Unsloth MCP Server logo

Unsloth MCP Server

by OtotaO

The Unsloth MCP Server provides an interface to leverage the Unsloth library for efficient fine-tuning of large language models. It allows users to load, fine-tune, generate text, and export models with optimized memory usage and speed.

View on GitHub

Last updated: N/A

What is Unsloth MCP Server?

The Unsloth MCP Server is a server application that exposes the functionality of the Unsloth library through a standardized interface. It enables users to interact with Unsloth's optimized fine-tuning and inference capabilities for large language models (LLMs) via MCP (Model Control Protocol).

How to use Unsloth MCP Server?

To use the server, first install Unsloth and build the server using npm install and npm run build. Then, configure your MCP settings to include the server, specifying the command to run the server and any necessary environment variables. You can then use the available tools like check_installation, list_supported_models, load_model, finetune_model, generate_text, and export_model by sending requests to the server with appropriate parameters.

Key features of Unsloth MCP Server

  • Optimize fine-tuning for Llama, Mistral, Phi, Gemma, and other models

  • 4-bit quantization for efficient training

  • Extended context length support

  • Simple API for model loading, fine-tuning, and inference

  • Export to various formats (GGUF, Hugging Face, etc.)

Use cases of Unsloth MCP Server

  • Fine-tuning LLMs on custom datasets with limited VRAM

  • Generating text with fine-tuned models

  • Exporting models to different formats for deployment

  • Evaluating the performance of different models and fine-tuning strategies

FAQ from Unsloth MCP Server

What Python versions are supported?

Python 3.10, 3.11, or 3.12 are supported. Python 3.13 is not supported.

What CUDA version is recommended?

CUDA 11.8 or 12.1+ is recommended.

What do I do if I get a CUDA Out of Memory error?

Reduce batch size, use 4-bit quantization, or try a smaller model.

How can I use a custom dataset for fine-tuning?

Format your dataset properly and host it on Hugging Face or provide a local path. Specify the dataset name as 'json' and provide the path to your data file in the data_files argument.

What formats can I export the fine-tuned model to?

You can export the model to GGUF, Ollama, vLLM, or Hugging Face formats.