FlashAttention - Tri Dao | Stanford MLSys #67

Episode 67 of the Stanford MLSys Seminar “Foundation Models Limited Series”!
Speaker: Tri Dao
Abstract:
Transformers are slow and memory-hungry on long sequences, since the time and memory complexity of self-attention are quadratic in sequence length. Approximate attention methods have attempted to address this problem by trading off model quality to reduce the compute complexity, but often do not achieve wall-clock speedup. We argue that a missing principle is making attention algorithms IO-aware -- accounting for reads and writes between levels of GPU memory. We propose FlashAttention, an IO-aware exact attention algorithm that uses tiling to reduce the number of memory reads/writes between GPU high bandwidth memory (HBM) and GPU on-chip SRAM. We analyze the IO complexity of FlashAttention, showing that it requires fewer HBM accesses than standard attention, and is optimal for a range of SRAM sizes. We also extend FlashAttention to block-sparse attention, yielding an approximate attention algorithm that is faster than any existing approximate attention method. FlashAttention trains Transformers faster than existing baselines: 15% end-to-end wall-clock speedup on BERT-large (seq. length 512) compared to the MLPerf 1.1 training speed record, 3× speedup on GPT-2 (seq. length 1K), and 2.4× speedup on long-range arena (seq. length 1K-4K). FlashAttention and block-sparse FlashAttention enable longer context in Transformers, yielding higher quality models (0.7 better perplexity on GPT-2 and 6.4 points of lift on long-document classification) and entirely new capabilities: the first Transformers to achieve better-than-chance performance on the Path-X challenge (seq. length 16K, 61.4% accuracy) and Path-256 (seq. length 64K, 63.1% accuracy).
This work received the Best Paper Award at the Hardware-Aware Efficient Training Workshop at ICML, 2022. FlashAttention is now widely used in some of the largest research labs and companies, in just 6 months after its release.
Paper: arxiv.org/abs/...
Github: github.com/Haz...
Bio:
Tri Dao is a PhD student in Computer Science at Stanford, co-advised by Christopher Ré and Stefano Ermon. He works at the interface of machine learning and systems, and his research interests include sequence models with long-range memory and structured matrices for compact deep learning models. His work has received the ICML 2022 Outstanding paper runner-up award.
Check out our website for the schedule: mlsys.stanford.edu
Join our mailing list to get weekly updates: groups.google....

[@anishbhanushali]

22:08 (basics of attention + memory hierarchy in GPU till here ) actual explainations starts

[@denizlarson8862]

good research and nicely explained

[@rfernand2]

Great work and presentation. Where else could this be applied?

[@shuminghu]

Why does tiling reduce HBM to SRAM transfer? Or is it through pipelining that transfer time overlap more with compute?

[@xianbiaoqi7009]

Good idea and nice talk.

[@aamirmirza2806]

Really nice well explained.

[@sskhdsk]

simple and effective

[@JazevoAudiosurf]

well explained

[@deepanshusingh2527]

This is utilised in inference as well? How fast compared to naive implementation?

[@TheAIEpiphany]

btw at 28:10 the animation got the order wrong compared to the paper's Algorithm 1, the inner loop should be going over queries not over values

[@for-ever-22]

These videos are amazing

[@kawingchan]

I am not familiar at all with CPU or GPU architecture, so i naturally wonder how much of this also applies to Apple GPU (MPS). It was mentioned this is already in pytorch, but i do doubt if it even get activated on MPS. I would love to know, maybe at high level, how it may (if possible) be ported to Apple GPU, which has this unified memory thing.

[@brandomiranda6703]

ML for theorem proving would also benefit with longer sequences! Reference Lemma proved in 300 BC...

[@brandomiranda6703]

11:09