unsloth

name	unsloth
description	Fast QLoRA/QLoRA fine-tuning with 2x faster training and 50% less memory. Supports Llama, Mistral, Gemma, Qwen, DeepSeek, Phi, Yi, Falcon. Flash Attention, 4-bit quantization. No quality loss.
tags	["qlora-finetuning","lora-finetuning","memory-efficient-tuning","quantized-llm-training","unsloth"]

Overview

Unsloth provides 2x faster QLoRA training with 50% less memory via optimized kernels. Supports Llama, Mistral, Gemma, Qwen 2.5, DeepSeek, Phi, Yi, and Falcon with Flash Attention.

Installation

uv pip install unsloth

QLoRA Fine-Tuning

from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Qwen2.5-7B-Instruct-bnb-4bit",
    max_seq_length=4096,
    dtype=torch.bfloat16,
    load_in_4bit=True,
)
model = FastLanguageModel.get_peft_model(
    model, r=16, target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
    lora_alpha=16, use_gradient_checkpointing="unsloth",
)
print(model.print_trainable_parameters())

Inference

FastLanguageModel.for_inference(model)
inputs = tokenizer(["Describe quantum computing."], return_tensors="pt").to("cuda")
print(tokenizer.decode(model.generate(**inputs, max_new_tokens=256)[0]))

References

• Unsloth GitHub
• Unsloth docs