NLP Lineage

11 July 2024 · Efficiency and infrastructure

FlashAttention-3

FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision

J. Shah, G. Bikshandi, Y. Zhang, et al. (6 authors)

Date
11 July 2024
What the date marks
arXiv submission
Kind
Infrastructure
Organisation
Colfax / Together AI / Princeton
Access
BSD-3-Clause
Other dates
NeurIPS 2024
Threads
In the guided tour
Chapter 9. Making it run anywhere
Learning phases
6. Local models and deployment

Notable: no later milestone in the data is documented as descending from it. 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

FlashAttention-2 was tuned for the previous GPU generation.

The problem

Hopper GPUs have asynchronous copy engines and FP8 units that the older kernel left idle.

What it introduced

Overlap computation with data movement and exploit low precision, roughly doubling attention throughput on H100s.

Why it matters

Attention kernels are now co-designed with each GPU generation, and this is the current reference implementation.

What modern models inherit

Hardware-specific attention kernels as standard infrastructure.

What came next

Folded into serving stacks and the training runs of most 2024 and 2025 models.

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

FlashAttention-3 has 13 documented ancestors across 6 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 5 turning points among them.

FlashAttention-3 has 13 documented ancestors across 6 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

FlashAttention-3 has 15 recorded ancestors, including indirect ones, across 6 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 5 turning points among them.

FlashAttention-3 has 15 recorded ancestors, including indirect ones, across 6 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

  1. 2024FlashAttention-3This milestone.
  2. 2022FlashAttentionFlashAttention-3 built directly on thisChapter 9. Making it run anywhere
  3. 2020BigBirdFlashAttention succeeded thisChapter 10. Reading more than you remember
  4. 2019Sparse Transformer Turning pointBigBird was influenced by thisChapter 4. Pretrain once, use everywhere
  5. 2017Transformer Turning pointFlashAttention built directly on thisChapter 3. Attention, then only attention
  6. 2016Layer normTransformer built directly on thisChapter 3. Attention, then only attention
  7. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  8. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  9. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  10. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  11. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  12. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  13. 1997LSTM Turning pointSeq2seq built directly on thisChapter 1. Language as probability
  14. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability
  1. 2024FlashAttention-3This milestone.
  2. 2022FlashAttentionFlashAttention-3 built directly on thisChapter 9. Making it run anywhere
  3. 2020BigBirdFlashAttention succeeded thisChapter 10. Reading more than you remember
  4. 2019Sparse Transformer Turning pointBigBird was influenced by thisChapter 4. Pretrain once, use everywhere
  5. 2019Transformer-XLSparse Transformer was influenced by this (indirect influence)Chapter 4. Pretrain once, use everywhere
  6. 2017Transformer Turning pointFlashAttention built directly on thisChapter 3. Attention, then only attention
  7. 2017ConvS2STransformer was influenced by this (indirect influence)Chapter 3. Attention, then only attention
  8. 2016Layer normTransformer built directly on thisChapter 3. Attention, then only attention
  9. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  10. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  11. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  12. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  13. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  14. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  15. 1997LSTM Turning pointSeq2seq built directly on thisChapter 1. Language as probability
  16. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability

After: what it led to

No documented descendants are recorded for FlashAttention-3 without weak links.

No documented descendants are recorded for FlashAttention-3 without weak links.

No recorded descendants are recorded for FlashAttention-3.

No recorded descendants are recorded for FlashAttention-3.

Direct links and their sources

Where it came from

Built directly on

What it led to

None recorded in the sources used.