Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

SARATHI: Efficient LLM Inference by Piggybacking Decodes with Chunked Prefills (2023)
Hanayo: Harnessing Wave-like Pipeline Parallelism for Enhanced Large Model Training Efficiency (2023)
FineQuant: Unlocking Efficiency with Fine-Grained Weight-Only Quantization for LLMs (2023)
GNNPipe: Scaling Deep GNN Training with Pipelined Model Parallelism (2023)
LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models (2023)

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

\n","updatedAt":"2023-10-02T13:42:55.960Z","author":{"_id":"63d3e0e8ff1384ce6c5dd17d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg","fullname":"Librarian Bot (Bot)","name":"librarian-bot","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":264}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7397905588150024},"editors":["librarian-bot"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2309.14509","authors":[{"_id":"6513924a606fc2019243eb7f","user":{"_id":"651736dd20b18e99b44dd5d5","avatarUrl":"/avatars/c054eb2567e027408340a03a7ca0b29d.svg","isPro":false,"fullname":"Sam Ade Jacobs","user":"samadejacobs","type":"user"},"name":"Sam Ade Jacobs","status":"claimed_verified","statusLastChangedAt":"2023-10-02T07:00:29.319Z","hidden":false},{"_id":"6513924a606fc2019243eb80","name":"Masahiro Tanaka","hidden":false},{"_id":"6513924a606fc2019243eb81","user":{"_id":"6480355ccc9e514b3b033a78","avatarUrl":"/avatars/ff86bd27261c755c20546b400392f256.svg","isPro":false,"fullname":"Chengming Zhang","user":"cc32","type":"user"},"name":"Chengming Zhang","status":"admin_assigned","statusLastChangedAt":"2023-09-27T10:31:09.240Z","hidden":false},{"_id":"6513924a606fc2019243eb82","user":{"_id":"6305b5e39d2531fabd195c5f","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6305b5e39d2531fabd195c5f/oGS5leXVrB8Ct_BF96VkR.jpeg","isPro":false,"fullname":"Zhang","user":"Minjia","type":"user"},"name":"Minjia Zhang","status":"admin_assigned","statusLastChangedAt":"2023-09-27T10:29:23.957Z","hidden":false},{"_id":"6513924a606fc2019243eb83","user":{"_id":"651442f5c75a3d4c44de4ac9","avatarUrl":"/avatars/c60185389bf6379e4360cb54615922ed.svg","isPro":false,"fullname":"Shuaiwen Leon Song","user":"leonangel991","type":"user"},"name":"Leon Song","status":"claimed_verified","statusLastChangedAt":"2023-09-27T15:21:16.142Z","hidden":false},{"_id":"6513924a606fc2019243eb84","user":{"_id":"63238e79d444e1d90888ea03","avatarUrl":"/avatars/cd0accb5ddec4f2ae23c94da4094b5e5.svg","isPro":false,"fullname":"Samyam Rajbhandari","user":"samyam","type":"user"},"name":"Samyam Rajbhandari","status":"admin_assigned","statusLastChangedAt":"2023-09-27T10:29:50.197Z","hidden":false},{"_id":"6513924a606fc2019243eb85","name":"Yuxiong He","hidden":false}],"publishedAt":"2023-09-25T20:15:57.000Z","submittedOnDailyAt":"2023-09-27T00:54:11.831Z","title":"DeepSpeed Ulysses: System Optimizations for Enabling Training of Extreme\n Long Sequence Transformer Models","submittedOnDailyBy":{"_id":"60f1abe7544c2adfd699860c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674929746905-60f1abe7544c2adfd699860c.jpeg","isPro":false,"fullname":"AK","user":"akhaliq","type":"user"},"summary":"Computation in a typical Transformer-based large language model (LLM) can be\ncharacterized by batch size, hidden dimension, number of layers, and sequence\nlength. Until now, system works for accelerating LLM training have focused on\nthe first three dimensions: data parallelism for batch size, tensor parallelism\nfor hidden size and pipeline parallelism for model depth or layers. These\nwidely studied forms of parallelism are not targeted or optimized for long\nsequence Transformer models. Given practical application needs for long\nsequence LLM, renewed attentions are being drawn to sequence parallelism.\nHowever, existing works in sequence parallelism are constrained by\nmemory-communication inefficiency, limiting their scalability to long sequence\nlarge models. In this work, we introduce DeepSpeed-Ulysses, a novel, portable\nand effective methodology for enabling highly efficient and scalable LLM\ntraining with extremely long sequence length. DeepSpeed-Ulysses at its core\npartitions input data along the sequence dimension and employs an efficient\nall-to-all collective communication for attention computation. Theoretical\ncommunication analysis shows that whereas other methods incur communication\noverhead as sequence length increases, DeepSpeed-Ulysses maintains constant\ncommunication volume when sequence length and compute devices are increased\nproportionally. Furthermore, experimental evaluations show that\nDeepSpeed-Ulysses trains 2.5X faster with 4X longer sequence length than the\nexisting method SOTA baseline.","upvotes":19,"discussionId":"6513924b606fc2019243eb9f","ai_summary":"DeepSpeed-Ulysses enables efficient and scalable training of Transformer-based LLMs with long sequences by partitioning input data along the sequence dimension and using all-to-all communication.","ai_keywords":["Transformer-based","large language model","LLM","data parallelism","tensor parallelism","pipeline parallelism","sequence parallelism","memory-communication inefficiency","DeepSpeed-Ulysses","all-to-all collective communication","attention computation","communication analysis"]},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"64c41fe15e43ae1ab6df3664","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64c41fe15e43ae1ab6df3664/jx-1c_2Iu14IuXguxKKne.jpeg","isPro":false,"fullname":"Dave Goldblatt","user":"davegoldblatt","type":"user"},{"_id":"636ac507e3ad78bc68b31cfe","avatarUrl":"/avatars/e6dd4027945909c7cf13c61807c78f23.svg","isPro":false,"fullname":"Anas Saeed","user":"SaeedAnas","type":"user"},{"_id":"6457885a75f8f7d26aa5bc44","avatarUrl":"/avatars/8ce57c4d60a1f1b5afa2c592207a8335.svg","isPro":false,"fullname":"allthingsdisaggregated","user":"lastweek","type":"user"},{"_id":"5df82bcada6d0311fd3d5402","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1589104979708-5df82bcada6d0311fd3d5402.jpeg","isPro":false,"fullname":"Chuanming Liu","user":"Chuanming","type":"user"},{"_id":"6538119803519fddb4a17e10","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6538119803519fddb4a17e10/ffJMkdx-rM7VvLTCM6ri_.jpeg","isPro":false,"fullname":"samusenps","user":"samusenps","type":"user"},{"_id":"6032802e1f993496bc14d9e3","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6032802e1f993496bc14d9e3/w6hr-DEQot4VVkoyRIBiy.png","isPro":false,"fullname":"Omar Sanseviero","user":"osanseviero","type":"user"},{"_id":"648726f6759899ef7becbc4b","avatarUrl":"/avatars/a4cdd89466f32f10be38fe573e0e474a.svg","isPro":false,"fullname":"Gabriel ","user":"Kinopsis","type":"user"},{"_id":"5fcb4ec4835012afdc38cb29","avatarUrl":"/avatars/689ccd722bf64220364b9601d0bc3a7b.svg","isPro":false,"fullname":"kiran","user":"kira","type":"user"},{"_id":"6040a00558b78f3a0047c23a","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6040a00558b78f3a0047c23a/_BsyJoaCBO3r6GnfSgnIA.jpeg","isPro":false,"fullname":"David Mataciunas","user":"DeividasM","type":"user"},{"_id":"644e1b1d9b4e87c31bab0a14","avatarUrl":"/avatars/88bb4c4a67dc8958069e9014f5e73a0b.svg","isPro":false,"fullname":"Michael Barry","user":"MichaelBarryUK","type":"user"},{"_id":"5fac1f245eec0323e9470ba4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1624935381963-5fac1f245eec0323e9470ba4.jpeg","isPro":false,"fullname":"Jason Wolosonovich","user":"wolosonovich","type":"user"},{"_id":"61f8f1b03c1583dba38d9bb4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1643704749290-noauth.jpeg","isPro":false,"fullname":"SuperSecureHuman","user":"SuperSecureHuman","type":"user"}],"acceptLanguages":["*"],"dailyPaperRank":3}">

