目录

1、简单的模版

2、简单的案例 

  2.1、python 执行.py 文件

  2.2、调式多个文件

  2.3、torchrun、deepspeed 调试

  2.4、accelerate launch (模块)

3、完整的案例


1、简单的模版

定义一个简单的模版如下:

{
    // 使用 IntelliSense 了解相关属性。 
    // 悬停以查看现有属性的描述。
    // 欲了解更多信息,请访问: https://go.microsoft.com/fwlink/?linkid=830387
    "version": "0.2.0",
    "configurations": [
        {
            "name": "Python 调试一",         // 可自定义
            "type": "debugpy",
            "request": "launch",
            "program": "运行脚本的程序",  // 使用.py 脚本路径(相对路径)、which torchrun、which deepspeed等命令查看位置
            "console": "integratedTerminal",
            "justMyCode": false,       // 调试允许进入他人的代码
            "env": {
                "PYTHONPATH": "${workspaceRoot}" // 设置vscode家路径为项目根路径, 搜索包时优先从该目录进行,防止发生import包错误
            },
            "args": [               // 参数,每个参数的参数值无论是否是数字都需用引号
                "--参数1","值1",  
                "--model_name_or_path","facebook/opt-350m",
                "--per_device_train_batch_size", "4",
                "--per_device_eval_batch_size", "4"
            ]
        }
    ]
}

2、简单的案例 

  2.1、python 执行.py 文件

bash 命令

# 加入当前目录的绝对路径
PYTHONPATH=$PWD
export PYTHONPATH
echo "当前bash执行目录: $PWD, 已经将PYTHONPATH设置为: $PYTHONPATH"

batch_dir=data/gpt3_generations_ceshi/

# 命令行python 进行执行
python self_instruct/bootstrap_instructions.py \
    --batch_dir ${batch_dir} \
    --num_instructions_to_generate 5

命令行 python 进行执行脚本,构建launch.json 思路

  • bash 为python执行脚本.py,直接修改"program"为.py脚本相对路径
  • 其他参数照抄
{
    // 使用 IntelliSense 了解相关属性。 
    // 悬停以查看现有属性的描述。
    // 欲了解更多信息,请访问: https://go.microsoft.com/fwlink/?linkid=830387
    "version": "0.2.0",
    "configurations": [
        {
            "name": "Python 调试",
            "type": "debugpy",
            "request": "launch",
            "program": "self_instruct/bootstrap_instructions.py",  // .py脚本文件相对路径位置
            "console": "integratedTerminal",
            "justMyCode": false,
            "env": {
                "PYTHONPATH": "${workspaceRoot}" // 设置vscode项目根路径,搜索包时优先从该目录进行,防止发生import包错误
            },
            "args": [
                "--batch_dir","data/gpt3_generations_ceshi",// TODO 修改官方线上数据集为自己的路径
                "--num_instructions_to_generate","5"    
            ]
        }
    ]
}

  2.2、调式多个文件

与调试单个文件同理,只是重复

{
    // 使用 IntelliSense 了解相关属性。 
    // 悬停以查看现有属性的描述。
    // 欲了解更多信息,请访问: https://go.microsoft.com/fwlink/?linkid=830387
    "version": "0.2.0",
    "configurations": [
        // 第一个文件
        {
            "name": "Python 调试 bootstrap_instructions.py",
            "type": "debugpy",
            "request": "launch",
            "program": "self_instruct/bootstrap_instructions.py",  // .py脚本文件相对路径位置
            "console": "integratedTerminal",
            "justMyCode": false,
            "env": {
                "PYTHONPATH": "${workspaceRoot}" // 设置vscode项目根路径,搜索包时优先从该目录进行,防止发生import包错误
            },
            "args": [
                "--batch_dir","data/gpt3_generations_ceshi",// TODO 修改官方线上数据集为自己的路径
                "--num_instructions_to_generate","5"
            ]
        },
        // 第二个文件
        {
            "name": "Python 调试 identify_clf_or_not.py",
            "type": "debugpy",
            "request": "launch",
            "program": "self_instruct/identify_clf_or_not.py",  // .py脚本文件相对路径位置
            "console": "integratedTerminal",
            "justMyCode": false,
            "env": {
                "PYTHONPATH": "${workspaceRoot}" // 设置vscode项目根路径,搜索包时优先从该目录进行,防止发生import包错误
            },
            "args": [
                "--batch_dir","data/gpt3_generations_ceshi",// TODO 修改官方线上数据集为自己的路径
                "--num_instructions_to_generate","5"
            ]
        }
    ]
}

