不同的保存模式（如默认、合并16位、合并4位等）。同时，将保存模型的逻辑抽离到独立的`save_model.py`文件中，以提高代码的可维护性和复用性。

README.md

@@ -1,15 +1,91 @@
 # 基于文档驱动的自适应编码大模型微调框架
+
 ## 简介
-本人的毕业设计
+
 ### 项目概述
+本项目是一个基于文档驱动的自适应编码大模型微调框架，通过深度解析私有库的文档以及其他资源，生成指令型语料，据此对大语言模型进行针对私有库的微调。
 
-* 通过深度解析私有库的文档以及其他资源，生成指令型语料，据此对大语言模型进行针对私有库的微调。
+### 核心功能
+- 文档解析与语料生成
+- 大语言模型高效微调
+- 交互式训练与推理界面
+- 训练过程可视化监控
 
-### 项目技术
+## 技术架构
 
-* 使用unsloth框架在GPU上实现大语言模型的qlora微调
-* 使用langchain框架编写工作流实现批量生成微调语料
-* 使用tinydb和sqlite实现数据的持久化
-* 使用gradio框架实现前端展示
+### 系统架构
+```
+[前端界面] -> [模型微调] -> [数据存储]
+    ↑               ↑             ↑
+    │               │             │
+[Gradio]     [unsloth/QLoRA]  [SQLite/TinyDB]
+```
 
-**施工中......**
+### 技术栈
+- **前端界面**: Gradio构建的交互式Web界面
+- **模型微调**: 基于unsloth框架的QLoRA高效微调
+- **数据存储**: SQLite(结构化数据) + TinyDB(非结构化数据)
+- **工作流引擎**: LangChain实现文档解析与语料生成
+- **训练监控**: TensorBoard集成
+
+## 功能模块
+
+### 1. 模型管理
+- 支持多种格式的大语言模型加载
+- 模型信息查看与状态管理
+- 模型卸载与内存释放
+
+### 2. 模型推理
+- 对话式交互界面
+- 流式响应输出
+- 历史对话管理
+
+### 3. 模型微调
+- 训练参数配置(学习率、batch size等)
+- LoRA参数配置(秩、alpha等)
+- 训练过程实时监控
+- 训练中断与恢复
+
+### 4. 数据集生成
+- 文档解析与清洗
+- 指令-响应对生成
+- 数据集质量评估
+- 数据集版本管理
+
+## 微调技术
+
+### QLoRA原理
+QLoRA(Quantized Low-Rank Adaptation)是一种高效的大模型微调技术，核心特点包括：
+1. **4-bit量化**: 将预训练模型量化为4-bit表示，大幅减少显存占用
+2. **低秩适配**: 通过低秩矩阵分解(LoRA)实现参数高效更新
+3. **内存优化**: 使用梯度检查点等技术进一步降低显存需求
+
+### 参数配置
+- **学习率**: 建议2e-5到2e-4
+- **LoRA秩**: 控制适配器复杂度(建议16-64)
+- **LoRA Alpha**: 控制适配器更新幅度(通常设为秩的1-2倍)
+
+### 训练监控
+- **TensorBoard集成**: 实时查看损失曲线、学习率等指标
+- **日志记录**: 训练过程详细日志保存
+- **模型检查点**: 定期保存中间权重
+
+## 快速开始
+
+1. 安装依赖:
+```bash
+pip install -r requirements.txt
+```
+
+2. 启动应用:
+```bash
+python main.py
+```
+
+3. 访问Web界面:
+```
+http://localhost:7860
+```
+
+## 许可证
+MIT License

db/__init__.py

@@ -1,11 +1,12 @@
-from .init_db import get_sqlite_engine, initialize_sqlite_db
+from .init_db import load_sqlite_engine, initialize_sqlite_db
 from .prompt_store import get_prompt_tinydb, initialize_prompt_store
-from .dataset_store import get_all_dataset
+from .dataset_store import get_all_dataset, save_dataset
 
 __all__ = [
-    "get_sqlite_engine",
+    "load_sqlite_engine",
     "initialize_sqlite_db",
     "get_prompt_tinydb",
     "initialize_prompt_store",
-    "get_all_dataset"
+    "get_all_dataset",
+    "save_dataset"
 ]

db/dataset_store.py

@@ -8,7 +8,7 @@ from tinydb.storages import MemoryStorage
 
 # 将项目根目录添加到系统路径中，以便能够导入项目中的其他模块
 sys.path.append(str(Path(__file__).resolve().parent.parent))
-from schema.dataset import dataset, dataset_item, Q_A
+from schema.dataset import Dataset, DatasetItem, Q_A
 
 def get_all_dataset(workdir: str) -> TinyDB:
     """
@@ -39,6 +39,30 @@ def get_all_dataset(workdir: str) -> TinyDB:
     return db
 
