flake8-no-module-forward-call 0.1.0

Flake8 plugin to prohibit calling .forward() on objects (use obj(inputs) instead so DDP backward hooks run)

flake8-no-model-forward-call

A flake8 plugin that forbids calling .forward() directly on PyTorch nn.Module objects.

Why?

Calling model.forward(inputs) bypasses nn.Module.__call__, which means PyTorch hooks are silently skipped — including the backward hooks that DistributedDataParallel (DDP) relies on for gradient synchronization.

# Bad — DDP backward hooks won't run, gradients won't sync
loss = model.forward(inputs)

# Good — goes through __call__, all hooks fire correctly
loss = model(inputs)

This bug is easy to introduce, hard to notice (training may appear to work), and can silently corrupt distributed training runs.

Installation

pip install flake8-no-model-forward-call

For development / editable install:

git clone https://github.com/mrsndmn/flake8-no-model-forward-call.git
cd flake8-no-model-forward-call
pip install -e .

Usage

Run flake8 with the check enabled:

flake8 --extend-select=NMF001 .

Or add it to your flake8 config so it runs automatically:

.flake8 / setup.cfg:

[flake8]
extend-select = NMF001
max-line-length = 128

pyproject.toml:

[tool.flake8]
extend-select = ["NMF001"]
max-line-length = 128

Error codes

Code	Message
NMF001	Do not call .forward() directly; use model(inputs) instead so DDP backward hooks run

Example

import torch.nn as nn

class MyModel(nn.Module):
    def forward(self, x):
        return x * 2

model = MyModel()

model.forward(x)   # NMF001 — flagged
model(x)           # OK

Running flake8:

example.py:10:1: NMF001 Do not call .forward() directly; use model(inputs) instead so DDP backward hooks run

Contributing

1. Fork the repository
2. Create a feature branch: git checkout -b my-feature
3. Install dev dependencies: pip install -e ".[dev]"
4. Run tests: pytest
5. Open a pull request

License

Apache 2.0 — see LICENSE.