sqlcoder实践

背景

Defog
llama-3

意义

翻译自然语言到sql，类似脑机接口，大模型重要应用领域

sql是数据库查询标准;关系数据库，工具(datax,sqoop，logstash,hive)，非关系数据库（MongoDB，图数据库）等都支持sql查询
BI，数字化运营，商业分析，大数据分析
智能问数
智能问答
没有大模型前智能问答方案：
开源项目 QABasedOnMedicaKnowledgeGraph
https://gitcode.com/liuhuanyong/QASystemOnMedicalKG/overview?utm_source=csdn_github_accelerator&isLogin=1

待完善

可靠性
复杂，不规范的数据库表
信息安全

llama-3-sqlcoder-8b

要求

能翻墙
Nvidia 显卡

模型下载

https://huggingface.co/defog/llama-3-sqlcoder-8b
https://aifasthub.com/models/defog

环境配置

cuda

检查电脑适配cuda版本

D:\working\code> nvidia-smi   
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 528.49       Driver Version: 528.49       CUDA Version: 12.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  NVIDIA GeForce ... WDDM  | 00000000:03:00.0  On |                  N/A |
| N/A   32C    P8     9W /  80W |    616MiB / 12288MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A      1476    C+G   C:\Windows\System32\dwm.exe     N/A      |
|    0   N/A  N/A      2572    C+G   ...wekyb3d8bbwe\Video.UI.exe    N/A      |
|    0   N/A  N/A      2964    C+G   ...d\runtime\WeChatAppEx.exe    N/A      |
|    0   N/A  N/A      4280    C+G   ...2txyewy\TextInputHost.exe    N/A      |
|    0   N/A  N/A      4656    C+G   ...artMenuExperienceHost.exe    N/A      |
|    0   N/A  N/A      7636    C+G   C:\Windows\explorer.exe         N/A      |
|    0   N/A  N/A      7924    C+G   ...icrosoft VS Code\Code.exe    N/A      |
|    0   N/A  N/A      8796    C+G   ...5n1h2txyewy\SearchApp.exe    N/A      |
|    0   N/A  N/A      9376    C+G   ...me\Application\chrome.exe    N/A      |
|    0   N/A  N/A     10540      C   ...rograms\Ollama\ollama.exe    N/A      |
|    0   N/A  N/A     11720    C+G   ...y\ShellExperienceHost.exe    N/A      |
|    0   N/A  N/A     13676    C+G   ...ontend\Docker Desktop.exe    N/A      |
+-----------------------------------------------------------------------------+

得到CUDA版本为12.0

下载

https://developer.nvidia.com/cuda-toolkit-archive

安装后的信息

Installed:
     - Nsight for Visual Studio 2022
     - Nsight Monitor
Not Installed:
     - Nsight for Visual Studio 2019
       Reason: VS2019 was not found
     - Nsight for Visual Studio 2017
       Reason: VS2017 was not found
     - Integrated Graphics Frame Debugger and Profiler
       Reason: see https://developer.nvidia.com/nsight-vstools
     - Integrated CUDA Profilers
       Reason: see https://developer.nvidia.com/nsight-vstools

查看版本

C:\Users\Administrator>nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Mon_Oct_24_19:40:05_Pacific_Daylight_Time_2022
Cuda compilation tools, release 12.0, V12.0.76
Build cuda_12.0.r12.0/compiler.31968024_0

