transformer-tricks 0.3.2

A collection of tricks to speed up LLMs, see our transformer-tricks papers on arXiv

Transformer Tricks

A collection of tricks to simplify and speed up transformer models:

• Slim attention: [paper], [video], [podcast], [notebook], [code-readme]
• Matrix-shrink [work in progress]: [paper]
• Flash normalization: [paper], [podcast], [notebook], [code-readme]
• Precomputing the first layer: [paper], [podcast]
• Removing weights from skipless transformers: [paper], [podcast], [notebook]

Many of these tricks follow a recent trend of removing parts from neural networks such as RMSNorm’s removal of mean centering from LayerNorm, PaLM's removal of bias-parameters, GPT’s removal of the encoder stack, and of course transformer’s revolutionary removal of recurrent layers. Specifically, our FlashNorm removes the weights from RMSNorm and merges them with the next linear layer. And slim attention removes the entire V-cache from the context memory for MHA transformers.

---

Installation

Install the transformer tricks package:

pip install transformer-tricks

---

Documentation

Follow the links below for documentation of the python code in this directory:

• Slim attention
• Flash normalization

---

Notebooks

The papers are accompanied by the following Jupyter notebooks:

• Slim attention:
• Flash normalization:
• Removing weights from skipless transformers:

---

Newsletter

Please subscribe to our [newsletter] on substack to get the latest news about this project. We will never send you more than one email per month.

---

Contributing

Before making a change to this repo, please do the following:

• Format your code by typing autopep8 *.py. It's using the config in pyproject.toml
• Whenever you change transformer_tricks.py, publish a new version of the package as follows:
  • First, update the version number in pyproject.toml and in requirements.txt
  • Then, push the package to PyPi by typing ./push_pypi.sh
  • Links for python package: pypi, stats, source of this readme
• Generate notebooks from python by typing util/gen_notebooks

---

Please give us a ⭐ if you like this repo, thanks!

---