A collection of tricks to simplify and speed up transformer models:
- Slim attention: [paper], [video], [podcast], [notebook], [code-readme]
- 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, NoPE’s removal of positional encoding, 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.
Install the transformer tricks package:
pip install transformer-tricksFollow the links below for documentation of the python code in this directory:
The papers are accompanied by the following Jupyter notebooks:
- Slim attention:
- Flash normalization:
- Removing weights from skipless transformers:
Please subsribe to our [newsletter] on substack to get the latest news about this project. We will never send you more than one email per month.
Before making a change to this repo, please do the following:
- Format your code by typing
autopep8 *.py. It's using the config inpyproject.toml. - Whenever you change
transformer_tricks.py, publish a new version of the package as follows:- First, update the version number in
pyproject.tomland inrequirements.txt - Then, push the package to PyPi by typing
./push_pypi.sh - Links for python package: pypi, stats, source of this readme
- First, update the version number in
- Whenever you modify
flashNorm_example.pyor another python file, generate the corresponding notebook as follows:jupytext --to ipynb flashNorm_example.py -o notebooks/flashNorm_example.ipynb