transformer-tricks 0.2.3

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

Setup

pip3 install transformer-tricks

Example

The example below converts SmolLM-135M to FlashNorm and measures perplexity of the original and the modified model.

import transformer_tricks as tt

# convert model and store the new model in ./SmolLM-135M_flashNorm_test
tt.flashify_repo('HuggingFaceTB/SmolLM-135M')

# run example inference of original and modified model
tt.hello_world('HuggingFaceTB/SmolLM-135M')
tt.hello_world('./SmolLM-135M_flashNorm_test')

# measure perplexity of original and modified model
tt.perplexity('HuggingFaceTB/SmolLM-135M', speedup=16)
tt.perplexity('./SmolLM-135M_flashNorm_test', speedup=16)

Results:

Once upon a time there was a curious little girl
Once upon a time there was a curious little girl
perplexity = 16.083
perplexity = 16.083

You can run the example in your browser by clicking on this notebook: . Hit "cancel" when it says "Notebook does not have secret access", because we don't need an HF_TOKEN for SmolLM.

Test FlashNorm

# setup
git clone https://github.com/OpenMachine-ai/transformer-tricks.git
cd python
pip3 install --quiet -r requirements.txt

# run tests
python3 flashNorm_test.py

Results:

Once upon a time there was a curious little girl
Once upon a time there was a curious little girl
Once upon a time there was a little girl named
Once upon a time there was a little girl named
perplexity = 16.083
perplexity = 16.083
perplexity = 12.086
perplexity = 12.086

To run llama and other LLMs that need an agreement (not SmolLM), you first have to type the following, which will ask for your hf_token:

huggingface-cli login

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
• Whenever you modify flashNorm_example.py, generate the corresponding notebook as follows:

  jupytext --to ipynb flashNorm_example.py -o ../notebooks/flashNorm_example.ipynb
  

Notes on python package

• Link to package here
• Link to stats here
• Source of this README file here

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