unsloth/README.md

Finetune Mistral, Llama 2-5x faster with 70% less memory!

✨ Finetune for Free

All notebooks are beginner friendly! Colab provides a free GPU. Kaggle provides 30 hours for free per week.

Unsloth supports	Free Notebooks	Performance	Memory use
Mistral 7b	▶️ Start on Colab	2.2x faster	62% less
Llama-2 7b	▶️ Start on Colab	2.2x faster	43% less
DPO - Zephyr	▶️ Start on Colab	1.9x faster	19% less
TinyLlama	▶️ Start on Colab	3.9x faster	74% less
CodeLlama 34b A100	▶️ Start on Colab	1.9x faster	27% less
Mistral 7b 2xT4	▶️ Start on Kaggle	5x faster	60% less

• This conversational notebook is useful for ShareGPT ChatML datatsets.
• Our raw text notebook is useful for text completion.

🦥 Unsloth.ai News

• 📣 DPO support is now included. More info on DPO.
• 📣 TinyLlama 1.1b on 3T tokens now works.
• 📣 We did a blog with 🤗Hugging Face! We're in 🤗Hugging Face's official docs! Check out the SFT docs and DPO docs.
• 📣 Now supports Llama, Yi, Mistral, CodeLlama, Qwen (llamafied), Deepseek and their derived models (Open Hermes etc). Llama 7, 13, 70b; CodeLlama 7, 13, 34, 70b; Yi 6, 34b are all supported!
• 📣 Download models 4x faster from 🤗Hugging Face! Eg: unsloth/mistral-7b-bnb-4bit See our HF collection for more!

🔗 Links and Resources

Type	Links
📜 Documentation	Read The Doc
💾 Installation	unsloth/README.md
Twitter (aka X)	Follow us on X
🥇 Benchmarking	Performance Tables
🌐 Released Models	Unsloth Releases
✍️ Blog	Read our Blogs

⭐ Key Features

• All kernels written in OpenAI's Triton language. Manual backprop engine.
• 0% loss in accuracy - no approximation methods - all exact.
• No change of hardware. Supports NVIDIA GPUs since 2018+. Minimum CUDA Capability 7.0 (V100, T4, Titan V, RTX 20, 30, 40x, A100, H100, L40 etc) Check your GPU! GTX 1070, 1080 works, but is slow.
• Works on Linux and Windows via WSL.
• Supports 4bit and 16bit QLoRA / LoRA finetuning via bitsandbytes.
• Open source trains 5x faster - see Unsloth Pro for 30x faster training!
• If you trained a model with 🦥Unsloth, you can use this cool sticker!  

🥇 Performance Benchmarking

• For the full list of reproducable benchmarking tables, go to our website

1 A100 40GB	🤗Hugging Face	Flash Attention	🦥Unsloth Open Source	🦥Unsloth Pro
Alpaca	1x	1.04x	1.98x	15.64x
LAION Chip2	1x	0.92x	1.61x	20.73x
OASST	1x	1.19x	2.17x	14.83x
Slim Orca	1x	1.18x	2.22x	14.82x

• Benchmarking table below was conducted by 🤗Hugging Face.

Free Colab T4	Dataset	🤗Hugging Face	Pytorch 2.1.1	🦥Unsloth	🦥 VRAM reduction
Llama-2 7b	OASST	1x	1.19x	1.95x	-43.3%
Mistral 7b	Alpaca	1x	1.07x	1.56x	-13.7%
Tiny Llama 1.1b	Alpaca	1x	2.06x	3.87x	-73.8%
DPO with Zephyr	Ultra Chat	1x	1.09x	1.55x	-18.6%

💾 Installation Instructions

Conda Installation

Select either pytorch-cuda=11.8 for CUDA 11.8 or pytorch-cuda=12.1 for CUDA 12.1. If you have mamba, use mamba instead of conda for faster solving. See this Github issue for help on debugging Conda installs.

conda install pytorch torchvision torchaudio pytorch-cuda=<12.1/11.8> -c pytorch -c nvidia

conda install xformers -c xformers -y

pip install bitsandbytes

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

Pip Installation

Do NOT use this if you have Anaconda. You must use the Conda install method, or else stuff will BREAK.