 
+def save_dataset(db: TinyDB, workdir: str, name: str = None) -> None:
+    """
+    将TinyDB中的数据集保存为单独的json文件
+    
+    Args:
+        db (TinyDB): 包含数据集对象的TinyDB实例
+        workdir (str): 工作目录路径
+        name (str, optional): 要保存的数据集名称，None表示保存所有
+    """
+    dataset_dir = os.path.join(workdir, "dataset")
+    os.makedirs(dataset_dir, exist_ok=True)
+    
+    datasets = db.all() if name is None else db.search(Query().name == name)
+    
+    for dataset in datasets:
+        try:
+            filename = f"{dataset.get(dataset['name'])}.json"
+            filepath = os.path.join(dataset_dir, filename)
+            
+            with open(filepath, "w", encoding="utf-8") as f:
+                json.dump(dataset, f, ensure_ascii=False, indent=2)
+        except Exception as e:
+            print(f"Error saving dataset {dataset.get('id', 'unknown')}: {str(e)}")
+
 if __name__ == "__main__":
     # 定义工作目录路径
     workdir = os.path.join(os.path.dirname(__file__), "..", "workdir")
@@ -47,4 +71,11 @@ if __name__ == "__main__":
     # 打印结果
     print(f"Found {len(datasets)} datasets:")
     for ds in datasets.all():
-        print(f"- {ds['name']} (ID: {ds['id']})")
+        print(f"- {ds['name']} (ID: {ds['id']})")
+    
+    # 询问要保存的数据集名称
+    name = input("输入要保存的数据集名称(直接回车保存所有): ").strip() or None
+    
+    # 保存数据集到文件
+    save_dataset(datasets, workdir, name)
+    print(f"Datasets {'all' if name is None else name} saved to json files")

db/init_db.py

@@ -14,7 +14,7 @@ from schema.dataset_generation import APIProvider
 # 全局变量，用于存储数据库引擎实例
 _engine: Optional[Engine] = None
 
-def get_sqlite_engine(workdir: str) -> Engine:
+def load_sqlite_engine(workdir: str) -> Engine:
     """
     获取数据库引擎实例。如果引擎尚未创建，则创建一个新的引擎并返回。
 
@@ -74,6 +74,6 @@ if __name__ == "__main__":
     # 定义工作目录路径
     workdir = os.path.join(os.path.dirname(__file__), "..", "workdir")
     # 获取数据库引擎
-    engine = get_sqlite_engine(workdir)
+    engine = load_sqlite_engine(workdir)
     # 初始化数据库
     initialize_sqlite_db(engine)

db/prompt_store.py

@@ -44,12 +44,12 @@ def initialize_prompt_store(db: TinyDB) -> None:
     # 检查数据库是否为空
     if not db.all():  # 如果数据库中没有数据
         db.insert(promptTempleta(
-            id=0,
+            id=1,
             name="default",
             description="默认提示词模板",
-            content="""项目名为:{ project_name }
+            content="""项目名为:{project_name}
 请依据以下该项目官方文档的部分内容，创造合适的对话数据集用于微调一个了解该项目的小模型的语料，要求兼顾文档中间尽可能多的信息点，使用中文
-文档节选：{ content }""").model_dump())
+文档节选：{document_slice}""").model_dump())
     # 如果数据库中已有数据，则跳过插入
 
 

frontend/dataset_generate_page.py

@@ -1,42 +1,189 @@
 import gradio as gr
-from tools import scan_docs_directory
-from global_var import get_docs, scan_docs_directory, get_prompt_store
+import sys
+import json
+from tinydb import Query
+from pathlib import Path
+from langchain.prompts import PromptTemplate
+from sqlmodel import Session, select
+from schema import Dataset, DatasetItem, Q_A
+from db.dataset_store import get_all_dataset
+
+sys.path.append(str(Path(__file__).resolve().parent.parent))
+from schema import APIProvider, LLMRequest, LLMResponse, DatasetItem
+from db import save_dataset
+from tools import call_openai_api, process_markdown_file, generate_json_example
+from global_var import get_docs, get_prompt_store, get_sql_engine, get_datasets, get_workdir
 
 def dataset_generate_page():
     with gr.Blocks() as demo:
         gr.Markdown("## 数据集生成")
         with gr.Row():
-            with gr.Column():
-                # 获取文档列表并设置初始值
+            with gr.Column(scale=1):
                 docs_list = [str(doc.name) for doc in get_docs()]
                 initial_doc = docs_list[0] if docs_list else None
+                prompts = get_prompt_store().all()
+                prompt_list = [f"{p['id']} {p['name']}" for p in prompts]
+                initial_prompt = prompt_list[0] if prompt_list else None
                 
