Model parallel transformers in JAX and Haiku

Overview

Table of contents

  1. Mesh Transformer JAX
    1. Updates
  2. Pretrained Models
    1. GPT-J-6B
      1. Links
      2. Acknowledgments
      3. License
      4. Model Details
      5. Zero-Shot Evaluations
  3. Architecture and Usage
    1. Fine-tuning
    2. JAX Dependency
  4. TODO

Mesh Transformer JAX

A haiku library using the xmap/pjit operators in JAX for model parallelism of transformers.

The parallelism scheme is similar to the original Megatron-LM, which is efficient on TPUs due to the high speed 2d mesh network. There is also an experimental model version which implements ZeRo style sharding.

This library is designed for scalability up to approximately 40B parameters on TPUv3s, beyond which different parallelism strategies should be used. See other implementations such as GPT-NeoX or DeepSpeed for that.

One future direction for research is integrating this codebase with swarm-jax, to achieve further scalability with pipeline parallelism.

Updates

12-07-21: Added guide to fine tuning

Pretrained Models

GPT-J-6B

A 6 billion parameter, autoregressive text generation model trained on The Pile.

Links

Slim weights (bf16 weights only, for inference, 9GB)

Full weights (including optimizer params, 61GB)

Colab demo

Web demo

Aran's blog post

Acknowledgments

This project would not have been possible without compute generously provided by the TPU Research Cloud with assistance from EleutherAI.

Thanks to the Cloud TPU team at Google for providing early access to the Cloud TPU VM alpha (now publicly available!)

Thanks to everyone who have helped out one way or another (listed alphabetically):

  • Aran Komatsuzaki for advice with experiment design and writing the blog posts.
  • James Bradbury for valuable assistance with debugging JAX issues.
  • Janko Prester for creating the web demo frontend.
  • Laurence Golding for adding some features to the web demo.
  • Leo Gao for running zero shot evaluations for the baseline models for the table.

License

The weights of GPT-J-6B are licensed under version 2.0 of the Apache License.

Model Details

Hyperparameter Value
n_parameters 6,053,381,344
n_layers 28*
d_model 4,096
d_ff 16,384
n_heads 16
d_head 256
n_ctx 2,048
n_vocab 50,257 (same tokenizer as GPT-2/3)
position encoding Rotary position encodings (RoPE)
RoPE dimensions 64

* each layer consists of one feedforward block and one self attention block

The model consists of 28 layers with a model dimension of 4096, and a feedforward dimension of 16384. The model dimension is split into 16 heads, each with a dimension of 256. Rotary position encodings (RoPE) was applied to 64 dimensions of each head. The model is trained with a tokenization vocabulary of 50257, using the same set of BPEs as GPT-2/GPT-3.

Zero-Shot Evaluations

Models roughly sorted by performance, or by FLOPs if not available.

Model Weights Training FLOPs LAMBADA PPL ↓ LAMBADA Acc ↑ Winogrande ↑ Hellaswag ↑ PIQA ↑ Dataset Size (GB)
Chance 0 ~a lot ~0% 50% 25% 25% 0
GPT-3-Ada‡ ----- 9.95 51.6% 52.9% 43.4% 70.5% -----
GPT-2-1.5B ----- 10.63 51.21% 59.4% 50.9% 70.8% 40
GPTNeo-1.3B‡ 3.0e21 7.50 57.2% 55.0% 48.9% 71.1% 825
Megatron-2.5B* 2.4e21 ----- 61.7% ----- ----- ----- 174
GPTNeo-2.7B‡ 6.8e21 5.63 62.2% 56.5% 55.8% 73.0% 825
GPT-3-1.3B*‡ 2.4e21 5.44 63.6% 58.7% 54.7% 75.1% ~800
GPT-3-Babbage‡ ----- 5.58 62.4% 59.0% 54.5% 75.5% -----
Megatron-8.3B* 7.8e21 ----- 66.5% ----- ----- ----- 174
GPT-3-2.7B*‡ 4.8e21 4.60 67.1% 62.3% 62.8% 75.6% ~800
Megatron-11B† 1.0e22 ----- ----- ----- ----- ----- 161
GPT-J-6B 1.5e22 3.99 69.7% 65.3% 66.1% 76.5% 825
GPT-3-6.7B*‡ 1.2e22 4.00 70.3% 64.5% 67.4% 78.0% ~800
GPT-3-Curie‡ ----- 4.00 69.3% 65.6% 68.5% 77.9% -----
GPT-3-13B*‡ 2.3e22 3.56 72.5% 67.9% 70.9% 78.5% ~800
GPT-3-175B*‡ 3.1e23 3.00 76.2% 70.2% 78.9% 81.0% ~800
GPT-3-Davinci‡ ----- 3.0 75% 72% 78% 80% -----
Gopher 230B* 6.31E+23 ----- 74.50% 70.10% 79.20% 81.80% 1344
MT-NLG 530B*‡ ----- ----- 76.6% 73.0% 80.2% 82.0% -----

