transformer-tricks 0.1.13

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

Setup

To use the latest transformer-tricks python package:

pip3 install transformer-tricks

If you want to use the latest version of tricks.py, do this:

pip3 install --quiet -r requirements.txt

To run llama and other LLMs that need an agreement (not SmolLM), you first have to type the following:

huggingface-cli login

Above will ask you for the hf_token, which is the same you use e.g. in colab

Test FlashNorm

python3 test_flashNorm.py

Above should return the following:

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
ppl: tensor(16.0831)
ppl: tensor(16.0831)
ppl: tensor(12.0864)
ppl: tensor(12.0864)

Use the transformer-tricks package

import transformer_tricks as tt

Example

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

import transformer_tricks as tt

# convert model to flashNorm
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')

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

Above should return the following:

Once upon a time there was a curious little girl
Once upon a time there was a curious little girl
ppl: tensor(16.0831)
ppl: tensor(16.0831)

Notes on python package

• Link to package here
• Link to stats here