+                # 初始化Dataframe的值
+                initial_dataframe_value = []
+                if initial_prompt:
+                    selected_prompt_id = int(initial_prompt.split(" ")[0])
+                    prompt_data = get_prompt_store().get(doc_id=selected_prompt_id)
+                    prompt_content = prompt_data["content"]
+                    prompt_template = PromptTemplate.from_template(prompt_content)
+                    input_variables = prompt_template.input_variables
+                    input_variables.remove("document_slice")
+                    initial_dataframe_value = [[var, ""] for var in input_variables]
+                # 从数据库获取API Provider列表
+                with Session(get_sql_engine()) as session:
+                    providers = session.exec(select(APIProvider)).all()
+                    api_list = [f"{p.id} {p.model_id}" for p in providers]
+                initial_api = api_list[0] if api_list else None
+                
+                api_dropdown = gr.Dropdown(
+                    choices=api_list,
+                    value=initial_api,
+                    label="选择API",
+                    interactive=True
+                )
                 doc_dropdown = gr.Dropdown(
                     choices=docs_list,
-                    value=initial_doc,  # 设置初始选中项
+                    value=initial_doc,
                     label="选择文档",
-                    allow_custom_value=True,
                     interactive=True
                 )
-                doc_state = gr.State(value=initial_doc)  # 用文档初始值初始化状态
-
-            with gr.Column():
-                # 获取模板列表并设置初始值
-                prompts = get_prompt_store().all()
-                prompt_choices = [f"{p['id']} {p['name']}" for p in prompts]
-                initial_prompt = prompt_choices[0] if prompt_choices else None
-                
                 prompt_dropdown = gr.Dropdown(
-                    choices=prompt_choices,
-                    value=initial_prompt,  # 设置初始选中项
+                    choices=prompt_list,
+                    value=initial_prompt,
                     label="选择模板",
-                    allow_custom_value=True,
                     interactive=True
                 )
-                prompt_state = gr.State(value=initial_prompt)  # 用模板初始值初始化状态
+                rounds_input = gr.Number(
+                    value=1,
+                    label="生成轮次",
+                    minimum=1,
+                    maximum=100,
+                    step=1,
+                    interactive=True
+                )
+                concurrency_input = gr.Number(
+                    value=1,
+                    label="并发数",
+                    minimum=1,
+                    maximum=10,
+                    step=1,
+                    interactive=True,
+                    visible=False
+                )
+                dataset_name_input = gr.Textbox(
+                    label="数据集名称",
+                    placeholder="输入数据集保存名称",
+                    interactive=True
+                )
+                prompt_choice = gr.State(value=initial_prompt)
+                generate_button = gr.Button("生成数据集",variant="primary")
+                doc_choice = gr.State(value=initial_doc)
+                output_text = gr.Textbox(label="生成结果", interactive=False)
+                api_choice = gr.State(value=initial_api)
+            with gr.Column(scale=2):
+                variables_dataframe = gr.Dataframe(
+                    headers=["变量名", "变量值"],
+                    datatype=["str", "str"],
+                    interactive=True,
+                    label="变量列表",
+                    value=initial_dataframe_value  # 设置初始化数据
+                )
+        def on_doc_change(selected_doc):
+            return selected_doc
+        
+        def on_api_change(selected_api):
+            return selected_api
+            
+        def on_prompt_change(selected_prompt):
+            if not selected_prompt:
+                return None, []
+            selected_prompt_id = int(selected_prompt.split(" ")[0])
+            prompt_data = get_prompt_store().get(doc_id=selected_prompt_id)
+            prompt_content = prompt_data["content"]
+            prompt_template = PromptTemplate.from_template(prompt_content)
+            input_variables = prompt_template.input_variables
+            input_variables.remove("document_slice")
+            dataframe_value = [] if input_variables is None else input_variables
+            dataframe_value = [[var, ""] for var in input_variables]
+            return selected_prompt, dataframe_value
 
-        # 绑定事件（保留原有逻辑，确保交互时更新）
-        doc_dropdown.change(lambda x: x, inputs=doc_dropdown, outputs=doc_state)
-        prompt_dropdown.change(lambda x: x, inputs=prompt_dropdown, outputs=prompt_state)
+        def on_generate_click(doc_state, prompt_state, api_state, variables_dataframe, rounds, concurrency, dataset_name, progress=gr.Progress()):
+            dataset_db = get_datasets()
+            if not dataset_db.search(Query().name == dataset_name):
+                raise gr.Error("数据集名称已存在")
+            
+            doc = [i for i in get_docs() if i.name == doc_state][0]
+            doc_files = doc.markdown_files
+            document_slice_list = [process_markdown_file(doc) for doc in doc_files]
+            prompt = [i for i in get_prompt_store().all() if i["id"] == int(prompt_state.split(" ")[0])][0]
+            prompt = PromptTemplate.from_template(prompt["content"])
+            with Session(get_sql_engine()) as session:
+                api_provider = session.exec(select(APIProvider).where(APIProvider.id == int(api_state.split(" ")[0]))).first()
 
-    return demo
+            variables_dict = {}
+            for _, row in variables_dataframe.iterrows():
+                var_name = row['变量名'].strip()
+                var_value = row['变量值'].strip()
+                if var_name:
+                    variables_dict[var_name] = var_value
+
+            prompt = prompt.partial(**variables_dict)
+            
+            dataset = Dataset(
+                name=dataset_name,
+                model_id=[api_provider.model_id],
+                source_doc=doc,
+                dataset_items=[]
+            )
+            
+            total_slices = len(document_slice_list)
+            for i, document_slice in enumerate(document_slice_list):
+                progress((i + 1) / total_slices, desc=f"处理文档片段 {i + 1}/{total_slices}")
+                request = LLMRequest(api_provider=api_provider,
+                                    prompt=prompt.format(document_slice=document_slice),
+                                    format=generate_json_example(DatasetItem))
+                call_openai_api(request, rounds)
+
+                for resp in request.response:
+                    try:
+                        content = json.loads(resp.content)
+                        dataset_item = DatasetItem(
+                            message=[Q_A(
+                                question=content.get("question", ""),
+                                answer=content.get("answer", "")
+                            )]
+                        )
+                        dataset.dataset_items.append(dataset_item)
+                    except json.JSONDecodeError as e:
+                        print(f"Failed to parse response: {e}")
+
+            # 保存数据集到TinyDB
+            dataset_db.insert(dataset.model_dump())
+
+            save_dataset(dataset_db,get_workdir(),dataset_name)
+            
+            return f"数据集 {dataset_name} 生成完成，共 {len(dataset.dataset_items)} 条数据"
+
+        doc_dropdown.change(on_doc_change, inputs=doc_dropdown, outputs=doc_choice)
+        prompt_dropdown.change(on_prompt_change, inputs=prompt_dropdown, outputs=[prompt_choice, variables_dataframe])
+        api_dropdown.change(on_api_change, inputs=api_dropdown, outputs=api_choice)
+
+        generate_button.click(
+            on_generate_click,
+            inputs=[doc_choice, prompt_choice, api_choice, variables_dataframe, rounds_input, concurrency_input, dataset_name_input],
+            outputs=output_text
+        )
+
+    return demo
+
+if __name__ == "__main__":
+    from global_var import init_global_var
+    init_global_var("workdir")
+    demo = dataset_generate_page()
+    demo.launch()