1. Find your CUDA version via

import torch; torch.version.cuda

2. For Pytorch 2.1.0: You can update Pytorch via Pip (interchange cu121 / cu118). Go to https://pytorch.org/ to learn more. Select either cu118 for CUDA 11.8 or cu121 for CUDA 12.1. If you have a RTX 3060 or higher (A100, H100 etc), use the "ampere" path. For Pytorch 2.1.1: go to step 3. For Pytorch 2.2.0: go to step 4.

pip install --upgrade --force-reinstall --no-cache-dir torch==2.1.0 triton \
  --index-url https://download.pytorch.org/whl/cu121

pip install "unsloth[cu118] @ git+https://github.com/unslothai/unsloth.git"
pip install "unsloth[cu121] @ git+https://github.com/unslothai/unsloth.git"
pip install "unsloth[cu118_ampere] @ git+https://github.com/unslothai/unsloth.git"
pip install "unsloth[cu121_ampere] @ git+https://github.com/unslothai/unsloth.git"

3. For Pytorch 2.1.1: Use the "ampere" path for newer RTX 30xx GPUs or higher.

pip install --upgrade --force-reinstall --no-cache-dir torch==2.1.1 triton \
  --index-url https://download.pytorch.org/whl/cu121

pip install "unsloth[cu118_torch211] @ git+https://github.com/unslothai/unsloth.git"
pip install "unsloth[cu121_torch211] @ git+https://github.com/unslothai/unsloth.git"
pip install "unsloth[cu118_ampere_torch211] @ git+https://github.com/unslothai/unsloth.git"
pip install "unsloth[cu121_ampere_torch211] @ git+https://github.com/unslothai/unsloth.git"

4. For Pytorch 2.2.0: Use the "ampere" path for newer RTX 30xx GPUs or higher.

pip install --upgrade --force-reinstall --no-cache-dir torch==2.2.0 triton \
  --index-url https://download.pytorch.org/whl/cu121

pip install "unsloth[cu118_torch220] @ git+https://github.com/unslothai/unsloth.git"
pip install "unsloth[cu121_torch220] @ git+https://github.com/unslothai/unsloth.git"
pip install "unsloth[cu118_ampere_torch220] @ git+https://github.com/unslothai/unsloth.git"
pip install "unsloth[cu121_ampere_torch220] @ git+https://github.com/unslothai/unsloth.git"

5. If you get errors, try the below first, then go back to step 1:

pip install --upgrade pip

📜 Documentation

• We support Huggingface's TRL, Trainer, Seq2SeqTrainer or even Pytorch code!
• We're in 🤗Hugging Face's official docs! Check out the SFT docs and DPO docs!

from unsloth import FastLanguageModel
import torch
from trl import SFTTrainer
from transformers import TrainingArguments
from datasets import load_dataset
max_seq_length = 2048 # Supports RoPE Scaling interally, so choose any!
# Get LAION dataset
url = "https://huggingface.co/datasets/laion/OIG/resolve/main/unified_chip2.jsonl"
dataset = load_dataset("json", data_files = {"train" : url}, split = "train")

# 4bit pre quantized models we support - 4x faster downloading!
fourbit_models = [
    "unsloth/mistral-7b-bnb-4bit",
    "unsloth/llama-2-7b-bnb-4bit",
    "unsloth/llama-2-13b-bnb-4bit",
    "unsloth/codellama-34b-bnb-4bit",
    "unsloth/tinyllama-bnb-4bit",
]
# Load Llama model
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/mistral-7b-bnb-4bit", # Supports Llama, Mistral - replace this!
    max_seq_length = max_seq_length,
    dtype = None,
    load_in_4bit = True,
)

