LaMDA-pytorch

Open-source pre-training implementation of Google's LaMDA research paper in PyTorch. The totally not sentient AI. This repository will cover the 2B parameter implementation of the pre-training architecture as that is likely what most can afford to train. You can review Google's latest blog post from 2022 which details LaMDA here. You can also view their previous blog post from 2021 on the model here.

Acknowledgement:

I have been greatly inspired by the work of Dr. Phil 'Lucid' Wang. Please check out his open-source implementations of multiple different transformer architectures and support his work.

Developer Updates

Developer updates can be found on:

• https://twitter.com/EnricoShippole
• https://www.linkedin.com/in/enrico-shippole-495521b8/

Basic Usage - Pre-training

lamda_base = LaMDA(
    num_tokens = 20000,
    dim = 512,
    dim_head = 64,
    depth = 12,
    heads = 8
)

lamda = AutoregressiveWrapper(lamda_base, max_seq_len = 512)

tokens = torch.randint(0, 20000, (1, 512)) # mock token data

logits = lamda(tokens)

print(logits)

Notes on training at scale:

• There may be issues with NaN for fp16 training.
• Pipeline parallelism should be used with ZeRO 1, not ZeRO 2.

About LaMDA:

• T5 Relative Positional Bias in Attention
• Gated GELU Activation in the Feed forward layer
• GPT-like Decoder Only architecture
• Autoregressive with Top-k sampling
• Sentencepiece Byte-pair encoded tokenizer

To Do list

• Make To Do list (we are now officially more organized than Character.AI devs)
• Add inference only if someone wants to open-source LaMDA model weights
• Add finetuning script
• Add Sentencepiece tokenizer training script and integration
• Find ways, modify architecture if necessary, to reduce memory usage
  • Experiment with quantization, to reduce required memory? (Update: CPU only for now.)
  • Vocabulary size results in big embedding = more required memory and inference time. Experiment with embedding.
  • Large sequence lengths can absolutely destroy memory. Find a way to have large sequence lengths without taking up so much memory.
• Experiment with improvements to the model:
  • Try to see if Mixture of Denoisers can be added to the model/training process
  • Parametrized GEGLU or perhaps other activation functions (Squared ReLU shows promising improvement for now, may revisit later)
• Redo config system to require less hardcoding and more flexibility (Many PCs have too little disk space to fully load The Pile, need to get dataset streaming somehow working with this whole process)
  • Switch to .yaml based config system (maybe?)
  • Attempt to have it to where data can be streamed in instead of having to download the entire dataset at once.
• ???
• Our wAIfus are real again!

LEGACY TODO (>implying the previous guy is ever gonna finish this):

• Finish building pre-training model architecture
• Add pre-training script
• Integrate Huggingface datasets
• Use The Pile from Eleuther AI.
• Build the GODEL dataset and upload to HuggingFace datasets
• Implement GPT-2 tokenizer
• Add Sentencepiece tokenizer training script and integration
• Add detailed documentation
• Add logging with Weights And Biases
• Add scaling with ColossalAI.
• Add finetuning script
• Add pip installer with PyPI
• Implement a JAX / Flax version as well
• Add inference only if someone wants to open-source LaMDA model weights

Author

• Enrico Shippole

Citations

@article{DBLP:journals/corr/abs-2201-08239,
  author    = {Romal Thoppilan and
               Daniel De Freitas and
               Jamie Hall and
               Noam Shazeer and
               Apoorv Kulshreshtha and
               Heng{-}Tze Cheng and
               Alicia Jin and
               Taylor Bos and
               Leslie Baker and
               Yu Du and
               YaGuang Li and
               Hongrae Lee and
               Huaixiu Steven Zheng and
               Amin Ghafouri and
               Marcelo Menegali and
               Yanping Huang and
               Maxim Krikun and
               Dmitry Lepikhin and
               James Qin and
               Dehao Chen and
               Yuanzhong Xu and
               Zhifeng Chen and
               Adam Roberts and
               Maarten Bosma and
               Yanqi Zhou and
               Chung{-}Ching Chang and
               Igor Krivokon and
               Will Rusch and
               Marc Pickett and
               Kathleen S. Meier{-}Hellstern and
               Meredith Ringel Morris and
               Tulsee Doshi and
               Renelito Delos Santos and
               Toju Duke and
               Johnny Soraker and
               Ben Zevenbergen and
               Vinodkumar Prabhakaran and
               Mark Diaz and
               Ben Hutchinson and
               Kristen Olson and
               Alejandra Molina and
               Erin Hoffman{-}John and
               Josh Lee and
               Lora Aroyo and
               Ravi Rajakumar and
               Alena Butryna and
               Matthew Lamm and
               Viktoriya Kuzmina and
               Joe Fenton and
               Aaron Cohen and
               Rachel Bernstein and
               Ray Kurzweil and
               Blaise Aguera{-}Arcas and
               Claire Cui and
               Marian Croak and
               Ed H. Chi and
               Quoc Le},
  title     = {LaMDA: Language Models for Dialog Applications},
  journal   = {CoRR},
  volume    = {abs/2201.08239},
  year      = {2022},
  url       = {https://arxiv.org/abs/2201.08239},
  eprinttype = {arXiv},
  eprint    = {2201.08239},
  timestamp = {Fri, 22 Apr 2022 16:06:31 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2201-08239.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

@misc{https://doi.org/10.48550/arxiv.1706.03762,
  doi = {10.48550/ARXIV.1706.03762},
  
  url = {https://arxiv.org/abs/1706.03762},
  
  author = {Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N. and Kaiser, Lukasz and Polosukhin, Illia},
  
  keywords = {Computation and Language (cs.CL), Machine Learning (cs.LG), FOS: Computer and information sciences, FOS: Computer and information sciences},
  
  title = {Attention Is All You Need},
  
  publisher = {arXiv},
  
  year = {2017},
  
  copyright = {arXiv.org perpetual, non-exclusive license}
}

@misc{https://doi.org/10.48550/arxiv.1910.10683,
  doi = {10.48550/ARXIV.1910.10683},
  
  url = {https://arxiv.org/abs/1910.10683},
  
  author = {Raffel, Colin and Shazeer, Noam and Roberts, Adam and Lee, Katherine and Narang, Sharan and Matena, Michael and Zhou, Yanqi and Li, Wei and Liu, Peter J.},
  
  keywords = {Machine Learning (cs.LG), Computation and Language (cs.CL), Machine Learning (stat.ML), FOS: Computer and information sciences, FOS: Computer and information sciences},
  
  title = {Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer},
  
  publisher = {arXiv},
  
  year = {2019},
  
  copyright = {arXiv.org perpetual, non-exclusive license}
}

@misc{https://doi.org/10.48550/arxiv.2002.05202,
  doi = {10.48550/ARXIV.2002.05202},
  
  url = {https://arxiv.org/abs/2002.05202},
  
  author = {Shazeer, Noam},
  
  keywords = {Machine Learning (cs.LG), Neural and Evolutionary Computing (cs.NE), Machine Learning (stat.ML), FOS: Computer and information sciences, FOS: Computer and information sciences},
  
  title = {GLU Variants Improve Transformer},
  
  publisher = {arXiv},
  
  year = {2020},
  
  copyright = {arXiv.org perpetual, non-exclusive license}
}

@article{DBLP:journals/corr/abs-2101-00027,
  author    = {Leo Gao and
               Stella Biderman and
               Sid Black and
               Laurence Golding and
               Travis Hoppe and
               Charles Foster and
               Jason Phang and
               Horace He and
               Anish Thite and
               Noa Nabeshima and
               Shawn Presser and
               Connor Leahy},
  title     = {The Pile: An 800GB Dataset of Diverse Text for Language Modeling},
  journal   = {CoRR},
  volume    = {abs/2101.00027},
  year      = {2021},
  url       = {https://arxiv.org/abs/2101.00027},
  eprinttype = {arXiv},
  eprint    = {2101.00027},
  timestamp = {Thu, 14 Oct 2021 09:16:12 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2101-00027.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

@article{DBLP:journals/corr/abs-1808-06226,
  author    = {Taku Kudo and
               John Richardson},
  title     = {SentencePiece: {A} simple and language independent subword tokenizer
               and detokenizer for Neural Text Processing},
  journal   = {CoRR},
  volume    = {abs/1808.06226},
  year      = {2018},
  url       = {http://arxiv.org/abs/1808.06226},
  eprinttype = {arXiv},
  eprint    = {1808.06226},
  timestamp = {Sun, 02 Sep 2018 15:01:56 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/abs-1808-06226.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

@inproceedings{sennrich-etal-2016-neural,
    title = "Neural Machine Translation of Rare Words with Subword Units",
    author = "Sennrich, Rico  and
      Haddow, Barry  and
      Birch, Alexandra",
    booktitle = "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = aug,
    year = "2016",
    address = "Berlin, Germany",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/P16-1162",
    doi = "10.18653/v1/P16-1162",
    pages = "1715--1725",
}