80% faster 50% less memory local QLoRA finetuning

Manual autograd engine - hand derived backprop steps.
QLoRA / LoRA 80% faster, 50% less memory.
All kernels written in OpenAI's Triton language.
0% loss in accuracy - no approximation methods - all exact.
No change of hardware necessary. Supports NVIDIA GPUs since 2018+. CUDA 7.5+. Tesla T4, RTX 20, 30, 40 series, A100, H100s
Flash Attention support via Xformers.
Supports 4bit and 16bit LoRA finetuning.
Train Slim Orca fully locally in 260 hours from 1301 hours (5x faster).
Open source version trains 5x faster or you can check out Unsloth Pro and Max codepaths for 30x faster training!

Try our Colab examples for the Alpaca 52K dataset or the Slim Orca 518K dataset.
Try our Kaggle example for the LAION OIG Chip2 dataset
Join our Discord!

Installation Instructions - Conda

Unsloth currently only supports Linux distros and Pytorch >= 2.1.

conda install cudatoolkit xformers bitsandbytes pytorch pytorch-cuda=12.1 \
  -c pytorch -c nvidia -c xformers -c conda-forge -y
pip install "unsloth[kaggle] @ git+https://github.com/unslothai/unsloth.git"

Installation Instructions - Pip

Find your CUDA version via

import torch; torch.version.cuda

Select either cu118 for CUDA 11.8 or cu121 for CUDA 12.1

pip install "unsloth[cu118] @ git+https://github.com/unslothai/unsloth.git"
pip install "unsloth[cu121] @ git+https://github.com/unslothai/unsloth.git"

We only support Pytorch 2.1: You can update Pytorch via Pip:

pip install --upgrade --force-reinstall --no-cache-dir torch triton \
  --index-url https://download.pytorch.org/whl/cu121

Change cu121 to cu118 for CUDA version 11.8 or 12.1. Go to https://pytorch.org/ to learn more.

Alpaca Example

from unsloth import FastLlamaModel
import torch
max_seq_length = 2048 # Can change to any number <= 4096
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.

# Load Llama model
model, tokenizer = FastLlamaModel.from_pretrained(
    model_name = "unsloth/llama-2-7b", # Supports any llama model
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = load_in_4bit,
    # token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
)

# Do model patching and add fast LoRA weights
model = FastLlamaModel.get_peft_model(
    model,
    r = 16,
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj",],
    lora_alpha = 16,
    lora_dropout = 0, # Currently only supports dropout = 0
    bias = "none",    # Currently only supports bias = "none"
    use_gradient_checkpointing = True,
    random_state = 3407,
    max_seq_length = max_seq_length,
)

trainer = .... Use Huggingface's Trainer and dataset loading

If you trained a model with Unsloth, we made a cool sticker!!

Future Milestones and limitations

Support sqrt gradient checkpointing which further slashes memory usage by 25%.
Does not support non Llama models - we do so in the future.

Performance comparisons on 1 Tesla T4 GPU:

Time taken for 1 epoch

System	GPU	Alpaca (52K)	LAION OIG (210K)	Open Assistant (10K)	SlimOrca (518K)
Huggingface	1 T4	23h 15m	56h 28m	8h 38m	391h 41m
Unsloth Open	1 T4	13h 7m (1.8x)	31h 47m (1.8x)	4h 27m (1.9x)	240h 4m (1.6x)
Unsloth Pro	1 T4	3h 6m (7.5x)	5h 17m (10.7x)	1h 7m (7.7x)	59h 53m (6.5x)
Unsloth Max	1 T4	2h 39m (8.8x)	4h 31m (12.5x)	0h 58m (8.9x)	51h 30m (7.6x)

Peak Memory Usage

System	GPU	Alpaca (52K)	LAION OIG (210K)	Open Assistant (10K)	SlimOrca (518K)
Huggingface	1 T4	7.3GB	5.9GB	14.0GB	13.3GB
Unsloth Open	1 T4	6.8GB	5.7GB	7.8GB	7.7GB
Unsloth Pro	1 T4	6.4GB	6.4GB	6.4GB	6.4GB
Unsloth Max	1 T4	11.4GB	12.4GB	11.9GB	14.4GB

Performance comparisons on 2 Tesla T4 GPUs via DDP: