Unsloth for DGX Spark

Unsloth Spark is a Docker-based environment for running Unsloth, a library that enables faster LLM fine-tuning with minimal memory usage.

This project take's NVIDIA's PyTorch container, upgrades a few libs per Unsloth's Dockerfile, then tries to emulate Unsloth's own unsloth/unsloth container by cloning in their notebooks and installing Jupyter.

I imagine this will not be needed if Unsloth creates their own native image. As of today, 11/19/2025, that is not the case.

Prerequisites

• DGX Spark or related 3rd party device

Quick Start

1. Clone this repository:

   git clone https://github.com/eleqtrizit/Unsloth-DGX-Spark.git
   cd unsloth_spark

2. Build the container:

   make build

3. Login to container bash:

   make run

4. Test (after make run)!

   cd work && python test_unsloth.py

## Makefile Commands

The project includes a Makefile with the following commands:

```sh
> make
Available targets:
  build            - Build the container
  build-no-cache   - Build the container without using cache
  run              - Run bash from the container
  clean            - Remove the container image
  jupyter          - Run JupyterLab from the container

Manual Docker Commands

Read the Makefile for example commands. Ask an LLM what the params mean if you're unsure.

Features

• Based on NVIDIA's PyTorch container (25.09-py3) for GPU optimization
• Pre-installed JupyterLab for interactive development
• Volume mounting for persistent work data

Example Usage

The test_unsloth.py script demonstrates how to:

1. Load a pre-quantized 4-bit model
2. Apply LoRA fine-tuning
3. Train on a dataset using SFTTrainer

Updating the Container

Simply rebuild. make build