frontend/model_manage_page.py

@@ -7,7 +7,7 @@ import torch
 
 sys.path.append(str(Path(__file__).resolve().parent.parent))
 from global_var import get_model, get_tokenizer, set_model, set_tokenizer
-from tools.model import get_model_name
+from train import get_model_name
 
 def model_manage_page():
     workdir = "workdir"  # 假设workdir是当前工作目录下的一个文件夹
@@ -30,6 +30,11 @@ def model_manage_page():
         with gr.Row():
             with gr.Column(scale=3):
                 save_model_name_input = gr.Textbox(label="保存模型名称", placeholder="输入模型保存名称")
+                save_method_dropdown = gr.Dropdown(
+                    choices=["default", "merged_16bit", "merged_4bit", "lora", "gguf", "gguf_q4_k_m", "gguf_f16"],
+                    label="保存模式",
+                    value="default"
+                )
             with gr.Column(scale=1):
                 save_button = gr.Button("保存模型", variant="secondary")
 
@@ -73,21 +78,12 @@ def model_manage_page():
         
         unload_button.click(fn=unload_model, inputs=None, outputs=state_output)
 
-        def save_model(save_model_name):
-            try:
-                global model, tokenizer
-                if model is None:
-                    return "没有加载的模型可保存"
-                
-                save_path = os.path.join(models_dir, save_model_name)
-                os.makedirs(save_path, exist_ok=True)
-                model.save_pretrained(save_path)
-                tokenizer.save_pretrained(save_path)
-                return f"模型已保存到 {save_path}"
-            except Exception as e:
-                return f"保存模型时出错: {str(e)}"
+        from train.save_model import save_model_to_dir
+
+        def save_model(save_model_name, save_method):
+            return save_model_to_dir(save_model_name, models_dir, get_model(), get_tokenizer(), save_method)
         
-        save_button.click(fn=save_model, inputs=save_model_name_input, outputs=state_output)
+        save_button.click(fn=save_model, inputs=[save_model_name_input, save_method_dropdown], outputs=state_output)
 
         def refresh_model_list():
             try:

frontend/prompt_manage_page.py

@@ -61,9 +61,11 @@ def prompt_manage_page():
         
     selected_row = None  # 保存当前选中行的全局变量
     
-    def select_record(evt: gr.SelectData):
+    def select_record(dataFrame ,evt: gr.SelectData):
         global selected_row
-        selected_row = evt.row_value
+        selected_row = dataFrame.iloc[evt.index[0]].tolist()
+        selected_row[0] = int(selected_row[0])
+        print(selected_row)
 
     with gr.Blocks() as demo:
         gr.Markdown("## 提示词模板管理")
@@ -102,7 +104,10 @@ def prompt_manage_page():
             outputs=[prompt_table, name_input, description_input, content_input]
         )
 
-        prompt_table.select(select_record, [], [], show_progress="hidden")
+        prompt_table.select(fn=select_record, 
+                              inputs=[prompt_table], 
+                              outputs=[], 
+                              show_progress="hidden")
 
         edit_button.click(
             fn=edit_prompt,

frontend/setting_page.py

@@ -68,9 +68,11 @@ def setting_page():
         
     selected_row = None  # 保存当前选中行的全局变量
     
-    def select_record(evt: gr.SelectData):
+    def select_record(dataFrame ,evt: gr.SelectData):
         global selected_row
-        selected_row = evt.row_value
+        selected_row = dataFrame.iloc[evt.index[0]].tolist()
+        selected_row[0] = int(selected_row[0])
+        print(selected_row)
 
     with gr.Blocks() as demo:
         gr.Markdown("## API Provider 管理")
@@ -109,7 +111,10 @@ def setting_page():
             outputs=[provider_table, model_id_input, base_url_input, api_key_input]  # 添加清空输入框的输出
         )
 
-        provider_table.select(select_record, [], [], show_progress="hidden")
+        provider_table.select(fn=select_record, 
+                              inputs=[provider_table], 
+                              outputs=[], 
+                              show_progress="hidden")
 
         edit_button.click(
             fn=edit_provider,

frontend/train_page.py

@@ -8,7 +8,8 @@ from transformers import TrainerCallback
 
 sys.path.append(str(Path(__file__).resolve().parent.parent))
 from global_var import get_model, get_tokenizer, get_datasets, get_workdir
-from tools import train_model, find_available_port
+from tools import find_available_port
+from train import train_model
 
 def train_page():
     with gr.Blocks() as demo:
@@ -16,17 +17,15 @@ def train_page():
         # 获取数据集列表并设置初始值
         datasets_list = [str(ds["name"]) for ds in get_datasets().all()]
         initial_dataset = datasets_list[0] if datasets_list else None
-
-        dataset_dropdown = gr.Dropdown(
-            choices=datasets_list,
-            value=initial_dataset,  # 设置初始选中项
-            label="选择数据集",
-            allow_custom_value=True,
-            interactive=True
-        )
-        
         with gr.Row():
             with gr.Column(scale=1):
+                dataset_dropdown = gr.Dropdown(
+                    choices=datasets_list,
+                    value=initial_dataset,  # 设置初始选中项
+                    label="选择数据集",
+                    allow_custom_value=True,
+                    interactive=True
+                )
                 # 新增超参数输入组件
                 learning_rate_input = gr.Number(value=2e-4, label="学习率")
                 per_device_train_batch_size_input = gr.Number(value=1, label="batch size", precision=0)
@@ -37,7 +36,7 @@ def train_page():
                 train_button = gr.Button("开始微调")
                 
