search

Lightning-AI / pytorch-lightning

Pretrain, finetune ANY AI model of ANY size on multiple GPUs, TPUs with zero code changes.
https://lightning.ai
Apache License 2.0
28.2k stars 3.37k forks source link

Shuffle order is the same across runs when using strategy='ddp' #20249

Open bogdanmagometa opened 1 month ago

bogdanmagometa commented 1 month ago

Bug description

The batches and their order are the same across different executions of the script when using strategy='ddp' and dataloader with shuffle=True

What version are you seeing the problem on?

v2.2

How to reproduce the bug

Say you have train.py that prints the current input on each training iteration and has shuffling enabled in the dataloader:

import torch
from torch.utils.data import TensorDataset, DataLoader
import torch.nn.functional as F
import lightning.pytorch as pl

class SomeLightningModule(pl.LightningModule):
    def __init__(self):
        super().__init__()
        self.p1 = torch.nn.Parameter(torch.tensor(0.0))
        self.p2 = torch.nn.Parameter(torch.tensor(0.0))
    def training_step(self, batch):
        x, y = batch
        print(x.item())
        return F.mse_loss(x * self.p1 + self.p2, y)
    def configure_optimizers(self):
        optimizer = torch.optim.Adam(
            self.parameters(),
        )

        return {
            "optimizer": optimizer,
        }

lightning_module = SomeLightningModule()

trainer = pl.Trainer(
    strategy='ddp',
    max_epochs=1,
)

train_dataset = TensorDataset(torch.arange(5).float(), torch.arange(5).float())
train_loader = DataLoader(train_dataset, shuffle=True)

trainer.fit(lightning_module, train_dataloaders=train_loader)

When strategy='ddp', the script will print the same numbers across different runs:

$ python3 train.py
4.0
0.0
1.0
3.0
2.0
$ python3 train.py
4.0
0.0
1.0
3.0
2.0

Such behavior can be unwanted, as people might want to try different orders of batches (e.g. to construct ensembles or get the average performance)

