Skip to content

Commit

Permalink
Update README.md (#520)
Browse files Browse the repository at this point in the history
  • Loading branch information
@borzunov
borzunov authored Sep 22, 2023
1 parent a2484b3 commit 1d9401d
Showing 1 changed file with 12 additions and 97 deletions.
109 changes: 12 additions & 97 deletions README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -8,14 +8,14 @@
<br>
</p>

Generate text with distributed **Llama 2 (70B)**, **Stable Beluga 2**, **Falcon**, **Guanaco-65B** or **BLOOM-176B** and fine‑tune them for your own tasks &mdash; right from your desktop computer or Google Colab:
Generate text with distributed **Llama 2** (70B), **Falcon** (40B+), **BLOOM** (176B) (or their derivatives), and fine‑tune them for your own tasks &mdash; right from your desktop computer or Google Colab:

```python
from transformers import AutoTokenizer
from petals import AutoDistributedModelForCausalLM

# Choose any model available at https://health.petals.dev
model_name = "petals-team/StableBeluga2"
model_name = "petals-team/StableBeluga2" # This one is fine-tuned Llama 2 (70B)

# Connect to a distributed network hosting model layers
tokenizer = AutoTokenizer.from_pretrained(model_name)
Expand All @@ -31,9 +31,9 @@ print(tokenizer.decode(outputs[0])) # A cat sat on a mat...
🚀 &nbsp;<b><a href="https://colab.research.google.com/drive/1uCphNY7gfAUkdDrTx21dZZwCOUDCMPw8?usp=sharing">Try now in Colab</a></b>
</p>

🦙 **Want to run Llama 2?** Request access to its weights at the ♾️ [Meta AI website](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) and 🤗 [Model Hub](https://huggingface.co/meta-llama/Llama-2-70b-hf), then run `huggingface-cli login` in the terminal before loading the model. Or just try it in our [chatbot app](https://chat.petals.dev).
🔏 **Privacy.** Your data will be processed with the help of other people in the public swarm. Learn more about privacy [here](https://github.com/bigscience-workshop/petals/wiki/Security,-privacy,-and-AI-safety). For sensitive data, you can set up a [private swarm](https://github.com/bigscience-workshop/petals/wiki/Launch-your-own-swarm) among people you trust.

🔏 **Privacy.** Your data will be processed by other people in the public swarm. Learn more about privacy [here](https://github.com/bigscience-workshop/petals/wiki/Security,-privacy,-and-AI-safety). For sensitive data, you can set up a [private swarm](https://github.com/bigscience-workshop/petals/wiki/Launch-your-own-swarm) among people you trust.
🦙 **Want to run Llama 2?** Request access to its weights at the ♾️ [Meta AI website](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) and 🤗 [Model Hub](https://huggingface.co/meta-llama/Llama-2-70b-hf), then run `huggingface-cli login` in the terminal before loading the model. Or just try it in our [chatbot app](https://chat.petals.dev).

💬 **Any questions?** Ping us in [our Discord](https://discord.gg/KdThf2bWVU)!

Expand Down Expand Up @@ -81,9 +81,8 @@ python3 -m petals.cli.run_server petals-team/StableBeluga2

## How does it work?

- Petals runs large language models like [Llama](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) and [BLOOM](https://huggingface.co/bigscience/bloom) **collaboratively** — you load a small part of the model, then join 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 &approx; 1 step/sec for BLOOM-176B. This is [up to 10x faster](https://github.com/bigscience-workshop/petals#benchmarks) than offloading, enough to build [chatbots](https://chat.petals.dev) 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.
- You load a small part of the model, then join a [network](https://health.petals.dev) of people serving the other parts. Single‑batch inference runs at up to **6 tokens/sec** for **Llama 2** (70B) and up to **4 tokens/sec** for **Falcon** (180B) — enough for [chatbots](https://chat.petals.dev) and interactive apps.
- 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** and **🤗 Transformers**.

<p align="center">
<img src="https://i.imgur.com/RTYF3yW.png" width="800">
Expand Down Expand Up @@ -113,99 +112,15 @@ Advanced guides:
- Launch a private swarm: [guide](https://github.com/bigscience-workshop/petals/wiki/Launch-your-own-swarm)
- Run a custom model: [guide](https://github.com/bigscience-workshop/petals/wiki/Run-a-custom-model-with-Petals)

## Benchmarks

The benchmarks below are for BLOOM-176B:

<table align="center">
<tr>
<th colspan="2">Network</th>
<th colspan="2">Single-batch inference<br>(steps/s)</th>
<th colspan="2">Parallel forward<br>(tokens/s)</th>
</tr>
<tr>
<th rowspan="2">Bandwidth</th>
<th rowspan="2">Round-trip<br>latency</th>
<th colspan="2">Sequence length</th>
<th colspan="2">Batch size</th>
</tr>
<tr align="center">
<td>128</td>
<td>2048</td>
<td>1</td>
<td>64</td>
</tr>
<tr>
<th colspan="6">Offloading, max. possible speed on 1x A100 <sup>1</sup></th>
</tr>
<tr align="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>
<tr align="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>
<th colspan="6">Petals on 14 heterogeneous servers across Europe and North America <sup>2</sup></th>
</tr>
<tr align="center">
<td colspan="2">Real world</td>
<td>0.83</td>
<td>0.79</td>
<td>32.6</td>
<td>179.4</td>
</tr>
<tr>
<th colspan="6">Petals on 3 servers, with one A100 each <sup>3</sup></th>
</tr>
<tr align="center">
<td>1 Gbit/s</td>
<td>&lt; 5 ms</td>
<td>1.71</td>
<td>1.54</td>
<td>70.0</td>
<td>253.6</td>
</tr>
<tr align="center">
<td>100 Mbit/s</td>
<td>&lt; 5 ms</td>
<td>1.66</td>
<td>1.49</td>
<td>56.4</td>
<td>182.0</td>
</tr>
<tr align="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
### Benchmarks

Please see **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.

## 📜 Citation
### 📜 Citation

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.](https://arxiv.org/abs/2209.01188)
Expand Down

0 comments on commit 1d9401d

Please sign in to comment.