Developer Guide

Everything you need to contribute to TorchEBM. This page takes you from a fresh clone to a merged pull request; the other three pages cover the code itself:

• Architecture: package layout and how the core abstractions compose.
• Code Guidelines: style, API design, performance rules, testing.
• Performance and Benchmarking: writing fast code, and measuring it.

Setup

Requirements: Python 3.9+, Git, a GitHub account.

# Fork on GitHub, then:
git clone https://github.com/<your-user>/torchebm.git
cd torchebm
python -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
pip install -e ".[dev]"

For docs work, additionally:

pip install -e ".[docs]"
mkdocs serve                     # live preview at http://127.0.0.1:8000

Workflow

1. Branch from master with a descriptive name.

   git checkout -b feat/adaptive-step-size
   

2. Code: follow the Code Guidelines. Mirror the package layout under tests/.
3. Test and format before every commit.

   black torchebm/ tests/
   isort torchebm/ tests/
   pytest tests/ -v
   

4. Commit using Conventional Commits (below).
5. Push and open a PR against master. Link any related issue.

Commit conventions (mandatory)

TorchEBM uses Conventional Commits:

<type>(<optional scope>): <summary>

Type	Use for
feat	A new user-facing feature
fix	A bug fix
perf	A change that improves performance
refactor	Code change that is neither a feature nor a fix
test	Adding or fixing tests
docs	Documentation only
style	Formatting, whitespace, no logic change
build	Build system, packaging
ci	CI configuration
chore	Other maintenance (deps, tooling)

Examples

feat(samplers): add adaptive step size to LangevinDynamics
fix(losses): correct gradient sign in EquilibriumMatching
perf(integrators): cache RK buffers on device once per integrate()

Breaking changes add a ! after the type and a BREAKING CHANGE: footer. Keep the summary under 72 chars, imperative mood, no trailing period.

Before opening the PR

• Tests pass: pytest tests/ -v
• Formatting applied: black and isort
• Public API changes documented in the relevant docstring
• If you touched an example: pytest -m examples tests/examples passes
• If performance-sensitive: benchmarked per Performance and Benchmarking
• Commit messages follow Conventional Commits

The PR description should say what changed and why, and reference any issue it closes (Closes #123).