20+ high-performance LLMs with recipes to pretrain, finetune, deploy at scale.
✅ From scratch implementations ✅ No abstractions ✅ Beginner friendly ✅ Flash attention ✅ FSDP ✅ LoRA, QLoRA, Adapter ✅ Reduce GPU memory (fp4/8/16/32) ✅ 1-1000+ GPUs/TPUs ✅ 20+ LLMs
Lightning AI • Quick start • Models • Finetune • Deploy • All workflows • Features • Recipes (YAML) • Tutorials
Every LLM is implemented from scratch with no abstractions and full control, making them blazing fast, minimal, and performant at enterprise scale.
✅ Enterprise ready - Apache 2.0 for unlimited enterprise use.
✅ Developer friendly - Easy debugging with no abstraction layers and single file implementations.
✅ Optimized performance - Models designed to maximize performance, reduce costs, and speed up training.
✅ Proven recipes - Highly-optimized training/finetuning recipes tested at enterprise scale.
Install LitGPT
pip install 'litgpt[all]'
Load and use any of the 20+ LLMs:
from litgpt import LLM
llm = LLM.load("microsoft/phi-2")
text = llm.generate("Fix the spelling: Every fall, the familly goes to the mountains.")
print(text)
# Corrected Sentence: Every fall, the family goes to the mountains.
✅ Optimized for fast inference
✅ Quantization
✅ Runs on low-memory GPUs
✅ No layers of internal abstractions
✅ Optimized for production scale
Advanced install options
Install from source:
git clone https://github.com/Lightning-AI/litgpt
cd litgpt
pip install -e '.[all]'Explore the full Python API docs.
Every model is written from scratch to maximize performance and remove layers of abstraction:
| Model | Model size | Author | Reference |
|---|---|---|---|
| Llama 3 | 8B, 70B | Meta AI | Meta AI 2024 |
| Llama 2 | 7B, 13B, 70B | Meta AI | Touvron et al. 2023 |
| Code Llama | 7B, 13B, 34B, 70B | Meta AI | Rozière et al. 2023 |
| Mixtral MoE | 8x7B | Mistral AI | Mistral AI 2023 |
| Mistral | 7B | Mistral AI | Mistral AI 2023 |
| CodeGemma | 7B | Google Team, Google Deepmind | |
| ... | ... | ... | ... |
See full list of 20+ LLMs
| Model | Model size | Author | Reference |
|---|---|---|---|
| CodeGemma | 7B | Google Team, Google Deepmind | |
| Code Llama | 7B, 13B, 34B, 70B | Meta AI | Rozière et al. 2023 |
| Danube2 | 1.8B | H2O.ai | H2O.ai |
| Dolly | 3B, 7B, 12B | Databricks | Conover et al. 2023 |
| Falcon | 7B, 40B, 180B | TII UAE | TII 2023 |
| FreeWilly2 (Stable Beluga 2) | 70B | Stability AI | Stability AI 2023 |
| Function Calling Llama 2 | 7B | Trelis | Trelis et al. 2023 |
| Gemma | 2B, 7B | Google Team, Google Deepmind | |
| Llama 2 | 7B, 13B, 70B | Meta AI | Touvron et al. 2023 |
| Llama 3 | 8B, 70B | Meta AI | Meta AI 2024 |
| LongChat | 7B, 13B | LMSYS | LongChat Team 2023 |
| MicroLlama | 300M | Ken Wang | MicroLlama repo |
| Mixtral MoE | 8x7B | Mistral AI | Mistral AI 2023 |
| Mistral | 7B | Mistral AI | Mistral AI 2023 |
| Nous-Hermes | 7B, 13B, 70B | NousResearch | Org page |
| OpenLLaMA | 3B, 7B, 13B | OpenLM Research | Geng & Liu 2023 |
| Phi 1.5 & 2 | 1.3B, 2.7B | Microsoft Research | Li et al. 2023 |
| Phi 3 | 3.8B | Microsoft Research | Abdin et al. 2024 |
| Platypus | 7B, 13B, 70B | Lee et al. | Lee, Hunter, and Ruiz 2023 |
| Pythia | {14,31,70,160,410}M, {1,1.4,2.8,6.9,12}B | EleutherAI | Biderman et al. 2023 |
| RedPajama-INCITE | 3B, 7B | Together | Together 2023 |
| StableCode | 3B | Stability AI | Stability AI 2023 |
| StableLM | 3B, 7B | Stability AI | Stability AI 2023 |
| StableLM Zephyr | 3B | Stability AI | Stability AI 2023 |
| TinyLlama | 1.1B | Zhang et al. | Zhang et al. 2023 |
| Vicuna | 7B, 13B, 33B | LMSYS | Li et al. 2023 |
Tip: You can list all available models by running the litgpt download list command.
