Text Generation
Transformers
Safetensors
English
stripedhyena
custom_code
File size: 2,358 Bytes
2f9e2c6
 
06bc59c
 
2f9e2c6
06bc59c
d2ec434
06bc59c
cabde32
27d2bf6
cabde32
 
e4713f6
 
cabde32
e4713f6
00e8992
 
 
dce36cc
cabde32
06bc59c
 
 
af9abca
6be2da5
e4713f6
00e8992
e4713f6
00e8992
f6851b9
 
 
 
51d76d4
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
---
license: apache-2.0
language:
- en
---

## StripedHyena-Hessian-7B (SH 7B)

<p align="center">
  <img src="https://cdn-uploads.huggingface.co/production/uploads/62a1306bbe7fa896d2c8de44/Bfjh77emDsWOY-VmfvU9C.png" width="60%" />
</p>

### About 

One of the focus areas at Together Research is new architectures for long context, improved training, and inference performance over the Transformer architecture. Spinning out of a research program from our team and academic collaborators, with roots in **signal processing-inspired sequence models**, we are excited to introduce the **StripedHyena** models. StripedHyena is the **first alternative model competitive with the best open-source Transformers** of similar sizes in short and long-context evaluations.

**StripedHyena-Hessian-7B (SH 7B)** is our **base model** for this release.

- Read more here in [our blog](https://www.together.ai/blog/stripedhyena-7b).
- Play with the model on our [playground](https://api.together.xyz/playground/language/togethercomputer/StripedHyena-Hessian-7B)!
- Dive into the details of our [standalone implementation](https://github.com/togethercomputer/stripedhyena), and our related research: [1](https://arxiv.org/abs/2302.10866), [2](https://arxiv.org/abs/2310.18780), [3](https://arxiv.org/abs/2311.05908).

### Model Architecture

StripedHyena is a hybrid architecture composed of multi-head, grouped-query attention and gated convolutions arranged in [Hyena](https://arxiv.org/abs/2302.10866) blocks, different from traditional decoder-only Transformers.  
  - Costant memory decoding in Hyena blocks via representation of convolutions as state-space models (modal or canonical form), or as truncated filters.
  - Low latency, faster decoding and higher throughput than Transformers. 
  - Improvement to training and inference-optimal scaling laws, compared to optimized Transformer architectures such as Llama-2.
  - Trained on sequences of up to 32k, allowing it to process longer prompts.

### Note

To use StripedHyena outside of the playground, you will need to install custom kernels. Please follow the instructions from the [standalone repository](https://github.com/togethercomputer/stripedhyena).

StripedHyena is a mixed precision model. Make sure to keep your `poles` and `residues` in `float32` precision, especially for longer prompts or training.