Lightning-AI/pytorch-lightning

Checkpoint do not save under multinode training mode when using Webdataset

Open

#16,893 opened on Feb 28, 2023

View on GitHub
 (1 comment) (0 reactions) (0 assignees)Python (3,233 forks)batch import
bughelp wantedquestionrepro needed

Repository metrics

Stars
 (26,687 stars)
PR merge metrics
 (Avg merge 9d 15h) (3 merged PRs in 30d)

Description

Bug description

I use webdataset with pytorch lightning. In the fact, I get a webdataset dataloader which is a IterableDataset type, and fit it into pytorch lightning trainer. It works fine under single node multi-gpu mode, but when I switch to multi-node mode. The checkpoint do not save. Anyone can help me? Thanks very much!!!!

My modelckpt config is like bellow:

    default_modelckpt_cfg = {
        'metrics_over_trainsteps_checkpoint':{
        "target": "pytorch_lightning.callbacks.ModelCheckpoint",
        "params": {
            "dirpath": ckptdir,
            "filename": "{step:09}",
            "every_n_train_steps": 50000,
            "save_top_k": -1,
        }
        },
    }

And I test in 2 nodes, per node 2 gpus, like this:

    trainer_kwargs["max_epochs"] = 2
    
    trainer_kwargs["accelerator"] = 'gpu'
    trainer_kwargs["devices"] = 2
    trainer_kwargs["strategy"] = "ddp"

Then I fit the webdataset dataloader into trainer:

trainer.fit(model, train_dataloaders=data.data['train'].dataloader)

I am so confuse why just not work when I change to multi node mode, how pytorch lightning decide to save ckpt?

Contributor guide

Research direction
Investigate the ModelCheckpoint callback behavior under DDP multinode with IterableDataset (Webdataset). Check if the callback's save condition (every n train steps) is evaluated on all ranks or only rank 0, and whether rank synchronization affects it. Also verify if the dataloader's iteration is consistent across nodes.
Tech stack
pythonpytorch
Domain
backenddistributed systems
Issue type
Bug
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
2
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
1-3 hours
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Clear
Prerequisites
PythonPyTorchDistributed training basics
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
50