                 # 训练状态输出
-                output = gr.Textbox(label="训练日志", interactive=False)
+                output = gr.Textbox(label="训练状态", interactive=False)
             with gr.Column(scale=3):
                 # 新增 TensorBoard iframe 显示框
                 tensorboard_iframe = gr.HTML(label="TensorBoard 可视化")
@@ -75,7 +74,7 @@ def train_page():
 
             # 动态生成 TensorBoard iframe
             tensorboard_url = f"http://localhost:{tensorboard_port}"
-            tensorboard_iframe_value = f'<iframe src="{tensorboard_url}" width="100%" height="500px"></iframe>'
+            tensorboard_iframe_value = f'<iframe src="{tensorboard_url}" width="100%" height="1000px"></iframe>'
             yield "训练开始...", tensorboard_iframe_value  # 返回两个值，分别对应 textbox 和 html
 
             try:
@@ -83,6 +82,8 @@ def train_page():
                             dataset, new_training_dir, 
                             learning_rate, per_device_train_batch_size, epoch,
                             save_steps, lora_rank)
+            except Exception as e:
+                raise gr.Error(str(e))
             finally:
                 # 确保训练结束后终止 TensorBoard 子进程
                 tensorboard_process.terminate()

global_var.py

@@ -1,18 +1,16 @@
-from db import get_sqlite_engine, get_prompt_tinydb, get_all_dataset
+from db import load_sqlite_engine, get_prompt_tinydb, get_all_dataset
 from tools import scan_docs_directory
 
 _prompt_store = None
 _sql_engine = None
-_docs = None
 _datasets = None
 _model = None
 _tokenizer = None
 _workdir = None
 def init_global_var(workdir="workdir"):
-    global _prompt_store, _sql_engine, _docs, _datasets, _workdir
+    global _prompt_store, _sql_engine, _datasets, _workdir
     _prompt_store = get_prompt_tinydb(workdir)
-    _sql_engine = get_sqlite_engine(workdir)
-    _docs = scan_docs_directory(workdir)
+    _sql_engine = load_sqlite_engine(workdir)
     _datasets = get_all_dataset(workdir)
     _workdir = workdir
 
@@ -34,19 +32,11 @@ def set_sql_engine(new_sql_engine):
     _sql_engine = new_sql_engine
 
 def get_docs():
-    return _docs
-
-def set_docs(new_docs):
-    global _docs
-    _docs = new_docs
-
+    global _workdir
+    return scan_docs_directory(_workdir)
 def get_datasets():
     return _datasets
 
-def set_datasets(new_datasets):
-    global _datasets
-    _datasets = new_datasets
-
 def get_model():
     return _model
 

main.py

@@ -1,5 +1,5 @@
 import gradio as gr
-import unsloth
+import train
 from frontend import *
 from db import initialize_sqlite_db, initialize_prompt_store
 from global_var import init_global_var, get_sql_engine, get_prompt_store
@@ -26,4 +26,4 @@ if __name__ == "__main__":
             with gr.TabItem("设置"):
                 setting_page()
     
-    app.launch()
+    app.launch(server_name="0.0.0.0")

requirements.txt

@@ -1,7 +1,7 @@
 openai>=1.0.0
 python-dotenv>=1.0.0
 pydantic>=2.0.0
-gradio>=5.0.0
+gradio>=5.25.0
 langchain>=0.3
 tinydb>=4.0.0
 unsloth>=2025.3.19

schema/__init__.py

@@ -1,4 +1,4 @@
 from .dataset import *
-from .dataset_generation import APIProvider, LLMResponse, LLMRequest
+from .dataset_generation import *
 from .md_doc import MarkdownNode
 from .prompt import promptTempleta

schema/dataset.py

@@ -2,7 +2,7 @@ from typing import Optional
 from pydantic import BaseModel, Field
 from datetime import datetime, timezone
 
-class doc(BaseModel):
+class Doc(BaseModel):
     id: Optional[int] = Field(default=None, description="文档ID")
     name: str = Field(default="", description="文档名称")
     path: str = Field(default="", description="文档路径")
@@ -13,18 +13,18 @@ class Q_A(BaseModel):
     question: str = Field(default="", min_length=1,description="问题")
     answer: str = Field(default="", min_length=1, description="答案")
 
-class dataset_item(BaseModel):
+class DatasetItem(BaseModel):
     id: Optional[int] = Field(default=None, description="数据集项ID")
     message: list[Q_A] = Field(description="数据集项内容")
 
-class dataset(BaseModel):
+class Dataset(BaseModel):
     id: Optional[int] = Field(default=None, description="数据集ID")
     name: str = Field(default="", description="数据集名称")
     model_id: Optional[list[str]] = Field(default=None, description="数据集使用的模型ID")
-    source_doc: Optional[doc] = Field(default=None, description="数据集来源文档")
+    source_doc: Optional[Doc] = Field(default=None, description="数据集来源文档")
     description: Optional[str] = Field(default="", description="数据集描述")
     created_at: datetime = Field(
         default_factory=lambda: datetime.now(timezone.utc),
         description="记录创建时间"
     )
-    dataset_items: list[dataset_item] = Field(default_factory=list, description="数据集项列表")
+    dataset_items: list[DatasetItem] = Field(default_factory=list, description="数据集项列表")

schema/dataset_generation.py

@@ -12,40 +12,36 @@ class APIProvider(SQLModel, table=True):
         description="记录创建时间"
     )
 
