- Petals runs large language models like [BLOOM-176B](https://huggingface.co/bigscience/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.
- Inference runs at ≈ 1 sec per step (token) — 10x faster than possible with 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 by executing custom paths through the model or accessing its hidden states. You get the comforts of an API with the flexibility of PyTorch.
- Single-batch inference runs at ≈ 1 sec per step (token) — [up to 10x faster](https://github.com/bigscience-workshop/petals#benchmarks) than offloading, enough for [chatbots](http://chat.petals.ml) 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.
This script uses Anaconda to install CUDA-enabled PyTorch.
If you don't have anaconda, you can get it from [here](https://www.anaconda.com/products/distribution).
If you don't want anaconda, you can install PyTorch [any other way](https://pytorch.org/get-started/locally/).
If you want to run models with 8-bit weights, please install **PyTorch with CUDA 11** or newer for compatility with [bitsandbytes](https://github.com/timDettmers/bitsandbytes).
__System requirements:__ Petals only supports Linux for now. If you don't have a Linux machine, consider running Petals in Docker (see our [image](https://hub.docker.com/r/learningathome/petals)) or, in case of Windows, in WSL2 ([read more](https://learn.microsoft.com/en-us/windows/ai/directml/gpu-cuda-in-wsl)). CPU is enough to run a client, but you probably need a GPU to run a server efficiently.
## 🛠️ Development
Petals uses pytest with a few plugins. To install them, run:
git clone https://github.com/bigscience-workshop/petals.git && cd petals
pip install -e .[dev]
```
To run minimalistic tests, you need to make a local swarm with a small model and some servers. You may find more information about how local swarms work and how to run them in [this tutorial](https://github.com/bigscience-workshop/petals/wiki/Launch-your-own-swarm).
After you're done, you can terminate the servers and ensure that no zombie processes are left with `pkill -f petals.cli.run_server && pkill -f p2p`.
The automated tests use a more complex server configuration that can be found [here](https://github.com/bigscience-workshop/petals/blob/main/.github/workflows/run-tests.yaml).
### Code style
We use [black](https://black.readthedocs.io/en/stable/the_black_code_style/current_style.html) and [isort](https://pycqa.github.io/isort/) for all pull requests.
Before committing your code, simply run `black . && isort .` and you will be fine.
If you don't use Anaconda, you can install PyTorch in [any other way](https://pytorch.org/get-started/locally/). If you want to run models with 8-bit weights, please install PyTorch with CUDA 11.x or newer for compatility with [bitsandbytes](https://github.com/timDettmers/bitsandbytes).
See the instructions for macOS and Windows, the full requirements, and troubleshooting advice in our [FAQ](https://github.com/bigscience-workshop/petals/wiki/FAQ:-Frequently-asked-questions#running-a-client).
<thcolspan="6">Offloading, max. possible speed on 1x A100 <sup>1</sup></th>
</tr>
<tralign="center">
<td>256 Gbit/s</td>
<td></td>
<td>0.18</td>
<td>0.18</td>
<td>2.7</td>
<td>170.3</td>
</tr>
<tralign="center">
<td>128 Gbit/s</td>
<td></td>
<td>0.09</td>
<td>0.09</td>
<td>2.4</td>
<td>152.8</td>
</tr>
<tr>
<thcolspan="6">Petals on 14 heterogeneous servers across Europe and North America <sup>2</sup></th>
</tr>
<tralign="center">
<tdcolspan="2">Real world</td>
<td>0.83</td>
<td>0.79</td>
<td>32.6</td>
<td>179.4</td>
</tr>
<tr>
<thcolspan="6">Petals on 3 servers, with one A100 each <sup>3</sup></th>
</tr>
<tralign="center">
<td>1 Gbit/s</td>
<td>< 5 ms</td>
<td>1.71</td>
<td>1.54</td>
<td>70.0</td>
<td>253.6</td>
</tr>
<tralign="center">
<td>100 Mbit/s</td>
<td>< 5 ms</td>
<td>1.66</td>
<td>1.49</td>
<td>56.4</td>
<td>182.0</td>
</tr>
<tralign="center">
<td>100 Mbit/s</td>
<td>100 ms</td>
<td>1.23</td>
<td>1.11</td>
<td>19.7</td>
<td>112.2</td>
</tr>
</table>
<sup>1</sup>**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.
<sup>2</sup>**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.
<sup>3</sup>**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](https://arxiv.org/pdf/2209.01188.pdf).
## 🛠️ Contributing
Please see our [FAQ](https://github.com/bigscience-workshop/petals/wiki/FAQ:-Frequently-asked-questions#contributing) on contributing.