Error messages and logs

# Error messages and logs here please

Environment

Current environment * CUDA: - GPU: - Graphics Device - available: True - version: 11.8 * Lightning: - lightning: 2.2.0.post0 - lightning-utilities: 0.10.1 - pytorch-lightning: 1.7.7 - torch: 2.1.2 - torchaudio: 2.1.2 - torchmetrics: 0.10.3 - torchvision: 0.16.2 * Packages: - absl-py: 1.3.0 - aiohttp: 3.8.3 - aiosignal: 1.3.1 - alphafold-colabfold: 2.3.6 - altair: 5.4.0 - anarci: 1.3 - antiberty: 0.1.3 - antlr4-python3-runtime: 4.9.3 - anyio: 3.5.0 - appdirs: 1.4.4 - argon2-cffi: 21.3.0 - argon2-cffi-bindings: 21.2.0 - asttokens: 2.0.5 - astunparse: 1.6.3 - async-lru: 2.0.4 - async-timeout: 4.0.2 - attrs: 22.1.0 - babel: 2.11.0 - backcall: 0.2.0 - beautifulsoup4: 4.12.2 - biopython: 1.79 - bleach: 4.1.0 - blinker: 1.5 - bottleneck: 1.3.5 - brotlipy: 0.7.0 - cached-property: 1.5.2 - cachetools: 5.2.0 - certifi: 2023.5.7 - cffi: 1.15.1 - charset-normalizer: 2.1.1 - chex: 0.1.86 - click: 8.1.3 - cmake: 3.28.3 - colabfold: 1.5.5 - colorama: 0.4.6 - comm: 0.1.2 - contextlib2: 21.6.0 - contourpy: 1.0.6 - cryptography: 38.0.3 - cycler: 0.11.0 - debugpy: 1.6.7 - decorator: 5.1.1 - deepspeed: 0.9.5 - defusedxml: 0.7.1 - dm-haiku: 0.0.12 - dm-tree: 0.1.8 - docker-pycreds: 0.4.0 - docstring-parser: 0.15 - einops: 0.8.0 - entrypoints: 0.4 - et-xmlfile: 1.1.0 - etils: 1.5.2 - exceptiongroup: 1.0.4 - executing: 0.8.3 - fastjsonschema: 2.16.2 - filelock: 3.13.1 - flatbuffers: 24.3.25 - flax: 0.8.5 - fonttools: 4.38.0 - frozenlist: 1.3.3 - fsspec: 2024.3.1 - gast: 0.6.0 - gdown: 5.1.0 - gemmi: 0.5.7 - gitdb: 4.0.9 - gitpython: 3.1.29 - gmpy2: 2.1.2 - google-auth: 2.14.1 - google-auth-oauthlib: 0.4.6 - google-pasta: 0.2.0 - grpcio: 1.49.1 - h5py: 3.11.0 - hjson: 3.1.0 - huggingface-hub: 0.22.2 - hydra-core: 1.3.2 - idna: 3.4 - immutabledict: 4.2.0 - importlib-metadata: 4.13.0 - importlib-resources: 6.1.2 - ipykernel: 6.25.0 - ipython: 8.15.0 - ipython-genutils: 0.2.0 - ipywidgets: 8.0.4 - jax: 0.3.25 - jaxlib: 0.3.25+cuda11.cudnn82 - jedi: 0.18.1 - jinja2: 3.1.2 - jmp: 0.0.4 - json5: 0.9.6 - jsonargparse: 4.27.5 - jsonschema: 4.17.3 - jupyter: 1.0.0 - jupyter-client: 7.4.9 - jupyter-console: 6.6.3 - jupyter-core: 5.5.0 - jupyter-events: 0.6.3 - jupyter-lsp: 2.2.0 - jupyter-server: 2.10.0 - jupyter-server-terminals: 0.4.4 - jupyterlab: 4.0.8 - jupyterlab-pygments: 0.1.2 - jupyterlab-server: 2.22.0 - jupyterlab-widgets: 3.0.9 - keras: 3.4.1 - kiwisolver: 1.4.4 - libclang: 18.1.1 - lightning: 2.2.0.post0 - lightning-utilities: 0.10.1 - lit: 18.1.1 - markdown: 3.4.1 - markdown-it-py: 3.0.0 - markupsafe: 2.1.1 - matplotlib: 3.6.2 - matplotlib-inline: 0.1.6 - mdurl: 0.1.2 - mistune: 2.0.4 - mkl-fft: 1.3.1 - mkl-random: 1.2.2 - mkl-service: 2.4.0 - ml-collections: 0.1.1 - ml-dtypes: 0.3.2 - mmcif-pdbx: 2.0.1 - mpi4py: 3.1.4 - mpmath: 1.3.0 - msgpack: 1.0.8 - multidict: 6.0.2 - munkres: 1.1.4 - namex: 0.0.8 - narwhals: 1.5.0 - nbclient: 0.8.0 - nbconvert: 7.10.0 - nbformat: 5.9.2 - nest-asyncio: 1.5.6 - networkx: 3.1 - ninja: 1.11.1 - notebook: 6.3.0 - notebook-shim: 0.2.3 - numexpr: 2.8.4 - numpy: 1.23.5 - oauthlib: 3.2.2 - omegaconf: 2.3.0 - openpyxl: 3.1.5 - opt-einsum: 