+class LLMParameters(SQLModel):
+    temperature: Optional[float] = None
+    max_tokens: Optional[int] = None
+    top_p: Optional[float] = None
+    frequency_penalty: Optional[float] = None
+    presence_penalty: Optional[float] = None
+    seed: Optional[int] = None
+
+class TokensUsage(SQLModel):
+    prompt_tokens: int = Field(default=0, description="提示词使用的token数量")
+    completion_tokens: int = Field(default=0, description="完成部分使用的token数量")
+    prompt_cache_hit_tokens: Optional[int] = Field(default=None, description="缓存命中token数量")
+    prompt_cache_miss_tokens: Optional[int] = Field(default=None, description="缓存未命中token数量")
+
 class LLMResponse(SQLModel):
     timestamp: datetime  = Field(
         default_factory=lambda: datetime.now(timezone.utc),
         description="响应的时间戳"
     )
     response_id: str = Field(..., description="响应的唯一ID")
-    tokens_usage: dict = Field(default_factory=lambda: {
-        "prompt_tokens": 0,
-        "completion_tokens": 0,
-        "prompt_cache_hit_tokens": None,
-        "prompt_cache_miss_tokens": None
-    }, description="token使用信息")
-    response_content: dict = Field(default_factory=dict, description="API响应的内容")
+    tokens_usage: TokensUsage = Field(default_factory=TokensUsage, description="token使用信息")
+    content: str = Field(default_factory=dict, description="API响应的内容")
     total_duration: float = Field(default=0.0, description="请求的总时长，单位为秒")
-    llm_parameters: dict = Field(default_factory=lambda: {
-        "temperature": None,
-        "max_tokens": None,
-        "top_p": None,
-        "frequency_penalty": None,
-        "presence_penalty": None,
-        "seed": None
-    }, description="API的生成参数")
+    llm_parameters: Optional[LLMParameters] = Field(default=None, description="LLM参数")
 
 class LLMRequest(SQLModel):
     prompt: str = Field(..., description="发送给API的提示词") 
-    provider_id: int = Field(foreign_key="apiprovider.id")
-    provider: APIProvider = Relationship()
+    api_provider: APIProvider = Field(..., description="API提供者的信息")
     format: Optional[str] = Field(default=None, description="API响应的格式")
     response: list[LLMResponse] = Field(default_factory=list, description="API响应列表")
     error: Optional[list[str]] = Field(default=None, description="API请求过程中发生的错误信息")
     total_duration: float = Field(default=0.0, description="请求的总时长，单位为秒")
-    total_tokens_usage: dict = Field(default_factory=lambda: {
-        "prompt_tokens": 0,
-        "completion_tokens": 0,
-        "prompt_cache_hit_tokens": None,
-        "prompt_cache_miss_tokens": None
-    }, description="token使用信息")
+    total_tokens_usage: TokensUsage = Field(default_factory=TokensUsage, description="token使用信息")

schema/prompt.py

@@ -1,6 +1,7 @@
-from pydantic import BaseModel, Field
+from pydantic import BaseModel, Field, field_validator
 from typing import Optional
 from datetime import datetime, timezone
+from langchain.prompts import PromptTemplate
 
 class promptTempleta(BaseModel):
     id: Optional[int] = Field(default=None, description="模板ID")
@@ -10,4 +11,10 @@ class promptTempleta(BaseModel):
     created_at: str = Field(
         default_factory=lambda: datetime.now(timezone.utc).isoformat(),
         description="记录创建时间"
-    )
+    )
+
+    @field_validator('content')
+    def validate_content(cls, value):
+        if not "document_slice" in PromptTemplate.from_template(value).input_variables:
+            raise ValueError("模板变量缺少 document_slice")
+        return value

tools/__init__.py

@@ -1,5 +1,5 @@
-from .parse_markdown import parse_markdown
-from .scan_doc_dir import *
-from .json_example import generate_example_json
-from .model import *
-from .socket import *
+from .parse_markdown import *
+from .document import *
+from .json_example import generate_json_example
+from .port import *
+from .reasoning import call_openai_api

tools/convert_json_to_dataset.py

@@ -1,19 +1,19 @@
 from typing import List
-from schema.dataset import dataset, dataset_item, Q_A
+from schema.dataset import Dataset, DatasetItem, Q_A
 import json
 
-def convert_json_to_dataset(json_data: List[dict]) -> dataset:
+def convert_json_to_dataset(json_data: List[dict]) -> Dataset:
     # 将JSON数据转换为dataset格式
     dataset_items = []
     item_id = 1  # 自增ID计数器
     for item in json_data:
         qa = Q_A(question=item["question"], answer=item["answer"])
-        dataset_item_obj = dataset_item(id=item_id, message=[qa])
+        dataset_item_obj = DatasetItem(id=item_id, message=[qa])
         dataset_items.append(dataset_item_obj)
         item_id += 1  # ID自增
     
     # 创建dataset对象
-    result_dataset = dataset(
+    result_dataset = Dataset(
         name="Converted Dataset",
         model_id=None,
         description="Dataset converted from JSON",

tools/document.py

@@ -4,7 +4,7 @@ from pathlib import Path
 
 # 添加项目根目录到sys.path
 sys.path.append(str(Path(__file__).resolve().parent.parent))
-from schema import doc
+from schema import Doc
 
 def scan_docs_directory(workdir: str):
     docs_dir = os.path.join(workdir, "docs")
@@ -21,7 +21,7 @@ def scan_docs_directory(workdir: str):
                 for file in files:
                     if file.endswith(".md"):
                         markdown_files.append(os.path.join(root, file))
-            to_return.append(doc(name=doc_name, path=doc_path, markdown_files=markdown_files))
+            to_return.append(Doc(name=doc_name, path=doc_path, markdown_files=markdown_files))
                     
     return to_return
 

tools/json_example.py

@@ -1,32 +1,29 @@
 from pydantic import BaseModel, create_model
-from typing import Any, Dict, List, Optional, get_args, get_origin
+from typing import Any, Dict, List, Optional, Union, get_args, get_origin
 import json
 from datetime import datetime, date
 