* represents evaluation numbers reported by their respective authors, all other numbers are provided by running the lm-evaluation-harness either with the released weights or with API access. Due to subtle implementation differences as well as different zero shot task framing, these might not be directly comparable. See this blog post for more details.

The Megatron-11B model provides no comparable metrics, and several implementations using the released weights do not reproduce the generation quality and evaluations. (see 1 2 3) Thus, evaluation was not attempted.

These models have been trained with data which contains possible test set contamination. The OpenAI GPT-3 models failed to deduplicate training data for certain test sets, while the GPT-Neo models as well as this one is trained on The Pile, which has not been deduplicated against any test sets.

Architecture and Usage

Most scripts in this repository are designed to be run on TPUs, which under the TPU-VM architecture are virtual machines which can run arbitrary code. Most scripts are designed to spin up a TPU, SSH into it to set up the dependencies and copy code over from the local directory, and then start a Ray worker which can accept RPC calls.

The TPUVMs handles running model training steps and evaluation, checkpoint save and loading, while the driver python program handles data loading and general orchestration (such as when to save checkpoints etc).

This means that most scripts (train.py, eval_harness.py etc) expect to be running on a GCE virtual machine in the same region as the TPUs, to minimize RPC latency and data transfer cost. Other scripts (usually ones which don't take a --tpu argument, such as device_sample.py, device_serve.py or device_train.py) expect to be run directly on a TPUVM. The device_* scripts only work on a v3-8 and not on larger pods.

Furthermore, there is an example (resharding_example.py) of how to convert the provided checkpoints (which have 8 shards in the case of GPT-J-6B) down to a smaller number, such as for when running on GPU(s).

Fine-tuning

To fine-tune the model, run device_train.py on a TPU VM. Using a TPU v3-8, you can fine-tune at a rate of ~5000 tokens/second, which should be sufficient for small-to-medium-size datasets.

Please read the step by step guide for thorough fine-tuning instructions.

JAX Dependency

Note this library has some specific requirements for JAX version. Specifically, to use the v1 models (including GPT-J 6B), jax==0.2.12 is required. This in turn depends on jaxlib==0.1.68. If this is not done, you will get cryptic xmap errors

However, to use the v2 model code (no publicly released weights), the newest JAX version can be used.

Citation

To cite this repository:

@misc{mesh-transformer-jax,
  author = {Wang, Ben},
  title = {{Mesh-Transformer-JAX: Model-Parallel Implementation of Transformer Language Model with JAX}},
  howpublished = {\url{https://github.com/kingoflolz/mesh-transformer-jax}},
  year = 2021,
  month = May
}

To cite the weights of GPT-J-6B:

@misc{gpt-j,
  author = {Wang, Ben and Komatsuzaki, Aran},
  title = {{GPT-J-6B: A 6 Billion Parameter Autoregressive Language Model}},
  howpublished = {\url{https://github.com/kingoflolz/mesh-transformer-jax}},
  year = 2021,
  month = May
}

If you use this repository or any of the pretrained weights to do something cool, we would love to hear about it. Feel free to open a github issue or reach out over email (in profile).