# Do model patching and add fast LoRA weights
model = FastLanguageModel.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, # Supports any, but = 0 is optimized
    bias = "none",    # Supports any, but = "none" is optimized
    use_gradient_checkpointing = True,
    random_state = 3407,
    max_seq_length = max_seq_length,
)

trainer = SFTTrainer(
    model = model,
    train_dataset = dataset,
    dataset_text_field = "text",
    max_seq_length = max_seq_length,
    tokenizer = tokenizer,
    args = TrainingArguments(
        per_device_train_batch_size = 2,
        gradient_accumulation_steps = 4,
        warmup_steps = 10,
        max_steps = 60,
        fp16 = not torch.cuda.is_bf16_supported(),
        bf16 = torch.cuda.is_bf16_supported(),
        logging_steps = 1,
        output_dir = "outputs",
        optim = "adamw_8bit",
        seed = 3407,
    ),
)
trainer.train()

DPO Support

DPO (Direct Preference Optimization), PPO, Reward Modelling all seem to work as per 3rd party independent testing from Llama-Factory. We have a preliminary Google Colab notebook for reproducing Zephyr on Tesla T4 here: notebook.

We're in 🤗Hugging Face's official docs! We're on the SFT docs and the DPO docs!

from unsloth import FastLanguageModel, PatchDPOTrainer
PatchDPOTrainer()
import torch
from transformers import TrainingArguments
from trl import DPOTrainer

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/zephyr-sft-bnb-4bit",
    max_seq_length = max_seq_length,
    dtype = None,
    load_in_4bit = True,
)

# Do model patching and add fast LoRA weights
model = FastLanguageModel.get_peft_model(
    model,
    r = 64,
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj",],
    lora_alpha = 64,
    lora_dropout = 0, # Supports any, but = 0 is optimized
    bias = "none",    # Supports any, but = "none" is optimized
    use_gradient_checkpointing = True,
    random_state = 3407,
    max_seq_length = max_seq_length,
)

dpo_trainer = DPOTrainer(
    model = model,
    ref_model = None,
    args = TrainingArguments(
        per_device_train_batch_size = 4,
        gradient_accumulation_steps = 8,
        warmup_ratio = 0.1,
        num_train_epochs = 3,
        fp16 = not torch.cuda.is_bf16_supported(),
        bf16 = torch.cuda.is_bf16_supported(),
        logging_steps = 1,
        optim = "adamw_8bit",
        seed = 42,
        output_dir = "outputs",
    ),
    beta = 0.1,
    train_dataset = YOUR_DATASET_HERE,
    # eval_dataset = YOUR_DATASET_HERE,
    tokenizer = tokenizer,
    max_length = 1024,
    max_prompt_length = 512,
)
dpo_trainer.train()

🥇 Detailed Benchmarking Tables

• Click "Code" for fully reproducible examples
• "Unsloth Equal" is a preview of our PRO version, with code stripped out. All settings and the loss curve remains identical.
• For the full list of benchmarking tables, go to our website

1 A100 40GB	🤗Hugging Face	Flash Attention 2	🦥Unsloth Open	Unsloth Equal	Unsloth Pro	Unsloth Max
Alpaca	1x	1.04x	1.98x	2.48x	5.32x	15.64x
code	Code	Code	Code	Code
seconds	1040	1001	525	419	196	67
memory MB	18235	15365	9631	8525
% saved		15.74	47.18	53.25

Llama-Factory 3rd party benchmarking

• Link to performance table. TGS: tokens per GPU per second. Model: LLaMA2-7B. GPU: NVIDIA A100 * 1. Batch size: 4. Gradient accumulation: 2. LoRA rank: 8. Max length: 1024.

Method	Bits	TGS	GRAM	Speed
HF	16	2392	18GB	100%
HF+FA2	16	2954	17GB	123%
Unsloth+FA2	16	4007	16GB	168%
HF	4	2415	9GB	101%
Unsloth+FA2	4	3726	7GB	160%

Performance comparisons between popular models

Click for specific model benchmarking tables (Mistral 7b, CodeLlama 34b etc.)