Finetune • Pretrain • Continued pretraining • Evaluate • Deploy • Test
Use the command line interface to run advanced workflows such as pretraining or finetuning on your own data.
After installing LitGPT, select the model and workflow to run (finetune, pretrain, evaluate, deploy, etc...):
# ligpt [action] [model]
litgpt download meta-llama/Meta-Llama-3-8B-Instruct
litgpt chat meta-llama/Meta-Llama-3-8B-Instruct
litgpt evaluate meta-llama/Meta-Llama-3-8B-Instruct
litgpt finetune meta-llama/Meta-Llama-3-8B-Instruct
litgpt pretrain meta-llama/Meta-Llama-3-8B-Instruct
litgpt serve meta-llama/Meta-Llama-3-8B-Instruct
Finetuning is the process of taking a pretrained AI model and further training it on a smaller, specialized dataset tailored to a specific task or application.
# 0) setup your dataset
curl -L https://huggingface.co/datasets/ksaw008/finance_alpaca/resolve/main/finance_alpaca.json -o my_custom_dataset.json
# 1) Download a pretrained model
litgpt download microsoft/phi-2
# 2) Finetune the model
litgpt finetune microsoft/phi-2 \
--data JSON \
--data.json_path my_custom_dataset.json \
--data.val_split_fraction 0.1 \
--out_dir out/custom-model
# 3) Test the model
litgpt chat out/custom-model/final
# 4) Deploy the model
litgpt serve out/custom-model/final
Deploy a pretrained or finetune LLM to use it in real-world applications. Deploy, automatically sets up a web server that can be accessed by a website or app.
# deploy an out-of-the-box LLM
litgpt download microsoft/phi-2
litgpt serve microsoft/phi-2
# deploy your own trained model
litgpt serve path/to/microsoft/phi-2/checkpointShow code to query server:
Test the server in a separate terminal and integrate the model API into your AI product:
# 3) Use the server (in a separate Python session)
import requests, json
response = requests.post(
"http://127.0.0.1:8000/predict",
json={"prompt": "Fix typos in the following sentence: Exampel input"}
)
print(response.json()["output"])
Evaluate an LLM to test its performance on various tasks to see how well it understands and generates text. Simply put, we can evaluate things like how well would it do in college-level chemistry, coding, etc... (MMLU, Truthful QA, etc...)
litgpt evaluate microsoft/phi-2 --tasks 'truthfulqa_mc2,mmlu'Read the full evaluation docs.
Test how well the model works via an interactive chat. Use the chat command to chat, extract embeddings, etc...
Here's an example showing how to use the Phi-2 LLM:
litgpt chat microsoft/phi-2
>> Prompt: What do Llamas eat?Full code:
# 1) Download the LLM
litgpt download list
litgpt download microsoft/phi-2
# 2) Test the model
litgpt chat microsoft/phi-2
>> Prompt: What do Llamas eat?The download of certain models requires an additional access token. You can read more about this in the download documentation.
Pretraining is the process of teaching an AI model by exposing it to a large amount of data before it is fine-tuned for specific tasks.
Show code:
mkdir -p custom_texts
curl https://www.gutenberg.org/cache/epub/24440/pg24440.txt --output custom_texts/book1.txt
curl https://www.gutenberg.org/cache/epub/26393/pg26393.txt --output custom_texts/book2.txt
# 1) Download a tokenizer
litgpt download EleutherAI/pythia-160m \
--tokenizer_only True
# 2) Pretrain the model
litgpt pretrain EleutherAI/pythia-160m \
--tokenizer_dir EleutherAI/pythia-160m \
--data TextFiles \
--data.train_data_path "custom_texts/" \
--train.max_tokens 10_000_000 \
--out_dir out/custom-model
# 3) Test the model
litgpt chat out/custom-model/finalRead the full pretraining docs
Continued pretraining is another way of finetuning that specializes an already pretrained model by training on custom data:
Show code:
mkdir -p custom_texts
curl https://www.gutenberg.org/cache/epub/24440/pg24440.txt --output custom_texts/book1.txt
curl https://www.gutenberg.org/cache/epub/26393/pg26393.txt --output custom_texts/book2.txt
# 1) Download a pretrained model
litgpt download EleutherAI/pythia-160m
# 2) Continue pretraining the model
litgpt pretrain EleutherAI/pythia-160m \
--tokenizer_dir EleutherAI/pythia-160m \
--initial_checkpoint_dir EleutherAI/pythia-160m \
--data TextFiles \
--data.train_data_path "custom_texts/" \
--train.max_tokens 10_000_000 \
--out_dir out/custom-model
# 3) Test the model
litgpt chat out/custom-model/finalRead the full continued pretraining docs
✅ State-of-the-art optimizations: Flash Attention v2, multi-GPU support via fully-sharded data parallelism, optional CPU offloading, and TPU and XLA support.
