Advanced Topics

This section covers advanced features of Molfun for scaling training, accelerating computation, deploying models, and building reproducible workflows.

Topics

Page	Description
Distributed Training	Multi-GPU training with DDP and FSDP, gradient checkpointing, mixed precision
Triton Kernels	Custom GPU kernels for analysis (RMSD, contact maps) and model operations (fused LayerNorm, GELU)
Model Export	Export models to ONNX and TorchScript for production deployment
HuggingFace Hub	Push and pull models from HuggingFace Hub, model cards, versioning
Pipeline Framework	Reproducible end-to-end workflows with YAML recipes

When to use what

flowchart TD
    A[I have a trained model] --> B{What next?}
    B --> C[Scale training] --> D[Distributed Training]
    B --> E[Speed up analysis] --> F[Triton Kernels]
    B --> G[Deploy to production] --> H[Model Export]
    B --> I[Share with others] --> J[HuggingFace Hub]
    B --> K[Automate everything] --> L[Pipeline Framework]

Getting started

If you are new to Molfun, start with the Getting Started guide first. These advanced topics assume familiarity with the core API (MolfunStructureModel, fine-tuning strategies, and data loading).