torch

torch是一个Python库，用于构建和训练深度学习和张量计算模型

去torch官网中查看老版本CUDA适配的torch版本：

https://pytorch.org/get-started/locally/

C:\Users\Administrator>pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
Looking in indexes: https://download.pytorch.org/whl/cu121
Requirement already satisfied: torch in c:\python312\lib\site-packages (2.3.0)
Collecting torchvision
  Downloading https://download.pytorch.org/whl/cu121/torchvision-0.18.1%2Bcu121-cp312-cp312-win_amd64.whl (5.7 MB)
     ---------------------------------------- 5.7/5.7 MB 5.9 MB/s eta 0:00:00
Collecting torchaudio
  Downloading https://download.pytorch.org/whl/cu121/torchaudio-2.3.1%2Bcu121-cp312-cp312-win_amd64.whl (4.1 MB)
     ---------------------------------------- 4.1/4.1 MB 7.2 MB/s eta 0:00:00
Requirement already satisfied: filelock in c:\python312\lib\site-packages (from torch) (3.14.0)
Requirement already satisfied: typing-extensions>=4.8.0 in c:\python312\lib\site-packages (from torch) (4.12.1)
Requirement already satisfied: sympy in c:\python312\lib\site-packages (from torch) (1.12.1)
Requirement already satisfied: networkx in c:\python312\lib\site-packages (from torch) (3.3)
Requirement already satisfied: jinja2 in c:\python312\lib\site-packages (from torch) (3.1.4)
Requirement already satisfied: fsspec in c:\python312\lib\site-packages (from torch) (2024.5.0)
Requirement already satisfied: mkl<=2021.4.0,>=2021.1.1 in c:\python312\lib\site-packages (from torch) (2021.4.0)
Requirement already satisfied: numpy in c:\python312\lib\site-packages (from torchvision) (1.26.4)
Collecting torch
  Downloading https://download.pytorch.org/whl/cu121/torch-2.3.1%2Bcu121-cp312-cp312-win_amd64.whl (2423.5 MB)
     ---------------------------------------- 2.4/2.4 GB 501.6 kB/s eta 0:00:00
Collecting pillow!=8.3.*,>=5.3.0 (from torchvision)
  Downloading https://download.pytorch.org/whl/pillow-10.2.0-cp312-cp312-win_amd64.whl (2.6 MB)
     ---------------------------------------- 2.6/2.6 MB 2.5 MB/s eta 0:00:00
Requirement already satisfied: intel-openmp==2021.* in c:\python312\lib\site-packages (from mkl<=2021.4.0,>=2021.1.1->torch) (2021.4.0)
Requirement already satisfied: tbb==2021.* in c:\python312\lib\site-packages (from mkl<=2021.4.0,>=2021.1.1->torch) (2021.12.0)
Requirement already satisfied: MarkupSafe>=2.0 in c:\python312\lib\site-packages (from jinja2->torch) (2.1.5)
Requirement already satisfied: mpmath<1.4.0,>=1.1.0 in c:\python312\lib\site-packages (from sympy->torch) (1.3.0)
Installing collected packages: pillow, torch, torchvision, torchaudio
  Attempting uninstall: torch
    Found existing installation: torch 2.3.0
    Uninstalling torch-2.3.0:
      Successfully uninstalled torch-2.3.0
Successfully installed pillow-10.2.0 torch-2.3.1+cu121 torchaudio-2.3.1+cu121 torchvision-0.18.1+cu121

transformers

pip install transformers

编写脚本

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import sqlparse
print("是否可用：", torch.cuda.is_available())        # 查看GPU是否可用
print("GPU数量：", torch.cuda.device_count())        # 查看GPU数量
print("torch方法查看CUDA版本：", torch.version.cuda)  # torch方法查看CUDA版本
print("GPU索引号：", torch.cuda.current_device())    # 查看GPU索引号
print("GPU名称：", torch.cuda.get_device_name(0))    # 根据索引号得到GPU名称
available_memory = torch.cuda.get_device_properties(0).total_memory
print("GPU内存大小 ：",available_memory)

model_name = "llama-3-sqlcoder-8b"
tokenizer = AutoTokenizer.from_pretrained(model_name)
if available_memory > 20e9:
    # if you have atleast 20GB of GPU memory, run load the model in float16
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        trust_remote_code=True,
        torch_dtype=torch.float16,
        device_map="auto",
        use_cache=True,
    )
else:
    # else, load in 4 bits – this is slower and less accurate
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        trust_remote_code=True,
        # torch_dtype=torch.float16,
        load_in_4bit=True,
        device_map="auto",
        use_cache=True,
    )

prompt = """<|begin_of_text|><|start_header_id|>user<|end_header_id|>

Generate a SQL query to answer this question: `{question}`

DDL statements:

CREATE TABLE products (
  product_id INTEGER PRIMARY KEY, -- Unique ID for each product
  name VARCHAR(50), -- Name of the product
  price DECIMAL(10,2), -- Price of each unit of the product
  quantity INTEGER  -- Current quantity in stock
);

