NLP Lineage

15 August 2022 · Efficiency and infrastructure

LLM.int8()

LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale

T. Dettmers, M. Lewis, Y. Belkada, L. Zettlemoyer

Date
15 August 2022
What the date marks
arXiv submission
Kind
Technique
Organisation
U. Washington / Meta AI / Hugging Face
Access
MIT
Other dates
NeurIPS 2022
Threads
In the guided tour
Chapter 9. Making it run anywhere
Learning phases
6. Local models and deployment

A milestone: 2 later milestones descend directly from it, and its descendants reach 2 of the 10 threads. How levels are decided

Read the primary source (arxiv.org)

Full summary in progress. Until then, this page shows the milestone's record from the research data.

What it changed

Before it

Running a 175B model meant a multi-GPU server holding 16-bit weights.

The problem

Naive 8-bit quantization destroyed accuracy because of rare outlier features.

What it introduced

Quantize almost everything to 8-bit while keeping outlier dimensions in 16-bit, with no measured degradation.

Why it matters

It halved the memory needed to run big models and started the quantization toolchain most local inference uses.

What modern models inherit

Quantized inference as the normal way to run a model you did not train.

What came next

GPTQ and AWQ pushed to 4-bit; QLoRA made 4-bit trainable.

Before and after

The milestones it descends from, and the milestones that descend from it. Work developed alongside it is not counted. Weak links are left out unless you include them.

Before: where it came from

LLM.int8() has 31 documented ancestors across 7 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 11 turning points among them.

LLM.int8() has 31 documented ancestors across 7 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

LLM.int8() has 33 recorded ancestors, including indirect ones, across 7 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 11 turning points among them.

LLM.int8() has 33 recorded ancestors, including indirect ones, across 7 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

  1. 2022LLM.int8()This milestone.
  2. 2022OPT-175BLLM.int8() was influenced by thisChapter 8. The weights get out
  3. 2020The PileOPT-175B built directly on thisChapter 5. Scale becomes a science
  4. 2020GPT-3 Turning pointOPT-175B built directly on thisChapter 5. Scale becomes a science
  5. 2020Scaling laws Turning pointGPT-3 was influenced by thisChapter 5. Scale becomes a science
  6. 2019C4 Turning pointThe Pile was influenced by thisChapter 4. Pretrain once, use everywhere
  7. 2019Megatron-LMOPT-175B was influenced by thisChapter 4. Pretrain once, use everywhere
  8. 2019Sparse Transformer Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  9. 2019GPT-2 Turning pointGPT-3 built directly on this
    Megatron-LM was influenced by this    Chapter 4. Pretrain once, use everywhere
  10. 2019WebTextGPT-2 was influenced by thisChapter 4. Pretrain once, use everywhere
  11. 2018BERT Turning pointMegatron-LM was influenced by thisChapter 4. Pretrain once, use everywhere
  12. 2018GPT-1BERT was influenced by this
    GPT-2 built directly on this    Chapter 4. Pretrain once, use everywhere
  13. 2018ELMoBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  14. 2018ULMFiTBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  15. 2017Mixed precisionMegatron-LM was influenced by thisChapter 9. Making it run anywhere
  16. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  17. 2017Transformer Turning pointMegatron-LM built directly on thisChapter 3. Attention, then only attention
  18. 2016Layer normTransformer built directly on this
    GPT-2 built directly on this    Chapter 3. Attention, then only attention
  19. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  20. 2015LM pretraining Turning pointBERT was influenced by thisChapter 2. Words become vectors
  21. 2015BPE tokenizationGPT-2 built directly on thisChapter 2. Words become vectors
  22. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  23. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  24. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  25. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  26. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  27. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  28. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  29. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  30. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  31. 1997LSTM Turning pointSeq2seq built directly on this
    ELMo built directly on this    Chapter 1. Language as probability
  32. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability
  1. 2022LLM.int8()This milestone.
  2. 2022OPT-175BLLM.int8() was influenced by thisChapter 8. The weights get out
  3. 2020The PileOPT-175B built directly on thisChapter 5. Scale becomes a science
  4. 2020GPT-3 Turning pointOPT-175B built directly on thisChapter 5. Scale becomes a science
  5. 2020Scaling laws Turning pointGPT-3 was influenced by thisChapter 5. Scale becomes a science
  6. 2019C4 Turning pointThe Pile was influenced by thisChapter 4. Pretrain once, use everywhere
  7. 2019Megatron-LMOPT-175B was influenced by thisChapter 4. Pretrain once, use everywhere
  8. 2019Sparse Transformer Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  9. 2019GPT-2 Turning pointGPT-3 built directly on this
    Megatron-LM was influenced by this    Chapter 4. Pretrain once, use everywhere
  10. 2019WebTextGPT-2 was influenced by this
    C4 was influenced by this (indirect influence)    Chapter 4. Pretrain once, use everywhere
  11. 2019Transformer-XLSparse Transformer was influenced by this (indirect influence)Chapter 4. Pretrain once, use everywhere
  12. 2018BERT Turning pointMegatron-LM was influenced by thisChapter 4. Pretrain once, use everywhere
  13. 2018GPT-1BERT was influenced by this
    GPT-2 built directly on this    Chapter 4. Pretrain once, use everywhere
  14. 2018ELMoBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  15. 2018ULMFiTBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  16. 2017Mixed precisionMegatron-LM was influenced by thisChapter 9. Making it run anywhere
  17. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  18. 2017Transformer Turning pointMegatron-LM built directly on thisChapter 3. Attention, then only attention
  19. 2017ConvS2STransformer was influenced by this (indirect influence)Chapter 3. Attention, then only attention
  20. 2016Layer normTransformer built directly on this
    GPT-2 built directly on this    Chapter 3. Attention, then only attention
  21. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  22. 2015LM pretraining Turning pointBERT was influenced by thisChapter 2. Words become vectors
  23. 2015BPE tokenizationGPT-2 built directly on thisChapter 2. Words become vectors
  24. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  25. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  26. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  27. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  28. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  29. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  30. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  31. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  32. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  33. 1997LSTM Turning pointSeq2seq built directly on this
    ELMo built directly on this    Chapter 1. Language as probability
  34. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability

After: what it led to

LLM.int8() has 2 documented descendants across 1 chapter, reaching forward to QLoRA in 2023. Showing all of them.

LLM.int8() has 2 documented descendants across 1 chapter, reaching forward to QLoRA in 2023. Showing all of them.

LLM.int8() has 3 recorded descendants, including indirect ones, across 1 chapter, reaching forward to QLoRA in 2023. Showing its direct descendants.

LLM.int8() has 3 recorded descendants, including indirect ones, across 1 chapter, reaching forward to QLoRA in 2023. Showing all of them.

  1. 2022LLM.int8()This milestone.
  2. 2022GPTQwas influenced by LLM.int8()Chapter 9. Making it run anywhere
  3. 2023QLoRAbuilt directly on LLM.int8()Chapter 9. Making it run anywhere
  1. 2022LLM.int8()This milestone.
  2. 2022GPTQwas influenced by LLM.int8()Chapter 9. Making it run anywhere
  3. 2023llama.cppwas influenced by GPTQ (indirect influence)Chapter 9. Making it run anywhere
  4. 2023QLoRAbuilt directly on LLM.int8()Chapter 9. Making it run anywhere

Direct links and their sources

Where it came from

Influenced by

What it led to

Built on directly by

Influenced