Got stuck when debugging a multi-process program in PyTorch

提供された example.py、launch.json、torchrun コマンドから始めて実行し、その後、両方の DDP プロセスにおける失敗する attach 構成と、動作する connect/listen 構成を比較します。ハングを再現し、debugpy 固有の挙動を切り分け、期待される非ブロッキングの結果を確認できれば完了です。

Before creating a new issue, please check the FAQ to see if your question is answered there.

Environment data

• debugpy version: 1.8.1 (run import debugpy; print(debugpy.__version__) if uncertain)
• OS and version: Centos
• Python version (& distribution if applicable, e.g. Anaconda): 3.10.13
• Using VS Code or Visual Studio: VS Code 1.90.2

Actual behavior

When I initiate debugging and press F10, the program gets stuck at the line with DDP (ddp_model = DDP(model, device_ids=[device_id]))

• launch.json

{
    "version": "0.2.0",
    "configurations": [
        {
            "name": "Attach Test",
            "type": "debugpy",
            "request": "attach",
            "listen": {
                "port": 5678,
            },
            "justMyCode": false,
        },
   ],
}

• example.py

import torch
import torch.distributed as dist
import torch.nn as nn
import torch.optim as optim
from torch.nn.parallel import DistributedDataParallel as DDP

class ToyModel(nn.Module):
    def __init__(self):
        super(ToyModel, self).__init__()
        self.net1 = nn.Linear(10, 10)
        self.relu = nn.ReLU()
        self.net2 = nn.Linear(10, 5)

    def forward(self, x):
        return self.net2(self.relu(self.net1(x)))

def demo_basic():
    dist.init_process_group("nccl")
    rank = dist.get_rank()
    print(f"Start running basic DDP example on rank {rank}.")

    # create model and move it to GPU with id rank
    device_id = rank % torch.cuda.device_count()
    model = ToyModel().to(device_id)

    import debugpy
    debugpy.connect(5678)
    debugpy.wait_for_client()
    debugpy.breakpoint()

    ddp_model = DDP(model, device_ids=[device_id])

    loss_fn = nn.MSELoss()
    optimizer = optim.SGD(ddp_model.parameters(), lr=0.001)
    optimizer.zero_grad()
    outputs = ddp_model(torch.randn(20, 10))
    labels = torch.randn(20, 5).to(device_id)
    loss_fn(outputs, labels).backward()
    optimizer.step()
    dist.destroy_process_group()

if __name__ == "__main__":
    demo_basic()

• launch command:

torchrun --nproc_per_node=2 example.py

Note: The configuration above was working before, but it suddenly stopped working recently. Additionally, I can debug normally by configuring 'connect' in launch.json and using 'listen' in the code, but this method requires setting different ports to avoid port conflicts, which is not very convenient.

Expected behavior

It will not get stuck.

Steps to reproduce:

1. XXX