---
title: "Fine-Tune Faster with Unsloth — Overview"
canonical: "https://build.nvidia.com/playbooks/unsloth/overview.md"
---

# Basic idea

Unsloth is a performance-focused library for fine-tuning large language models. It reduces training time and memory use compared with standard parameter-efficient fine-tuning workflows, with a streamlined Python interface.

- **Faster training** — custom kernels and optimized math paths aim for higher throughput on a single GPU
- **Lower memory use** — 4-bit and 16-bit quantization options help fit larger models in available memory
- **Broad model support** — works with many open LLMs (for example Llama, Mistral, Qwen, DeepSeek) and exports to formats used by tools such as Ollama, vLLM, GGUF, and Hugging Face
- **Minimal boilerplate** — notebooks and helpers reduce setup for LoRA and QLoRA fine-tuning

# What you'll accomplish

You'll set up Unsloth for optimized fine-tuning of large language models on your **hardware platform**, with parameter-efficient methods such as LoRA and QLoRA for faster training and reduced memory use.

# What to know before starting

**Required:**

- Python package management with pip and virtual environments or containers
- Hugging Face Transformers basics (loading models, tokenizers, datasets)
- GPU fundamentals (CUDA vs CPU, memory constraints, device availability)
- Basic LLM training concepts (loss, checkpoints)

**Optional:**

- Familiarity with prompt engineering and base model interaction
- LoRA / QLoRA parameter-efficient fine-tuning knowledge

# Supported hardware platforms

Use the matrix below to confirm your hardware platform, recommended default local settings, and whether multi-node applies.

| Hardware platform | OS | Memory | Recommended default local settings | Multi-node capable hardware |
| :---- | :---- | :---- | :---- | :---- |
| **DGX Spark** | DGX OS (Linux) | 128 GB Unified Memory | `nvcr.io/nvidia/pytorch:25.11-py3` | — |

# Prerequisites

**Hardware requirements**

- Supported hardware platform — see Supported hardware platforms matrix above
- Sufficient memory and free storage for model and dataset downloads (plan for several GB or more depending on model size)

**Software requirements**

- Docker installed with GPU support: `docker --version`
- NVIDIA Container Toolkit configured (`nvidia-smi` works inside a GPU container)
- CUDA toolkit available for verification: `nvcc --version` (expect CUDA 13.0)
- Network access to download container images, models, and datasets

# Ancillary files

All required assets can be found [in this playbook repository](https://github.com/NVIDIA/dgx-spark-playbooks/blob/main/nvidia/playbook-unsloth/).

- `assets/test_unsloth.py` — validation script that runs a short LoRA fine-tuning job to confirm Unsloth is installed correctly

# Time & risk

- **Estimated time:** 60 MIN (first run includes container pull, package install, and a short validation train)
- **Risk level:** Medium
- Triton compiler or CUDA toolkit mismatches can block kernel compilation
- Memory pressure may require smaller batch sizes or shorter sequence lengths
- Large model or dataset downloads can fail on limited network or disk space
- **Rollback:** Exit and remove the container; optionally uninstall packages with `pip uninstall unsloth unsloth_zoo torch torchvision` if you installed outside the ephemeral container
- **Last Updated:** 07/31/2026
- Set up Unsloth for optimized LoRA/QLoRA fine-tuning with a containerized validation workflow