Mistral 7b

1 A100 40GB	Hugging Face	Flash Attention 2	Unsloth Open	Unsloth Equal	Unsloth Pro	Unsloth Max
Mistral 7B Slim Orca	1x	1.15x	2.15x	2.53x	4.61x	13.69x
code	Code	Code	Code	Code
seconds	1813	1571	842	718	393	132
memory MB	32853	19385	12465	10271
% saved		40.99	62.06	68.74

CodeLlama 34b

1 A100 40GB	Hugging Face	Flash Attention 2	Unsloth Open	Unsloth Equal	Unsloth Pro	Unsloth Max
Code Llama 34B	OOM ❌	0.99x	1.87x	2.61x	4.27x	12.82x
code	▶️ Code	Code	Code	Code
seconds	1953	1982	1043	748	458	152
memory MB	40000	33217	27413	22161
% saved		16.96	31.47	44.60

1 Tesla T4

1 T4 16GB	Hugging Face	Flash Attention	Unsloth Open	Unsloth Pro Equal	Unsloth Pro	Unsloth Max
Alpaca	1x	1.09x	1.69x	1.79x	2.93x	8.3x
code	▶️ Code	Code	Code	Code
seconds	1599	1468	942	894	545	193
memory MB	7199	7059	6459	5443
% saved		1.94	10.28	24.39

2 Tesla T4s via DDP

2 T4 DDP	Hugging Face	Flash Attention	Unsloth Open	Unsloth Equal	Unsloth Pro	Unsloth Max
Alpaca	1x	0.99x	4.95x	4.44x	7.28x	20.61x
code	▶️ Code	Code	Code
seconds	9882	9946	1996	2227	1357	480
memory MB	9176	9128	6904	6782
% saved		0.52	24.76	26.09

Performance comparisons on 1 Tesla T4 GPU:

Click for Time taken for 1 epoch

One Tesla T4 on Google Colab bsz = 2, ga = 4, max_grad_norm = 0.3, num_train_epochs = 1, seed = 3047, lr = 2e-4, wd = 0.01, optim = "adamw_8bit", schedule = "linear", schedule_steps = 10

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

Click for Performance Comparisons on 2 Tesla T4 GPUs via DDP:

**Time taken for 1 epoch**

Two Tesla T4s on Kaggle bsz = 2, ga = 4, max_grad_norm = 0.3, num_train_epochs = 1, seed = 3047, lr = 2e-4, wd = 0.01, optim = "adamw_8bit", schedule = "linear", schedule_steps = 10

System	GPU	Alpaca (52K)	LAION OIG (210K)	Open Assistant (10K)	SlimOrca (518K) *
Huggingface	2 T4	84h 47m	163h 48m	30h 51m	1301h 24m *
Unsloth Pro	2 T4	3h 20m (25.4x)	5h 43m (28.7x)	1h 12m (25.7x)	71h 40m (18.1x) *
Unsloth Max	2 T4	3h 4m (27.6x)	5h 14m (31.3x)	1h 6m (28.1x)	54h 20m (23.9x) *

Peak Memory Usage on a Multi GPU System (2 GPUs)

System	GPU	Alpaca (52K)	LAION OIG (210K)	Open Assistant (10K)	SlimOrca (518K) *
Huggingface	2 T4	8.4GB | 6GB	7.2GB | 5.3GB	14.3GB | 6.6GB	10.9GB | 5.9GB *
Unsloth Pro	2 T4	7.7GB | 4.9GB	7.5GB | 4.9GB	8.5GB | 4.9GB	6.2GB | 4.7GB *
Unsloth Max	2 T4	10.5GB | 5GB	10.6GB | 5GB	10.6GB | 5GB	10.5GB | 5GB *

• Slim Orca bsz=1 for all benchmarks since bsz=2 OOMs. We can handle bsz=2, but we benchmark it with bsz=1 for consistency.

Credits

1. RandomInternetPreson for confirming WSL support
2. 152334H for experimental DPO support
3. atgctg for syntax highlighting