TODO

  • disentangle heads and shards
  • test/benchmark on TPU
  • implement gradient checkpointing
  • fix initialization
  • mixed precision
  • deal with preemptible TPUs
  • test and validate generation
  • shard activations instead of replicating for memory efficiency (in v2)
  • support ZeRO style sharding (in v2)
Owner
Ben Wang
Ben Wang
Kestrel Threat Hunting Language

Kestrel Threat Hunting Language What is Kestrel? Why we need it? How to hunt with XDR support? What is the science behind it? You can find all the ans

Open Cybersecurity Alliance 201 Dec 16, 2022
Legal text retrieval for python

legal-text-retrieval Overview This system contains 2 steps: generate training data containing negative sample found by mixture score of cosine(tfidf)

Nguyễn Minh Phương 22 Dec 06, 2022
PyTorch Implementation of Meta-StyleSpeech : Multi-Speaker Adaptive Text-to-Speech Generation

StyleSpeech - PyTorch Implementation PyTorch Implementation of Meta-StyleSpeech : Multi-Speaker Adaptive Text-to-Speech Generation. Status (2021.06.09

Keon Lee 142 Jan 06, 2023
A framework for cleaning Chinese dialog data

A framework for cleaning Chinese dialog data

Yida 136 Dec 20, 2022
A desktop GUI providing an audio interface for GPT3.

Jabberwocky neil_degrasse_tyson_with_audio.mp4 Project Description This GUI provides an audio interface to GPT-3. My main goal was to provide a conven

16 Nov 27, 2022
VoiceFixer VoiceFixer is a framework for general speech restoration.

VoiceFixer VoiceFixer is a framework for general speech restoration. We aim at the restoration of severly degraded speech and historical speech. Paper

Leo 174 Jan 06, 2023
Web mining module for Python, with tools for scraping, natural language processing, machine learning, network analysis and visualization.

Pattern Pattern is a web mining module for Python. It has tools for: Data Mining: web services (Google, Twitter, Wikipedia), web crawler, HTML DOM par

Computational Linguistics Research Group 8.4k Dec 30, 2022
Pytorch version of BERT-whitening

BERT-whitening This is the Pytorch implementation of "Whitening Sentence Representations for Better Semantics and Faster Retrieval". BERT-whitening is

Weijie Liu 255 Dec 27, 2022
The swas programming language

The Swas programming language This is a language that was made for fun. Installation Step 0: Make sure you have python installed Step 1. Clone this re

Swas.py 19 Jul 18, 2022
Extract city and country mentions from Text like GeoText without regex, but FlashText, a Aho-Corasick implementation.

flashgeotext ⚡ 🌍 Extract and count countries and cities (+their synonyms) from text, like GeoText on steroids using FlashText, a Aho-Corasick impleme

Ben 57 Dec 16, 2022
Simple GUI where you can enter an article and get a crisp summarized version.

Text-Summarization-using-TextRank-BART Simple GUI where you can enter an article and get a crisp summarized version. How to run: Clone the repo Instal

Rohit P 4 Sep 28, 2022
Python functions for summarizing and improving voice dictation input.

Helpmespeak Help me speak uses Python functions for summarizing and improving voice dictation input. Get started with OpenAI gpt-3 OpenAI is a amazing

Margarita Humanitarian Foundation 6 Dec 17, 2022
Abhijith Neil Abraham 2 Nov 05, 2021
HAIS_2GNN: 3D Visual Grounding with Graph and Attention

HAIS_2GNN: 3D Visual Grounding with Graph and Attention This repository is for the HAIS_2GNN research project. Tao Gu, Yue Chen Introduction The motiv

Yue Chen 1 Nov 26, 2022
The SVO-Probes Dataset for Verb Understanding

The SVO-Probes Dataset for Verb Understanding This repository contains the SVO-Probes benchmark designed to probe for Subject, Verb, and Object unders

DeepMind 20 Nov 30, 2022
Ecco is a python library for exploring and explaining Natural Language Processing models using interactive visualizations.

Visualize, analyze, and explore NLP language models. Ecco creates interactive visualizations directly in Jupyter notebooks explaining the behavior of Transformer-based language models (like GPT2, BER

Jay Alammar 1.6k Dec 25, 2022
Simple Text-To-Speech Bot For Discord

Simple Text-To-Speech Bot For Discord This is a very simple TTS bot for discord made with python. For this bot you need FFMPEG, see installation to se

1 Sep 26, 2022
Data loaders and abstractions for text and NLP

torchtext This repository consists of: torchtext.data: Generic data loaders, abstractions, and iterators for text (including vocabulary and word vecto

3.2k Dec 30, 2022