How to make VScode launch.json for a Python module

Question

I'm researching self-supervised machine learning code.

And I have wanted to debug the code with python debugger not pdb.set_trace(). This is python command for ubuntu terminal.

python -m torch.distributed.launch --nproc_per_node=1 main_swav.py \
--data_path /dataset/imagenet/train \
--epochs 400 \
--base_lr 0.6 \
--final_lr 0.0006 \
--warmup_epochs 0 \
--batch_size 8 \
--size_crops 224 96 \
--nmb_crops 2 6 \
--min_scale_crops 0.14 0.05 \
--max_scale_crops 1. 0.14 \
--use_fp16 true \
--freeze_prototypes_niters 5005 \
--queue_length 380 \
--epoch_queue_starts 15\
--workers 10

In order to debug the code with VScode, I tried to revise launch.json like below as referring stackoverflow question

{
    "version": "0.2.0",
    "configurations": [
        {
            "name": "Python: Current File",
            "type": "python",
            "module": "torch.distributed.launch --nproc_per_node=1 main_swav.py",
            "request": "launch",
            "console": "integratedTerminal",
            "args": ["--data_path", "/dataset/imagenet/train"]
        }
    ]
}

I knew this would not work...

Could you give me some advice?

args is used to pass command line arguments along to the app being launched. program is used to specify the python file. View debug python in vscode before you start debugging. — Molly Wang-MSFT
– Molly Wang-MSFT, Commented May 14, 2021 at 7:37
Note that currently (2024): "type" : "python" is deprecated. Use "debugpy" instead: See code.visualstudio.com/docs/python/debugging#_module — Daraan
– Daraan, Commented Mar 21, 2024 at 10:30

Dj Sushi · Accepted Answer · 2024-03-21 10:28:58Z

24

Specify the module you want to run with "module": "torch.distributed.launch"

You can ignore the -m flag. Put everything else under the args key.

Note: Make sure to include --nproc_per_node and the name of file (main_swav.py) in the list of arguments

{
    "version": "0.2.0",
    "configurations": [
        {
            "name": "Python: Current File",
            "type": "debugpy",
            "module": "torch.distributed.launch",
            "request": "launch",
            "console": "integratedTerminal",
            "args": [
                "--nproc_per_node", "1", 
                "main_swav.py",
                "--data_path", "/dataset/imagenet/train",
            ]
        }
    ]
}

Read more here: https://code.visualstudio.com/docs/python/debugging#_module

edited Mar 21, 2024 at 10:28

Dj Sushi

7061 gold badge9 silver badges21 bronze badges

answered Jul 13, 2021 at 20:36

Matt Spataro

3423 silver badges9 bronze badges

Sign up to request clarification or add additional context in comments.

6 Comments

aviator Over a year ago

Any update in this for the newer torchrun?

ringo Over a year ago

@aviator set your module to torch.distributed.run

임정섭 Over a year ago

@ringo That would cast away the benefit from torchrun and accompany efforts setting various distribution related environment variables. Why not just run as program:{torchrun_path}?

ringo Over a year ago

I don't think so. torchrun is just a simple Python script file that runs torch.distributed.run.main, and running torch.distributed.run as a module does the same thing.

Roberto Amoroso Over a year ago

If you have the program entry (e.g., "program": "${workspaceFolder}/main.py") remove it and add It in the args after "--nproc_per_node", "1",

|

azizbro · Accepted Answer · 2023-02-20 04:45:45Z

This is an example of my launch.json that I use to debug Python modules.

It has an additional configuration to debug "current file" (not as module) which is useful to keep.

{
linkid=830387
  "version": "0.2.0",
  "configurations": [
    {
      "name": "Python: Module",
      "type": "python",
      "request": "launch",
      "module": "path.to.module",
      "args": ["run_example --arg <arg>"],
      "justMyCode": true
    },
    {
      "name": "Python: Current File",
      "type": "python",
      "request": "launch",
      "program": "${file}",
      "console": "integratedTerminal",
      "justMyCode": true
    }
  ]
}

This would replicate a terminal command to run a Python module like so:

python -m path.to.module run_example --arg <arg>

cwin · Accepted Answer · 2023-09-30 23:01:42Z

3

For the new torchrun (Torch 2.0.1) here is launch.json:

"module" is no longer used
the full path to torchrun is needed (at least for me)

{
    "name": "working",
    "type": "python",
    "request": "launch",
    "program": "/home/user_x/anaconda3/envs/env_y/bin/torchrun",
    "console": "integratedTerminal",
    "justMyCode": true,
    "args": [
        "example_chat_completion.py",
        "--nproc_per_node", "1",
        "--ckpt_dir", "llama-2-7b-chat/",
        "--tokenizer_path", "tokenizer.model",
        "--max_seq_len", "512",
        "--max_batch_size", "6",
    ]
},

edited Sep 30, 2023 at 23:01

answered Sep 29, 2023 at 19:22

cwin

213 bronze badges

Comments

Zhang Kin · Accepted Answer · 2025-05-07 09:04:08Z

Related issues: https://github.com/microsoft/debugpy/issues/1311

First the .vscode/launch.json setting is:

{
    // kill -9 $(pgrep -f "python3 -m debugpy" | xargs echo)
    // python -m debugpy --listen 5678 --wait-for-client xxx.py
    // python -m debugpy --listen 5678 --wait-for-client -m torch.distributed.run --nproc_per_node 1 --nnodes 1 xxx.py
    "version": "0.2.0",
    "configurations": [
        {
            "name": "LDebug",
            "type": "python",
            "request": "attach",
            "connect": {
            "host": "localhost",
            "port": 5678
            }
        }
    ]
}

For torchrun debug command:

python -m debugpy --listen 5678 --wait-for-client -m torch.distributed.run --nproc_per_node 1 --nnodes 1 xxx.py

For normal python debug command:

python -m debugpy xxx.py

after excute command, press F1/or click start debug to link with the port.

Undefined443 · Accepted Answer · 2025-07-21 13:15:16Z

0

For torch>=2.7, use torch.distributed.run:

{
  "name": "train",
  "type": "debugpy",
  "request": "launch",
  "module": "torch.distributed.run",
  "args": [
    "--nproc-per-node=1",
    "--standalone",
    "train.py"
  ]
}

Reference: torchrun (Elastic Launch) — PyTorch 2.7 documentation

edited Jul 21 at 13:15

answered Mar 30 at 12:23

Undefined443

494 bronze badges

Collectives™ on Stack Overflow

How to make VScode launch.json for a Python module

5 Answers 5

6 Comments

Comments

Comments

Comments

Comments

Your Answer

Linked

Hot Network Questions

Collectives™ on Stack Overflow

5 Answers 5

6 Comments

Comments

Comments

Comments

Comments

Your Answer

Sign up or log in

Post as a guest

Linked

Related