-def generate_example_json(model: type[BaseModel]) -> str:
+def generate_json_example(model: type[BaseModel], include_optional: bool = False,list_length = 2) -> str:
     """
     根据 Pydantic V2 模型生成示例 JSON 数据结构。
     """
-
     def _generate_example(field_type: Any) -> Any:
         origin = get_origin(field_type)
         args = get_args(field_type)
 
         if origin is list or origin is List:
-            if args:
-                return [_generate_example(args[0])]
-            else:
-                return []
+            # 生成多个元素，这里生成 3 个
+            result = [_generate_example(args[0]) for _ in range(list_length)] if args else []
+            result.append("......")
+            return result
         elif origin is dict or origin is Dict:
-            if len(args) == 2 and args[0] is str:
+            if len(args) == 2:
                 return {"key": _generate_example(args[1])}
-            else:
-                return {}
-        elif origin is Optional or origin is type(None):
-            if args:
-                return _generate_example(args[0])
-            else:
-                return None
+            return {}
+        elif origin is Union:
+            # 处理 Optional 类型（Union[T, None]）
+            non_none_args = [arg for arg in args if arg is not type(None)]
+            return _generate_example(non_none_args[0]) if non_none_args else None
         elif field_type is str:
             return "string"
         elif field_type is int:
@@ -39,14 +36,23 @@ def generate_example_json(model: type[BaseModel]) -> str:
             return datetime.now().isoformat()
         elif field_type is date:
             return date.today().isoformat()
-        elif issubclass(field_type, BaseModel):
-            return generate_example_json(field_type)
+        elif isinstance(field_type, type) and issubclass(field_type, BaseModel):
+            return json.loads(generate_json_example(field_type, include_optional))
         else:
-            return "unknown"  # 对于未知类型返回 "unknown"
+            # 处理直接类型注解（非泛型）
+            if field_type is type(None):
+                return None
+            try:
+                if issubclass(field_type, BaseModel):
+                    return json.loads(generate_json_example(field_type, include_optional))
+            except TypeError:
+                pass
+            return "unknown"
 
     example_data = {}
     for field_name, field in model.model_fields.items():
-        example_data[field_name] = _generate_example(field.annotation)
+        if include_optional or not isinstance(field.default, type(None)):
+            example_data[field_name] = _generate_example(field.annotation)
 
     return json.dumps(example_data, indent=2, default=str)
 
@@ -55,9 +61,7 @@ if __name__ == "__main__":
     from pathlib import Path
     # 添加项目根目录到sys.path
     sys.path.append(str(Path(__file__).resolve().parent.parent))
-    from schema import Q_A
-    class Q_A_list(BaseModel):
-        Q_As: List[Q_A]
+    from schema import Dataset
 
     print("示例 JSON:")
-    print(generate_example_json(Q_A_list))
+    print(generate_json_example(Dataset))

tools/port.py

@@ -7,4 +7,9 @@ def find_available_port(start_port):
         with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
             if s.connect_ex(('localhost', port)) != 0:  # 端口未被占用
                 return port
-            port += 1  # 如果端口被占用，尝试下一个端口
+            port += 1  # 如果端口被占用，尝试下一个端口
+
+if __name__ == "__main__":
+    start_port = 6006  # 起始端口号
+    available_port = find_available_port(start_port)
+    print(f"Available port: {available_port}")