当vscode需要跳转到项目的子目录执行程序时,这个时候可以使用 “cwd”参数,将项目根目录改变,例如跳转至子目录“test1”中,需要在上述脚本中“env”参数前加入下面的一行命令即可:

"cwd": "${workspaceFolder}/test1",

  2.3、torchrun、deepspeed 调试

bash 命令

# 加入当前目录的绝对路径
PYTHONPATH=$PWD
export PYTHONPATH
echo "当前bash执行目录: $PWD, 已经将PYTHONPATH设置为: $PYTHONPATH"

batch_dir=data/gpt3_generations_ceshi/

# 命令行python 进行执行
deepspeed --num_gpus 1 self_instruct/bootstrap_instructions.py \
    --batch_dir ${batch_dir} \
    --num_instructions_to_generate 5

命令行 deepspeed/torchrun 进行执行脚本,构建launch.json 思路

  • 构建launch.json脚本时需要找到“deepspeed”命令的路径,bash命令行:which deepspeed,直接修改"program"为该路径。
  • self_instruct/bootstrap_instructions.py 是执行的脚本的相对路径,不在主目录中,因此我们需要加入 "PYTHONPATH": "${workspaceRoot}" 指定项目目录到环境变量中,以防代码运行时出现 import 错误
{
    // 使用 IntelliSense 了解相关属性。 
    // 悬停以查看现有属性的描述。
    // 欲了解更多信息,请访问: https://go.microsoft.com/fwlink/?linkid=830387
    "version": "0.2.0",
    "configurations": [
        {
            "name": "Python 调试一阶段LORA",
            "type": "debugpy",
            "request": "launch",
            "program": "/opt/conda/envs/dsc/bin/deepspeed",  // which deepspeed 查看位置
            "console": "integratedTerminal",
            "justMyCode": false,
            "env": {
                "PYTHONPATH": "${workspaceRoot}" // 设置vscode项目根路径,搜索包时优先从该目录进行,防止发生import包错误
            },
            "args": [
                "--num_gpus", "1", 
                "self_instruct/bootstrap_instructions.py",   // 给定脚本地址(相对路径)
                "--batch_dir","data/gpt3_generations_ceshi",  
                "--num_instructions_to_generate","5"
            ]
        }
    ]
}

  2.4、accelerate launch (模块)

# bash
accelerate launch --config_file "examples/sft/configs/deepspeed_config_z3_qlora.yaml"  examples/sft/train.py \
    --seed 100 \
    --model_name_or_path "/workspace/Llama-2-7b-chat-hf" \
    --dataset_name "smangrul/ultrachat-10k-chatml" \
    --chat_template_format "chatml" \
    --add_special_tokens False \
    --append_concat_token False \
    --splits "train,test" \
    2>&1 | tee -a examples/sft/qlora_ds_zero3_log.out

launch.json 

{
    // 使用 IntelliSense 了解相关属性。 
    // 悬停以查看现有属性的描述。
    // 欲了解更多信息,请访问: https://go.microsoft.com/fwlink/?linkid=830387
    "version": "0.2.0",
    "configurations": [
        {
            "name": "Python ds_z3_qlora_multigpu 微调",
            "type": "debugpy",
            "request": "launch",
            "module": "accelerate.commands.launch",          //调试accelerate launch
            "console": "integratedTerminal",
            "justMyCode": false,
            "env": {
                "PYTHONPATH": "${workspaceRoot}"
            },
            "args": [
                "--config_file", "examples/sft/configs/deepspeed_config_z3_qlora.yaml",
                "examples/sft/train.py",
                "--seed", "100",
                "--model_name_or_path", "/workspace/Llama-2-7b-chat-hf",
                "--dataset_name", "smangrul/ultrachat-10k-chatml",
                "--chat_template_format", "chatml",
                "--add_special_tokens", "False",
                "--append_concat_token", "False",
                "--splits", "train,test"
              ]
        }
    ]
}

