Run large language models at home, BitTorrent-style.
Fine-tuning and inference up to 10x faster than offloading

Generate text with distributed LLaMA 2 (70B, 70B-Chat), LLaMA-65B, Guanaco-65B or BLOOM-176B and fine‑tune them for your own tasks — right from your desktop computer or Google Colab:

from transformers import AutoTokenizer
from petals import AutoDistributedModelForCausalLM

model_name = "enoch/llama-65b-hf"
# You can also use "meta-llama/Llama-2-70b-hf", "meta-llama/Llama-2-70b-chat-hf",
# "bigscience/bloom", or "bigscience/bloomz"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoDistributedModelForCausalLM.from_pretrained(model_name)
# Embeddings & prompts are on your device, transformer blocks are distributed across the Internet

inputs = tokenizer("A cat sat", return_tensors="pt")["input_ids"]
outputs = model.generate(inputs, max_new_tokens=5)
print(tokenizer.decode(outputs[0]))  # A cat sat on a mat...

🚀  Try now in Colab

📋 Make sure you follow the model's terms of use (see LLaMA 2, LLaMA and BLOOM licenses).

🔏 Your data will be processed by other people in the public swarm. Learn more about privacy here. For sensitive data, you can set up a private swarm among people you trust.

Run these commands in an Anaconda env (requires Linux and Python 3.8+):

conda install pytorch pytorch-cuda=11.7 -c pytorch -c nvidia
pip install --upgrade petals
python -m petals.cli.run_server enoch/llama-65b-hf --adapters timdettmers/guanaco-65b

Or run our Docker image (works on Linux, macOS, and Windows with WSL2):

sudo docker run -p 31330:31330 --ipc host --gpus all --volume petals-cache:/cache --rm learningathome/petals:main \
    python -m petals.cli.run_server --port 31330 enoch/llama-65b-hf --adapters timdettmers/guanaco-65b

This will host a part of LLaMA-65B with optional Guanaco adapters on your machine. You can also host meta-llama/Llama-2-70b-hf, meta-llama/Llama-2-70b-chat-hf, bigscience/bloom, bigscience/bloomz, and other compatible models from 🤗 Model Hub, or add support for new model architectures.

🔒 Hosting a server does not allow others to run custom code on your computer. Learn more about security here.

💬 See FAQ to learn how to use multple GPUs, restart the server on reboot, etc. If you have any issues or feedback, ping us in our Discord!

Basic tutorials:

• Getting started: tutorial
• Prompt-tune LLaMA-65B for text semantic classification: tutorial
• Prompt-tune BLOOM to create a personified chatbot: tutorial

Useful tools and advanced guides:

• Chatbot web app (connects to Petals via an HTTP/WebSocket endpoint): source code
• Monitor for the public swarm: source code
• Launch your own swarm: guide
• Run a custom foundation model: guide

Learning more:

• Frequently asked questions: FAQ
• In-depth system description: paper

• Petals runs large language models like LLaMA and BLOOM collaboratively — you load a small part of the model, then team up with people serving the other parts to run inference or fine-tuning.
• Single-batch inference runs at up to 6 steps/sec for LLaMA 2 (70B) and ≈ 1 step/sec for BLOOM-176B. This is up to 10x faster than offloading, enough for chatbots and other interactive apps. Parallel inference reaches hundreds of tokens/sec.
• Beyond classic language model APIs — you can employ any fine-tuning and sampling methods, execute custom paths through the model, or see its hidden states. You get the comforts of an API with the flexibility of PyTorch.

📚  See FAQ            📜  Read paper

Here's how to install Petals with Anaconda on Linux:

conda install pytorch pytorch-cuda=11.7 -c pytorch -c nvidia
pip install --upgrade petals

If you don't use Anaconda, you can install PyTorch in any other way. If you want to run models with 8-bit weights, please install PyTorch with CUDA 11.x or newer for compatility with bitsandbytes.

See the instructions for macOS and Windows, the full requirements, and troubleshooting advice in our FAQ.

The benchmarks below are for BLOOM-176B:

¹ An upper bound for offloading performance. We base our offloading numbers on the best possible hardware setup for offloading: CPU RAM offloading via PCIe 4.0 with 16 PCIe lanes per GPU and PCIe switches for pairs of GPUs. We assume zero latency for the upper bound estimation. In 8-bit, the model uses 1 GB of memory per billion parameters. PCIe 4.0 with 16 lanes has a throughput of 256 Gbit/s, so offloading 176B parameters takes 5.5 seconds. The throughput is twice as slow (128 Gbit/s) if we have two GPUs behind the same PCIe switch.

² A real-world distributed setting with 14 servers holding 2× RTX 3060, 4× 2080Ti, 2× 3090, 2× A4000, and 4× A5000 GPUs. These are personal servers and servers from university labs, spread across Europe and North America and connected to the Internet at speeds of 100–1000 Mbit/s. 4 servers operate from under firewalls.

³ An optimistic setup that requires least communication. The client nodes have 8 CPU cores and no GPU.

We provide more evaluations and discuss these results in more detail in Section 3.3 of our paper.

Please see our FAQ on contributing.

Alexander Borzunov, Dmitry Baranchuk, Tim Dettmers, Max Ryabinin, Younes Belkada, Artem Chumachenko, Pavel Samygin, and Colin Raffel. Petals: Collaborative Inference and Fine-tuning of Large Models. arXiv preprint arXiv:2209.01188, 2022.

@article{borzunov2022petals,
  title = {Petals: Collaborative Inference and Fine-tuning of Large Models},
  author = {Borzunov, Alexander and Baranchuk, Dmitry and Dettmers, Tim and Ryabinin, Max and Belkada, Younes and Chumachenko, Artem and Samygin, Pavel and Raffel, Colin},
  journal = {arXiv preprint arXiv:2209.01188},
  year = {2022},
  url = {https://arxiv.org/abs/2209.01188}
}

This project is a part of the BigScience research workshop.