Papers

arxiv:2309.14509

DeepSpeed Ulysses: System Optimizations for Enabling Training of Extreme Long Sequence Transformer Models

Published on Sep 25, 2023

· Submitted by

AK on Sep 27, 2023

#3 Paper of the day

Upvote

Authors:

Sam Ade Jacobs ,

Chengming Zhang ,

Minjia Zhang ,

Leon Song ,

Samyam Rajbhandari ,

Abstract

DeepSpeed-Ulysses enables efficient and scalable training of Transformer-based LLMs with long sequences by partitioning input data along the sequence dimension and using all-to-all communication.

AI-generated summary

Computation in a typical Transformer-based large language model (LLM) can be characterized by batch size, hidden dimension, number of layers, and sequence length. Until now, system works for accelerating LLM training have focused on the first three dimensions: data parallelism for batch size, tensor parallelism for hidden size and pipeline parallelism for model depth or layers. These widely studied forms of parallelism are not targeted or optimized for long sequence Transformer models. Given practical application needs for long sequence LLM, renewed attentions are being drawn to sequence parallelism. However, existing works in sequence parallelism are constrained by memory-communication inefficiency, limiting their scalability to long sequence large models. In this work, we introduce DeepSpeed-Ulysses, a novel, portable and effective methodology for enabling highly efficient and scalable LLM training with extremely long sequence length. DeepSpeed-Ulysses at its core partitions input data along the sequence dimension and employs an efficient all-to-all collective communication for attention computation. Theoretical communication analysis shows that whereas other methods incur communication overhead as sequence length increases, DeepSpeed-Ulysses maintains constant communication volume when sequence length and compute devices are increased proportionally. Furthermore, experimental evaluations show that DeepSpeed-Ulysses trains 2.5X faster with 4X longer sequence length than the existing method SOTA baseline.