OpenParable
Skip to content
← Back to skills

Mflux PR

Development
mflux-pr

Mflux PR is an Agent Skill for development published in the filipstrand/mflux repository on GitHub: a SKILL.md instruction file an AI agent can load to do this task.

Make a clean PR in mflux (inspect diff, quick verification, commit, push, open PR) using repo conventions.

View on GitHub ↗
QUICK START

How to use this skill

Bring this guide into your coding agent with a prompt tailored to the tool you use.

  1. Open your project in Codex.
  2. Copy and paste the prompt below.
  3. Review files and risks before approving any action.
Prompt to paste
Inspect the Agent Skill at https://github.com/filipstrand/mflux/blob/HEAD/.cursor/skills/mflux-pr/SKILL.md and all supporting files. Report provenance, permissions, scripts, dependencies, and security risks. Propose an exact installation into .agents/skills/mflux-pr, but wait for my approval before writing files or running scripts.

Copying this prompt does not install or run the skill. Review third-party files before use.

mflux pull request workflow

When to Use

  • You’re about to open a PR (or want a safe sequence to do it).

Instructions

  • If you run tests as part of PR hygiene, prefer fast tests first:
    • make test-fast
  • Keep commits focused and messages consistent with repo history.
  • If the PR changes CLI defaults, public APIs, or model behavior, check for README/example drift before opening the PR.
  • Always ask for permission before pushing to the remote repository.
  • If gh isn’t available, fall back to the GitHub web UI (or stop and ask).

Pre-merge checklist (model port PRs)

Use after the core port lands and you are polishing for merge. For the full integration surfaces tick list (LoRA key formats, save routing, tokenizer edge cases, etc. learned from past closed PRs), see mflux-model-porting → Integration surfaces checklist.

Correctness

  1. make lint and make test-fast
  2. Slow golden tests for the new model:
    MFLUX_PRESERVE_TEST_OUTPUT=1 uv run pytest tests/image_generation/test_generate_image_<model>.py -m slow -v
    
  3. Optional but high-signal: diffusers side-by-side + latent injection (mflux-debugging, mflux-manual-testing)

Cross-model diff audit

List files changed outside src/mflux/models/<model>/:

CategoryExpected
pyproject.toml, cli/defaults/defaults.py, ModelConfig, mflux-save routingRequired wiring
README.md table + attributionRequired
Training runner.py, example JSON, .gitignore JSON exceptionsIf training supported
Shared VAE/callback/training one-linersOnly if required; document blast radius in PR
Personal .gitignore, unrelated formattingRemove

Verify quantized README disk claims with measurement:

du -sh ~/.cache/huggingface/hub/models--<org>--<Model>*
mflux-save --model <alias> --quantize 8 --path /tmp/model-q8 && du -sh /tmp/model-q8

Docs / examples

  • Model README matches a recent port (e.g. Flux2): hero image, turbo + base CLI, feature section, disk warning, Notes, Training.
  • Main README.md model table row (correct release date).
  • Showcase asset if other models have one (src/mflux/assets/; may need git add -f when *.jpg is gitignored).

PR description callouts

  • Shared code touched and why (shared VAE, callbacks, training runner, etc.).
  • Reference pipeline features intentionally not ported (optional preprocessors, extra encoders, components omitted from mflux weight downloads).
  • Known non-parity with diffusers (RNG, sigma schedule, optional modules) if golden tests lock mflux-native sampling.