Architecture

Molfun is built on three principles:

Modularity -- Every component (attention, blocks, structure modules, embedders, losses) lives behind an abstract interface and a registry. Swap one implementation for another with a single line of code.

Extensibility -- Adding a new attention mechanism, training strategy, or model backend never requires modifying existing code. Register your class with a decorator and the framework picks it up.

Scientific rigor -- Protein structure prediction demands reproducible experiments. Molfun's training framework enforces consistent infrastructure (EMA, gradient accumulation, checkpointing, early stopping) across all fine-tuning strategies so results are comparable.

---

Architecture guides

Guide	What you will learn
System Overview	Logical layers, subsystem map, request flow
Design Patterns	Registry, Strategy, Adapter, Template Method, Facade
Module System	4 module families, registries, builder, swapper
Training Framework	Fine-tuning strategies, training loop, checkpointing
Data Pipeline	Sources, parsers, datasets, splits, storage