mergoo: Combine Fine-Tuned LLM Experts into Routed Models

Overview

Mergoo is a Python library for composing specialized LLMs or LoRA adapters into a model that routes inputs among experts. It addresses the work of bringing together domain or task-specific model knowledge while retaining the option to train the composed model.

It may suit practitioners already using Hugging Face model workflows who want to experiment with mixture-of-experts or mixture-of-adapters designs. The repository documents support for several model families and training paths; check the README and its notebooks for implementation details.

Key Features

• Compose fully fine-tuned models as a Mixture-of-Experts model.
• Combine LoRA fine-tunes as a Mixture-of-Adapters model.
• Apply routing selectively across layers.
• Configure the number of experts used per token.
• Save a composed model checkpoint for loading and further training.
• Integrate with Hugging Face Trainer, TRL's SFTTrainer, and PEFT.
• Documented base-model support includes Llama, Mistral, Phi-3, and BERT.
• Supports CPU, MPS, and GPU devices.

Use Cases

• A model-development team can combine math- and code-specialized fine-tunes when it wants a single routed model to use both areas of expertise.
• Teams with multiple LoRA adapters for related support tasks can combine them into a mixture of adapters instead of managing each adapter independently.
• Researchers can explore layer-wise routing and expert composition with Hugging Face training workflows.
• Practitioners can compose a model first, then fine-tune its routers or the full merged model for a target dataset.

Getting Started

Install from PyPI:

pip install mergoo

For configuration, model composition, and training examples, see the README and linked notebooks.

Considerations

• Composing and training large language models can require substantial memory and compute. The README lists CPU, MPS, and GPU support, but does not specify resource requirements.
• The documented support is limited to named model families and Hugging Face-based training integrations. Verify compatibility with your specific model and workflow before adopting it.
• The roadmap lists router load-balancing loss and lazy tensor loading as incomplete, so these capabilities should not be assumed.
• The LGPL-3.0 license has conditions worth reviewing for your distribution and integration plans.