tools/reasoning.py

@@ -0,0 +1,123 @@
+import sys
+import asyncio
+import openai
+from pathlib import Path
+from datetime import datetime, timezone
+from typing import Optional
+sys.path.append(str(Path(__file__).resolve().parent.parent))
+from schema import APIProvider, LLMRequest, LLMResponse, TokensUsage, LLMParameters
+
+async def call_openai_api(llm_request: LLMRequest, rounds: int = 1, llm_parameters: Optional[LLMParameters] = None) -> LLMRequest:
+    start_time = datetime.now(timezone.utc)
+    client = openai.AsyncOpenAI(
+        api_key=llm_request.api_provider.api_key,
+        base_url=llm_request.api_provider.base_url
+    )
+    
+    total_duration = 0.0
+    total_tokens = TokensUsage()
+    prompt = llm_request.prompt
+    round_start = datetime.now(timezone.utc)
+    if llm_request.format:
+        prompt += "\n请以JSON格式返回结果" + llm_request.format
+
+    for i in range(rounds):
+        round_start = datetime.now(timezone.utc)
+        try:
+            messages = [{"role": "user", "content": prompt}]
+            create_args = {
+                "model": llm_request.api_provider.model_id,
+                "messages": messages,
+                "temperature": llm_parameters.temperature if llm_parameters else None,
+                "max_tokens": llm_parameters.max_tokens if llm_parameters else None,
+                "top_p": llm_parameters.top_p if llm_parameters else None,
+                "frequency_penalty": llm_parameters.frequency_penalty if llm_parameters else None,
+                "presence_penalty": llm_parameters.presence_penalty if llm_parameters else None,
+                "seed": llm_parameters.seed if llm_parameters else None
+            }    # 处理format参数
+
+            if llm_request.format:
+                create_args["response_format"] = {"type": "json_object"}
+
+            response = await client.chat.completions.create(**create_args)
+            
+            round_end = datetime.now(timezone.utc)
+            duration = (round_end - round_start).total_seconds()
+            total_duration += duration
+            
+            # 处理可能不存在的缓存token字段
+            usage = response.usage
+            cache_hit = getattr(usage, 'prompt_cache_hit_tokens', None)
+            cache_miss = getattr(usage, 'prompt_cache_miss_tokens', None)
+            
+            tokens_usage = TokensUsage(
+                prompt_tokens=usage.prompt_tokens,
+                completion_tokens=usage.completion_tokens,
+                prompt_cache_hit_tokens=cache_hit,
+                prompt_cache_miss_tokens=cache_miss if cache_miss is not None else usage.prompt_tokens
+            )
+            
+            # 累加总token使用量
+            total_tokens.prompt_tokens += tokens_usage.prompt_tokens
+            total_tokens.completion_tokens += tokens_usage.completion_tokens
+            if tokens_usage.prompt_cache_hit_tokens:
+                total_tokens.prompt_cache_hit_tokens = (total_tokens.prompt_cache_hit_tokens or 0) + tokens_usage.prompt_cache_hit_tokens
+            if tokens_usage.prompt_cache_miss_tokens:
+                total_tokens.prompt_cache_miss_tokens = (total_tokens.prompt_cache_miss_tokens or 0) + tokens_usage.prompt_cache_miss_tokens
+            
+            llm_request.response.append(LLMResponse(
+                response_id=response.id,
+                tokens_usage=tokens_usage,
+                content = response.choices[0].message.content,
+                total_duration=duration,
+                llm_parameters=llm_parameters
+            ))
+        except Exception as e:
+            round_end = datetime.now(timezone.utc)
+            duration = (round_end - round_start).total_seconds()
+            total_duration += duration
+            
+            llm_request.response.append(LLMResponse(
+                response_id=f"error-round-{i+1}",
+                content={"error": str(e)},
+                total_duration=duration
+            ))
+            if llm_request.error is None:
+                llm_request.error = []
+            llm_request.error.append(str(e))
+    
+    # 更新总耗时和总token使用量
+    llm_request.total_duration = total_duration
+    llm_request.total_tokens_usage = total_tokens
+    
+    return llm_request
+
+if __name__ == "__main__":
+    from json_example import generate_json_example
+    from sqlmodel import Session, select
+    from global_var import get_sql_engine, init_global_var
+    from schema import DatasetItem
+
+    init_global_var("workdir")
+    api_state = "1 deepseek-chat"
+    with Session(get_sql_engine()) as session:
+        api_provider = session.exec(select(APIProvider).where(APIProvider.id == int(api_state.split(" ")[0]))).first()
+    llm_request = LLMRequest(
+        prompt="测试，随便说点什么",
+        api_provider=api_provider,
+        format=generate_json_example(DatasetItem)
+    )
+
+    # # 单次调用示例
+    # result = asyncio.run(call_openai_api(llm_request))
+    # print(f"\n单次调用结果 - 响应数量: {len(result.response)}")
+    # for i, resp in enumerate(result.response, 1):
+    #     print(f"响应{i}: {resp.response_content}")
+
+    # 多次调用示例
+    params = LLMParameters(temperature=0.7, max_tokens=100)
+    result = asyncio.run(call_openai_api(llm_request, 3,params))
+    print(f"\n3次调用结果 - 总耗时: {result.total_duration:.2f}s")
+    print(f"总token使用: prompt={result.total_tokens_usage.prompt_tokens}, completion={result.total_tokens_usage.completion_tokens}")
+    for i, resp in enumerate(result.response, 1):
+        print(f"响应{i}: {resp.content}")

train/__init__.py

@@ -0,0 +1 @@
+from .model import *

train/save_model.py

@@ -0,0 +1,48 @@
+import os
+from global_var import get_model, get_tokenizer
+
+def save_model_to_dir(save_model_name, models_dir, model, tokenizer, save_method="default"):
+    """
+    保存模型到指定目录
+    :param save_model_name: 要保存的模型名称
+    :param models_dir: 模型保存的基础目录
+    :param model: 要保存的模型
+    :param tokenizer: 要保存的tokenizer
+    :param save_method: 保存模式选项
+        - "default": 默认保存方式
+        - "merged_16bit": 合并为16位
+        - "merged_4bit": 合并为4位
+        - "lora": 仅LoRA适配器
+        - "gguf": 保存为GGUF格式
+        - "gguf_q4_k_m": 保存为q4_k_m GGUF格式
+        - "gguf_f16": 保存为16位GGUF格式
+    :return: 保存结果消息或错误信息
+    """
+    try:
+        if model is None:
+            return "没有加载的模型可保存"
+            
+        save_path = os.path.join(models_dir, save_model_name)
+        os.makedirs(save_path, exist_ok=True)
+        
+        if save_method == "default":
+            model.save_pretrained(save_path)
+            tokenizer.save_pretrained(save_path)
+        elif save_method == "merged_16bit":
+            model.save_pretrained_merged(save_path, tokenizer, save_method="merged_16bit")
+        elif save_method == "merged_4bit":
+            model.save_pretrained_merged(save_path, tokenizer, save_method="merged_4bit_forced")
+        elif save_method == "lora":
+            model.save_pretrained_merged(save_path, tokenizer, save_method="lora")
+        elif save_method == "gguf":
+            model.save_pretrained_gguf(save_path, tokenizer)
+        elif save_method == "gguf_q4_k_m":
+            model.save_pretrained_gguf(save_path, tokenizer, quantization_method="q4_k_m")
+        elif save_method == "gguf_f16":
+            model.save_pretrained_gguf(save_path, tokenizer, quantization_method="f16")
+        else:
+            return f"不支持的保存模式: {save_method}"
+            
+        return f"模型已保存到 {save_path} (模式: {save_method})"
+    except Exception as e:
+        return f"保存模型时出错: {str(e)}"