✅ Pretrain, finetune, and deploy
✅ Reduce compute requirements with low-precision settings: FP16, BF16, and FP16/FP32 mixed.
✅ Lower memory requirements with quantization: 4-bit floats, 8-bit integers, and double quantization.
✅ Configuration files for great out-of-the-box performance.
✅ Parameter-efficient finetuning: LoRA, QLoRA, Adapter, and Adapter v2.
✅ Exporting to other popular model weight formats.
✅ Many popular datasets for pretraining and finetuning, and support for custom datasets.
✅ Readable and easy-to-modify code to experiment with the latest research ideas.
LitGPT comes with validated recipes (YAML configs) to train models under different conditions. We've generated these recipes based on the parameters we found to perform the best for different training conditions.
Browse all training recipes here.
litgpt finetune \
--config https://raw.githubusercontent.com/Lightning-AI/litgpt/main/config_hub/finetune/llama-2-7b/lora.yaml✅ Use configs to customize training
Configs let you customize training for all granular parameters like:
# The path to the base model's checkpoint directory to load for finetuning. (type: <class 'Path'>, default: checkpoints/stabilityai/stablelm-base-alpha-3b)
checkpoint_dir: checkpoints/meta-llama/Llama-2-7b-hf
# Directory in which to save checkpoints and logs. (type: <class 'Path'>, default: out/lora)
out_dir: out/finetune/qlora-llama2-7b
# The precision to use for finetuning. Possible choices: "bf16-true", "bf16-mixed", "32-true". (type: Optional[str], default: null)
precision: bf16-true
...✅ Example: LoRA finetuning config
# The path to the base model's checkpoint directory to load for finetuning. (type: <class 'Path'>, default: checkpoints/stabilityai/stablelm-base-alpha-3b)
checkpoint_dir: checkpoints/meta-llama/Llama-2-7b-hf
# Directory in which to save checkpoints and logs. (type: <class 'Path'>, default: out/lora)
out_dir: out/finetune/qlora-llama2-7b
# The precision to use for finetuning. Possible choices: "bf16-true", "bf16-mixed", "32-true". (type: Optional[str], default: null)
precision: bf16-true
# If set, quantize the model with this algorithm. See ``tutorials/quantize.md`` for more information. (type: Optional[Literal['nf4', 'nf4-dq', 'fp4', 'fp4-dq', 'int8-training']], default: null)
