AI: Megatron the Transformer, and its related language models

 

What’s in my AI? A Comprehensive Analysis of Datasets Used to Train GPT-1, GPT-2, GPT-3, GPT-NeoX-20B, Megatron-11B, MT-NLG, and Gopher

Alan D. Thompson
LifeArchitect.ai
March 2022
26 pages incl title page, references, appendix.

Read more…

What is Megatron?

Megatron is a large, powerful transformer developed by the Applied Deep Learning Research team at NVIDIA, based on work by Google.

How to use it

Play with the Megatron-11B model at Adam Daniel King’s InferKit.com.

Viz: Megatron MT-NLG (530B, September 2021)

Megatron-Turing Natural Language Generation model (MT-NLG). MT-NLG is the successor to Microsoft Turing NLG 17B and NVIDIA Megatron-LM 8.3B. The MT-NLG model is three times larger than GPT-3 (530B vs 175B).

Download source (PDF)
Contents: View the data (Google sheets)

Viz: Evolution of Megatron (2019-2021)


* RealNews is practically the same as Common Crawl News (CC-News). RealNews is 120GB from 5,000 domains from Common Crawl Dec/2016-Mar/2019. CC-News is 76GB from Common Crawl Sep/2016-Feb/2019. They are shown with different colours here (amber/blue) for interest only.

Download source (PDF)
Contents: View the data (Google sheets)

Timeline


November 2018: Google open sources BERT. Trained in four days.

Name BERT
B
idirectional Encoder Representations from Transformers.
Lab Google
Parameters 345M
Dataset sources English Wikipedia (12GB)
+ BookCorpus (4GB).
Dataset total size 16GB


July 2019: Facebook AI and University of Washington introduce RoBERTa.

Name RoBERTa.
Robustly optimized BERT approach.
Lab FAIR (Facebook AI Research) + UW
Parameters 125M (RoBERTa-base)
Dataset sources Trained with BERT original dataset:
English Wikipedia (12GB)
+ BookCorpus (4GB)
+ CC-News, 63 million English news articles from Sep/2016-Feb/2019 (76GB).
+ OpenWebText/Reddit upvoted (38GB).
+ Stories, 1M story documents from the CC (31GB).
Dataset total size 161GB


August 2019: NVIDIA introduces Megatron-LM. Trained in 53 minutes.
8.3 billion parameter transformer language model with data parallelism trained on 512 GPUs.

Name Megatron-LM
Lab NVIDIA
Parameters 8.3B (8,300M)
Dataset sources WikipediaOpenWebTextRealNews, + CC-Stories.
Dataset total size 174GB

 

April 2020: Facebook AI Research labs introduce Megatron-11b (RoBERTa).
Megatron-11b is a unidirectional language model with 11B parameters based on Megatron-LM. Following the original Megatron work, FAIR trained the model using intra-layer model parallelism with each layer’s parameters split across 8 GPUs.

Name Megatron-11B
Lab FAIR (Facebook AI Research)
Parameters 11B (11,000M)
Dataset sources Same as RoBERTa. Trained with BERT original dataset:
English Wikipedia (12GB)
+ BookCorpus (4GB)
+ CC-News, 63 million English news articles from Sep/2016-Feb/2019 (76GB).
+ OpenWebText/Reddit upvoted (38GB).
+ Stories, 1M story documents from the CC (31GB).
Dataset total size 161GB


 
October 2021: NVIDIA and Microsoft introduce Megatron-Turing NLG 530B (The Pile).
Megatron-Turing Natural Language Generation model (MT-NLG). MT-NLG is the successor to Microsoft Turing NLG 17B and NVIDIA Megatron-LM 8.3B. The MT-NLG model is three times larger than GPT-3 (530B vs 175B). Following the original Megatron work, NVIDIA and Microsoft trained the model on over 4,000 GPUs.

Name Megatron MT-NLG
Lab NVIDIA and Microsoft
Parameters 530B (530,000M)
Dataset sources Trained with The Pile v1 + more, totalling 15 datasets:
Books3
OpenWebText2 (Reddit links)
Stack Exchange
PubMed Abstracts
Wikipedia
Gutenberg (PG-19)
BookCorpus2
NIH ExPorter
Pile-CC
ArXiv
GitHub
+ Common Crawl 2020
+ Common Crawl 2021
+ RealNews, from 5000 news domains (120GB).
+ CC-Stories, 1M story documents from the CC (31GB).
Dataset total size >825GB
(My estimate is 1.86TB or 1,863GB)
Contents: View the data (Google sheets)

“We live in a time where AI advancements are far outpacing Moore’s law. We continue to see more computation power being made available with newer generations of GPUs, interconnected at lightning speeds. At the same time, we continue to see hyperscaling of AI models leading to better performance, with seemingly no end in sight.”
— NVIDIA and Microsoft (October 2021)


 
December 2021: Meta AI introduces Fairseq.
NOTE: Fairseq is not related to Megatron, and the two use different technologies for training. The only link is in the datasets, which is the primary focus of this page. 13B trained in 2,363 GPU-days (assume 1,024 GPUs, for a total of ~3 days).

Name Fairseq.
Lab Meta, previously known as FAIR (Facebook AI Research)
Parameters 13B and 1.1T
Dataset sources Same as RoBERTa (+CC100). Trained with BERT original dataset:
English Wikipedia (12GB)
+ BookCorpus (4GB)
+ CC-News, 63 million English news articles from Sep/2016-Feb/2019 (76GB).
+ OpenWebText/Reddit upvoted (38GB).
+ Stories, 1M story documents from the CC (31GB).
+ new addition of English CC100 in Wikipedia style from Jan/2018-Dec/2018 (292GB).
Dataset total size 453GB

Get The Memo

by Dr Alan D. Thompson · Be inside the lightning-fast AI revolution.
Bestseller. 10,000+ readers from 142 countries. Microsoft, Tesla, Google...
Artificial intelligence that matters, as it happens, in plain English.
Get The Memo.

Dr Alan D. Thompson is an AI expert and consultant, advising Fortune 500s and governments on post-2020 large language models. His work on artificial intelligence has been featured at NYU, with Microsoft AI and Google AI teams, at the University of Oxford’s 2021 debate on AI Ethics, and in the Leta AI (GPT-3) experiments viewed more than 4.5 million times. A contributor to the fields of human intelligence and peak performance, he has held positions as chairman for Mensa International, consultant to GE and Warner Bros, and memberships with the IEEE and IET. Technical highlights.

This page last updated: 21/Mar/2022. https://lifearchitect.ai/megatron/