Home

Awesome

StripedHyena

Minimal implementation of a StripedHyena model.

<p align="center"> <img src="https://github.com/togethercomputer/stripedhyena/assets/34561392/ce11c785-c8c5-43b2-a34a-41ee9cba3958" width="60%" /> </p>

About

StripedHyena is the first alternative model architecture competitive with the best open-source Transformers of similar sizes in short and long-context evaluations.

StripedHyena is a deep signal processing, hybrid architecture composed of rotary (grouped) attention and gated convolutions arranged in Hyena blocks, with improved scaling over decoder-only Transformers. StripedHyena is designed to leverage the specialization of each of its layer classes, with Hyena layers implementing the bulk of the computation required for sequence processing and attention layers supplementing the ability to perform targeted pattern recall.

Models

Biology: Evo-1-7B

Evo is a biological foundation model capable of long-context modeling and design.

Evo uses the StripedHyena architecture to enable modeling of sequences at a single-nucleotide, byte-level resolution with near-linear scaling of compute and memory relative to context length. Evo has 7 billion parameters and is trained on OpenGenome, a prokaryotic whole-genome dataset containing ~300 billion tokens.

Language: StripedHyena-7B

StripedHyena-Nous-7B (SH-N 7B) is our chat model for this release, and was developed with our collaborators at Nous Research.

SH-N 7B uses this prompt format: ### Instruction:\n{prompt}\n\n### Response:\n{response}

Quick Start

The most direct way to test StripedHyena models is via our playground, which includes a variety of architecture-specific optimizations.

Playground:

Standalone

Checkpoints

We provide a checkpoint for StripedHyena-Hessian 7B, our base model. Download pytorch-model.bin from the HuggingFace repository. As an alternative, we also provide HuggingFace compatible checkpoints for AutoClasses.

Environment Setup

To run our standalone StripedHyena implementation, you will need to install the packages in requirements.txt, as well as rotary and normalization kernels from flash_attn.

The easiest way to ensure all requirements are installed is to build a Docker image using Dockerfile, or follow the steps detailed in the Dockerfile itself in a different virtual environment. For example, to build a Docker image, run:

docker build --tag sh:test .

Installing the dependencies and kernels could take several minutes. Then run the container interactively with:

docker run -it --gpus all --network="host" --shm-size 900G -v=<path_to_this_repo>:/mnt:rw --rm sh:test

Environment Setup

Once the environment is set up, you will be able to generate text with:

python generate.py --config_path ./configs/7b-sh-32k-v1.yml \
--checkpoint_path <path_to_ckpt> --cached_generation \
--prompt_file ./test_prompt.txt

If you are generating with prompt.txt, set prefill_style: fft in the config. For very long prompts, you may want to opt for prefill_style: recurrence, which will be slower but use less memory.

If the installation was correct, test prompt will generate the following paragraph

The four species of hyenas are the striped hyena (Hyaena hyaena), the brown hyena (Parahyaena brunnea), the spotted hyena (Crocuta crocuta), and the aardwolf (Proteles cristata).\n\nThe striped hyena is the most widespread species, occurring in Africa, the Middle East, and Asia.

HuggingFace

We also provide an entry script to generate with StripedHyena models hosted on HuggingFace. The model ids are:

Choose your model id, then run the following command:

python generate_transformers.py --model-name <model_id> --input-file ./test_prompt.txt

Testing Correctness

We report lm-evaluation-harness (10-shot) scores to use as a proxy for (standalone) model correctness in your environment.

More extensive benchmarks results are provided in the blog post and on HuggingFace.

Optional Dependencies

The standalone implementation provides integration with some custom kernels for StripedHyena such as FlashFFTConv (see the model config 7b-sh-32k-v1.yml for more information). These additional kernels are not required to run the model.

Issues

Several issues can be resolved by reinstalling the latest version of flash_attn (pip freeze | grep flash-attn should return a version >= 2.0.0).

StripedHyena is a mixed precision model. Make sure to keep your poles and residues in float32 precision.

Cite

If have found the pretrained models or architecture useful for you research or application, consider citing:

@software{stripedhyena,
  title        = {{StripedHyena: Moving Beyond Transformers with Hybrid Signal Processing Models}},
  author       = { Poli, Michael and Wang, Jue and Massaroli, Stefano and Quesnelle, Jeffrey and Carlow, Ryan and Nguyen, Eric and Thomas, Armin},
  month        = 12,
  year         = 2023,
  url          = { https://github.com/togethercomputer/stripedhyena },
  doi          = { 10.57967/hf/1595 },
}