Question about the `# Copied from` system

[delmalih]

Hi team! 👋

While working on improving docstrings and type hints across scheduler files (issue #9567), I've noticed the # Copied from pattern used extensively throughout the codebase.

Examples:

• Functions like betas_for_alpha_bar are duplicated across multiple schedulers
• Output classes like DDPMSchedulerOutput are copied with name replacements (e.g., DDPM->EulerDiscrete)

My question: What's the rationale behind this duplication system instead of:

1. Using a shared utils.py or common.py file for common functions
2. Using class inheritance for similar Output classes

I understand there might be good architectural reasons (module independence, API stability, avoiding circular dependencies, etc.), but this isn't documented anywhere that I could find.

Suggested action: Regardless of the answer, I think we should either:

• Option A: Refactor to use inheritance/shared utilities (if the current system is legacy)
• Option B: Document this design decision in:
    - A CONTRIBUTING.md or architecture doc
    - Comments in the utils/check_copies.py script itself
    - Another README in the diffusers directory

This would help future contributors (like me! 😅) understand why this pattern exists and how to work with it properly when improving documentation. What do you think?

Thanks for maintaining such a great library! 🚀

[DN6]

Hi @delmalih Here's a bit of background about the #Copied From pattern inherited from transformers.
https://huggingface.co/blog/transformers-design-philosophy

We are actively working to refine this pattern, because it is very likely that it has been overused in certain places, and perhaps we can streamline some aspects of the library through inheritance and shared modules.

[delmalih]

Thanks @DN6 for the context!

Update: I just discovered that the # Copied from mechanism is actually already documented in docs/source/en/conceptual/contribution.md! 🎉

I noticed we have two very similar contribution files:

• CONTRIBUTING.md (root)
• docs/source/en/conceptual/contribution.md

They cover mostly the same content but can get out of sync. Would it make sense to make one a symlink to the other to avoid maintaining two versions? Or is there a reason they need to stay separate?

Regarding overuse identification:
I'm still happy to help audit the codebase to identify places where the pattern has been overused. I could create a follow-up issue listing candidates for refactoring, such as:

• Utility functions that are truly generic (like betas_for_alpha_bar) and unlikely to diverge
• Output dataclasses that are identical across multiple schedulers
• Helper functions that don't contain model-specific logic

Let me know if this would be useful! 🚀