Causal Transformer Decoder

This repository contains the code for the causal transformer decoder, which is the autoregressive version of the Pytorch TransformerDecoder.

Explanation of what it does can be found here: https://scale.com/blog/pytorch-improvements

To install the python package directly:

pip install git+https://github.com/alexmt-scale/causal-transformer-decoder.git

To test the consistency of the implementation:

The Causal Transformer Decoder is supposed to return the same output as the Pytorch TransformerDecoder when generating sentences, provided the input is the same.

python -m pytest

To run benchmarks against Pytorch implementations:

python -m benchmarks.benchmark_transformers

And against HuggingFace GPT-2 (our implementation should be as fast as theirs):

python -m benchmarks.gpt_generation_benchmark

License

The source code for the site is licensed under the MIT license, which you can find in the LICENSE file.