quantize: bnb.nf4
# How many devices/GPUs to use. (type: Union[int, str], default: 1)
devices: 1
# How many nodes to use. (type: int, default: 1)
num_nodes: 1
# The LoRA rank. (type: int, default: 8)
lora_r: 32
# The LoRA alpha. (type: int, default: 16)
lora_alpha: 16
# The LoRA dropout value. (type: float, default: 0.05)
lora_dropout: 0.05
# Whether to apply LoRA to the query weights in attention. (type: bool, default: True)
lora_query: true
# Whether to apply LoRA to the key weights in attention. (type: bool, default: False)
lora_key: false
# Whether to apply LoRA to the value weights in attention. (type: bool, default: True)
lora_value: true
# Whether to apply LoRA to the output projection in the attention block. (type: bool, default: False)
lora_projection: false
# Whether to apply LoRA to the weights of the MLP in the attention block. (type: bool, default: False)
lora_mlp: false
# Whether to apply LoRA to output head in GPT. (type: bool, default: False)
lora_head: false
# Data-related arguments. If not provided, the default is ``litgpt.data.Alpaca``.
data:
class_path: litgpt.data.Alpaca2k
init_args:
mask_prompt: false
val_split_fraction: 0.05
prompt_style: alpaca
ignore_index: -100
seed: 42
num_workers: 4
download_dir: data/alpaca2k
# Training-related arguments. See ``litgpt.args.TrainArgs`` for details
train:
# Number of optimizer steps between saving checkpoints (type: Optional[int], default: 1000)
save_interval: 200
# Number of iterations between logging calls (type: int, default: 1)
log_interval: 1
# Number of samples between optimizer steps across data-parallel ranks (type: int, default: 128)
global_batch_size: 8
# Number of samples per data-parallel rank (type: int, default: 4)
micro_batch_size: 2
# Number of iterations with learning rate warmup active (type: int, default: 100)
lr_warmup_steps: 10
# Number of epochs to train on (type: Optional[int], default: 5)
epochs: 4
# Total number of tokens to train on (type: Optional[int], default: null)
max_tokens:
# Limits the number of optimizer steps to run (type: Optional[int], default: null)
max_steps:
# Limits the length of samples (type: Optional[int], default: null)
max_seq_length: 512
# Whether to tie the embedding weights with the language modeling head weights (type: Optional[bool], default: null)
tie_embeddings:
# (type: float, default: 0.0003)
learning_rate: 0.0002
# (type: float, default: 0.02)
weight_decay: 0.0
# (type: float, default: 0.9)
beta1: 0.9
# (type: float, default: 0.95)
beta2: 0.95
# (type: Optional[float], default: null)
max_norm:
# (type: float, default: 6e-05)
min_lr: 6.0e-05
# Evaluation-related arguments. See ``litgpt.args.EvalArgs`` for details
eval:
# Number of optimizer steps between evaluation calls (type: int, default: 100)
interval: 100
# Number of tokens to generate (type: Optional[int], default: 100)
max_new_tokens: 100
# Number of iterations (type: int, default: 100)
max_iters: 100
# The name of the logger to send metrics to. (type: Literal['wandb', 'tensorboard', 'csv'], default: csv)
logger_name: csv
# The random seed to use for reproducibility. (type: int, default: 1337)
seed: 1337✅ Override any parameter in the CLI:
litgpt finetune \
--config https://raw.githubusercontent.com/Lightning-AI/litgpt/main/config_hub/finetune/llama-2-7b/lora.yaml \
--lora_r 4
LitGPT powers many great AI projects, initiatives, challenges and of course enterprises. Please submit a pull request to be considered for a feature.
📊 SAMBA: Simple Hybrid State Space Models for Efficient Unlimited Context Language Modeling
The Samba project by researchers at Microsoft is built on top of the LitGPT code base and combines state space models with sliding window attention, which outperforms pure state space models.
🏆 NeurIPS 2023 Large Language Model Efficiency Challenge: 1 LLM + 1 GPU + 1 Day
The LitGPT repository was the official starter kit for the NeurIPS 2023 LLM Efficiency Challenge, which is a competition focused on finetuning an existing non-instruction tuned LLM for 24 hours on a single GPU.
🦙 TinyLlama: An Open-Source Small Language Model
LitGPT powered the TinyLlama project and TinyLlama: An Open-Source Small Language Model research paper.
🍪 MicroLlama: MicroLlama-300M
MicroLlama is a 300M Llama model pretrained on 50B tokens powered by TinyLlama and LitGPT.
🔬 Pre-training Small Base LMs with Fewer Tokens
The research paper "Pre-training Small Base LMs with Fewer Tokens", which utilizes LitGPT, develops smaller base language models by inheriting a few transformer blocks from larger models and training on a tiny fraction of the data used by the larger models. It demonstrates that these smaller models can perform comparably to larger models despite using significantly less training data and resources.
We welcome all individual contributors, regardless of their level of experience or hardware. Your contributions are valuable, and we are excited to see what you can accomplish in this collaborative and supportive environment.
🚀 Get started
⚡️ Finetuning, incl. LoRA, QLoRA, and Adapters
🤖 Pretraining
💬 Model evaluation
📘 Supported and custom datasets
🧹 Quantization
🤯 Tips for dealing with out-of-memory (OOM) errors
🧑🏽💻 Using cloud TPUs
This implementation extends on Lit-LLaMA and nanoGPT, and it's powered by Lightning Fabric ⚡.
- @karpathy for nanoGPT
- @EleutherAI for GPT-NeoX and the Evaluation Harness
- @TimDettmers for bitsandbytes
- @Microsoft for LoRA
- @tridao for Flash Attention 2
LitGPT is released under the Apache 2.0 license.
If you use LitGPT in your research, please cite the following work:
@misc{litgpt-2023,
author = {Lightning AI},
title = {LitGPT},
howpublished = {\url{https://github.com/Lightning-AI/litgpt}},
year = {2023},
}