CREATE TABLE customers (
   customer_id INTEGER PRIMARY KEY, -- Unique ID for each customer
   name VARCHAR(50), -- Name of the customer
   address VARCHAR(100) -- Mailing address of the customer
);

CREATE TABLE salespeople (
  salesperson_id INTEGER PRIMARY KEY, -- Unique ID for each salesperson
  name VARCHAR(50), -- Name of the salesperson
  region VARCHAR(50) -- Geographic sales region
);

CREATE TABLE sales (
  sale_id INTEGER PRIMARY KEY, -- Unique ID for each sale
  product_id INTEGER, -- ID of product sold
  customer_id INTEGER,  -- ID of customer who made purchase
  salesperson_id INTEGER, -- ID of salesperson who made the sale
  sale_date DATE, -- Date the sale occurred
  quantity INTEGER -- Quantity of product sold
);

CREATE TABLE product_suppliers (
  supplier_id INTEGER PRIMARY KEY, -- Unique ID for each supplier
  product_id INTEGER, -- Product ID supplied
  supply_price DECIMAL(10,2) -- Unit price charged by supplier
);

-- sales.product_id can be joined with products.product_id
-- sales.customer_id can be joined with customers.customer_id
-- sales.salesperson_id can be joined with salespeople.salesperson_id
-- product_suppliers.product_id can be joined with products.product_id<|eot_id|><|start_header_id|>assistant<|end_header_id|>

The following SQL query best answers the question `{question}`:
```sql
"""


def generate_query(question):
    updated_prompt = prompt.format(question=question)
    inputs = tokenizer(updated_prompt, return_tensors="pt").to("cuda")
    generated_ids = model.generate(
        **inputs,
        num_return_sequences=1,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.eos_token_id,
        max_new_tokens=400,
        do_sample=False,
        num_beams=1,
        temperature=0.0,
        top_p=1,
    )
    outputs = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)

    torch.cuda.empty_cache()
    torch.cuda.synchronize()
    # empty cache so that you do generate more results w/o memory crashing
    # particularly important on Colab – memory management is much more straightforward
    # when running on an inference service
    # return sqlparse.format(outputs[0].split("[SQL]")[-1], reindent=True)
    return outputs[0].split("```sql")[1].split(";")[0]

question = "What was our revenue by product in the new york region last month?"
generated_sql = generate_query(question)
print(sqlparse.format(generated_sql, reindent=True))

运行过程

D:\working\code> & C:/Python312/python.exe d:/working/code/sqlcode_v3.py
是否可用： True
GPU数量： 1
torch方法查看CUDA版本： 12.1
GPU索引号： 0
GPU名称： NVIDIA GeForce RTX 3060 Laptop GPU
GPU内存大小 ： 12884377600
Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
The `load_in_4bit` and `load_in_8bit` arguments are deprecated and will be removed in the future versions. Please, pass a `BitsAndBytesConfig` object in `quantization_config` argument instead.
Loading checkpoint shards: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:26<00:00,  6.59s/it]
C:\Python312\Lib\site-packages\transformers\generation\configuration_utils.py:515: UserWarning: `do_sample` is set to `False`. However, `temperature` is set to `0.0` -- this flag is only used in sample-based generation modes. You should set `do_sample=True` or unset `temperature`.
  warnings.warn(
C:\Python312\Lib\site-packages\bitsandbytes\nn\modules.py:426: UserWarning: Input type into Linear4bit is torch.float16, but bnb_4bit_compute_dtype=torch.float32 (default). This 
will lead to slow inference or training speed.
  warnings.warn(
C:\Python312\Lib\site-packages\transformers\models\llama\modeling_llama.py:649: UserWarning: 1Torch was not compiled with flash attention. (Triggered internally at ..\aten\src\ATen\native\transformers\cuda\sdp_utils.cpp:455.)
  attn_output = torch.nn.functional.scaled_dot_product_attention(

SELECT p.name,
       SUM(s.quantity * p.price) AS total_revenue
FROM products p
JOIN sales s ON p.product_id = s.product_id
JOIN salespeople sp ON s.salesperson_id = sp.salesperson_id
WHERE sp.region = 'New York'
  AND s.sale_date >= CURRENT_DATE - INTERVAL '1 month'
GROUP BY p.name