3、完整的案例

{
    // 使用 IntelliSense 了解相关属性。 
    // 悬停以查看现有属性的描述。
    // 欲了解更多信息,请访问: https://go.microsoft.com/fwlink/?linkid=830387
    "version": "0.2.0",
    "configurations": [
        // py 脚本
        {
            "name": "Python lora 微调",
            "type": "debugpy",
            "request": "launch",
            "program": "finetune_demo/finetune_hf.py",
            "console": "integratedTerminal",
            "justMyCode": false,
            "env": {
                "PYTHONPATH": "${workspaceRoot}"
            },
            "args": [
                "/workspace/AdvertiseGen_fix",  
                "/workspace/chatglm3-6b",
                "finetune_demo/configs/lora.yaml"
            ]
        },
        // torchrun 分布式
        {
            "name": "Python lora_ds 微调",
            "type": "debugpy",
            "request": "launch",
            "program": "/opt/conda/envs/llm/bin/torchrun",
            "console": "integratedTerminal",
            "justMyCode": false,
            "env": {
                "PYTHONPATH": "${workspaceRoot}"
            },
            "args": [
                "--nproc_per_node","1",
                "finetune_demo/finetune_hf.py",
                "/workspace/AdvertiseGen_fix",  
                "/workspace/chatglm3-6b",
                "finetune_demo/configs/lora.yaml"
            ]
        }
    ]
}

GitHub 加速计划 / js / json
19
5
下载
适用于现代 C++ 的 JSON。
最近提交(Master分支:4 个月前 )
f06604fc * :page_facing_up: bump the copyright years Signed-off-by: Niels Lohmann <mail@nlohmann.me> * :page_facing_up: bump the copyright years Signed-off-by: Niels Lohmann <mail@nlohmann.me> * :page_facing_up: bump the copyright years Signed-off-by: Niels Lohmann <niels.lohmann@gmail.com> --------- Signed-off-by: Niels Lohmann <mail@nlohmann.me> Signed-off-by: Niels Lohmann <niels.lohmann@gmail.com> 28 天前
d23291ba * add a ci step for Json_Diagnostic_Positions Signed-off-by: Harinath Nampally <harinath922@gmail.com> * Update ci.cmake to address review comments Signed-off-by: Harinath Nampally <harinath922@gmail.com> * address review comment Signed-off-by: Harinath Nampally <harinath922@gmail.com> * fix typo in the comment Signed-off-by: Harinath Nampally <harinath922@gmail.com> * fix typos in ci.cmake Signed-off-by: Harinath Nampally <harinath922@gmail.com> * invoke the new ci step from ubuntu.yml Signed-off-by: Harinath Nampally <harinath922@gmail.com> * issue4561 - use diagnostic positions for exceptions Signed-off-by: Harinath Nampally <harinath922@gmail.com> * fix ci_test_documentation check Signed-off-by: Harinath Nampally <harinath922@gmail.com> * address review comments Signed-off-by: Harinath Nampally <harinath922@gmail.com> * fix ci check failures for unit-diagnostic-postions.cpp Signed-off-by: Harinath Nampally <harinath922@gmail.com> * improvements based on review comments Signed-off-by: Harinath Nampally <harinath922@gmail.com> * fix const correctness string Signed-off-by: Harinath Nampally <harinath922@gmail.com> * further refinements based on reviews Signed-off-by: Harinath Nampally <harinath922@gmail.com> * add one more test case for full coverage Signed-off-by: Harinath Nampally <harinath922@gmail.com> * ci check fix - add const Signed-off-by: Harinath Nampally <harinath922@gmail.com> * add unit tests for json_diagnostic_postions only Signed-off-by: Harinath Nampally <harinath922@gmail.com> * fix ci_test_diagnostics Signed-off-by: Harinath Nampally <harinath922@gmail.com> * fix ci_test_build_documentation check Signed-off-by: Harinath Nampally <harinath922@gmail.com> --------- Signed-off-by: Harinath Nampally <harinath922@gmail.com> 28 天前
Logo

旨在为数千万中国开发者提供一个无缝且高效的云端环境,以支持学习、使用和贡献开源项目。

更多推荐