3.3.0 - optax: 0.2.2 - optree: 0.11.0 - orbax-checkpoint: 0.5.20 - overrides: 7.4.0 - packaging: 21.3 - pandas: 1.5.3 - pandocfilters: 1.5.0 - parso: 0.8.3 - path: 16.2.0 - pathtools: 0.1.2 - pdb2pqr: 3.6.1 - pexpect: 4.8.0 - pickleshare: 0.7.5 - pillow: 9.2.0 - pip: 22.3.1 - platformdirs: 3.10.0 - ply: 3.11 - pmw: 2.0.1 - pooch: 1.6.0 - prody: 2.2.0 - prometheus-client: 0.14.1 - promise: 2.3 - prompt-toolkit: 3.0.43 - propka: 3.5.1 - protobuf: 4.21.9 - psutil: 5.9.4 - ptyprocess: 0.7.0 - pure-eval: 0.2.2 - py-cpuinfo: 9.0.0 - py3dmol: 2.0.4 - pyasn1: 0.4.8 - pyasn1-modules: 0.3.0 - pycollada: 0.8 - pycparser: 2.21 - pydantic: 1.10.11 - pydeprecate: 0.3.2 - pygments: 2.15.1 - pyjwt: 2.6.0 - pykerberos: 1.2.4 - pymol: 2.5.5 - pyopenssl: 22.1.0 - pyparsing: 3.0.9 - pyqt5: 5.15.7 - pyqt5-sip: 12.11.0 - pyrsistent: 0.20.0 - pysocks: 1.7.1 - python-dateutil: 2.8.2 - python-json-logger: 2.0.7 - pytorch-lightning: 1.7.7 - pytz: 2022.7 - pyu2f: 0.1.5 - pyyaml: 6.0 - pyzmq: 25.1.0 - qtconsole: 5.5.1 - qtpy: 2.4.1 - regex: 2023.12.25 - requests: 2.28.1 - requests-oauthlib: 1.3.1 - rfc3339-validator: 0.1.4 - rfc3986-validator: 0.1.1 - rich: 13.7.1 - rjieba: 0.1.11 - rsa: 4.9 - safetensors: 0.4.2 - scipy: 1.10.1 - seaborn: 0.13.2 - send2trash: 1.8.2 - sentry-sdk: 1.11.0 - setproctitle: 1.3.2 - setuptools: 59.5.0 - shortuuid: 1.0.11 - sip: 6.7.12 - six: 1.16.0 - smmap: 3.0.5 - sniffio: 1.2.0 - soupsieve: 2.5 - stack-data: 0.2.0 - sympy: 1.12 - tabulate: 0.9.0 - tensorboard: 2.16.2 - tensorboard-data-server: 0.7.2 - tensorboard-plugin-wit: 1.8.1 - tensorflow-cpu: 2.16.2 - tensorflow-io-gcs-filesystem: 0.37.0 - tensorstore: 0.1.63 - termcolor: 2.4.0 - terminado: 0.17.1 - tinycss2: 1.2.1 - tmtools: 0.2.0 - tokenizers: 0.15.2 - toml: 0.10.2 - tomli: 2.0.1 - toolz: 0.12.0 - torch: 2.1.2 - torchaudio: 2.1.2 - torchmetrics: 0.10.3 - torchvision: 0.16.2 - tornado: 6.3.3 - tqdm: 4.64.1 - trainable-folding: 0.0.0 - traitlets: 5.7.1 - transformers: 4.39.3 - triton: 2.1.0 - tunedabs: 0.0.1 - typeshed-client: 2.5.1 - typing-extensions: 4.10.0 - unicodedata2: 15.0.0 - urllib3: 1.26.11 - wandb: 0.13.5 - wcwidth: 0.2.5 - webencodings: 0.5.1 - websocket-client: 0.58.0 - werkzeug: 2.2.2 - wheel: 0.40.0 - widgetsnbextension: 4.0.5 - wrapt: 1.16.0 - yarl: 1.8.1 - zipp: 3.10.0 * System: - OS: Linux - architecture: - 64bit - ELF - processor: x86_64 - python: 3.9.13 - release: 3.10.0-693.17.1.el7.x86_64 - version: #1 SMP Thu Jan 25 20:13:58 UTC 2018

More info

No response

OberonDixon commented 3 weeks ago

I am having a similar problem. I am using lightning 2.4.0 and a LightningDataModule that wraps a custom pytorch Dataset. I have verified that my LightningDataModule, when initialized outside a distributed training run, will in fact shuffle - my DataLoader initialization in LightningDataModule.train_dataloader has shuffle=True - and I have also checked that my torch.initial_seed() is being set differently between runs and torch.backends.cudnn.deterministic and torch.backends.cudnn.benchmark are set to False. However, with Trainer accelerator='auto' and devices='auto', running on a >1 GPU slurm job, I always get the exact same order of batches.