서브메뉴
검색
Scheduling and Codesigning Machine Learning Models for GEMMs
Scheduling and Codesigning Machine Learning Models for GEMMs
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202103135
- ISBN
- 9798311951364
- DDC
- 330
- 저자명
- Fu, Daniel Y.
- 서명/저자
- Scheduling and Codesigning Machine Learning Models for GEMMs
- 발행사항
- [Sl] : Stanford University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 194 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-12, Section: B.
- 주기사항
- Advisor: Fatahalian, Kayvon;Ré, Chris.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2025.
- 초록/해제
- 요약Large-scale training of machine learning (ML) models on hardware accelerators like GPUs has led to substantial progress in recent years. Of critical importance is achieving high FLOP utilization during training. However, efficiently scheduling a diverse range of machine learning operators onto the specialized compute cores present in these hardware accelerators remains a challenge. For example, tensor core units, which are specialized to only compute matrix-matrix multiplication (GEMMs), compose up to 94% of the available FLOPs on Nvidia H100 GPUs. This specialization presents a challenge for modern ML models, which often interleave GEMMs with non-linearities (e.g., exponentiation), reduction operations (e.g., row-wise average), or non-GEMM operations (e.g., FFT).This thesis investigates how to schedule and codesign core machine learning operations onto specialized GEMM units, with a focus on Nvidia tensor cores (though the findings have been generalized to other hardware accelerators). We present two case studies on scheduling two core ML operations, attention and state space models (SSMs), followed by a study on codesigning a new architecture, Monarch Mixer, to utilize tensor cores from the ground-up. First, we focus on scheduling the attention operation, the core primitive at the heart of the Transformer architecture. We find that a simple algebraic reformulation of the softmax operation (combination of exponentiation and row-wise reduction) allows a tiling-based algorithm to achieve high FLOP utilization onto tensor cores. Second, we focus on SSMs, a promising next-generation architecture with better asymptotic complexity than Transformers. SSMs rely on a non-GEMM operation at their core, an FFT convolution. We show that an alternate algebraic formulation of the FFT, called a Monarch decomposition, allows the FFT convolution to be rewritten as a series of tensor core operations. Finally, we conduct a study on codesigning a new architecture, Monarch Mixer, to utilize tensor cores from the ground-up. We show that a generalized Monarch decomposition allows for a high-quality, asymptotically-efficient, and GEMM-friendly ML architecture. Overall, this thesis demonstrates that algebraic reformulation of ML operations is a powerful tool for scheduling diverse ML operations onto GEMMs.
- 일반주제명
- Sparsity
- 일반주제명
- Decomposition
- 일반주제명
- Benchmarks
- 일반주제명
- Computer science
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 86-12B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017357124
■00520260202103135
■006m o d
■007cr#unu||||||||
■020 ▼a9798311951364
■035 ▼a(MiAaPQ)AAI31974593
■035 ▼a(MiAaPQ)Stanfordcs793yc4618
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a330
■1001 ▼aFu, Daniel Y.
■24510▼aScheduling and Codesigning Machine Learning Models for GEMMs
■260 ▼a[Sl]▼bStanford University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a194 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-12, Section: B.
■500 ▼aAdvisor: Fatahalian, Kayvon;Ré, Chris.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2025.
■520 ▼aLarge-scale training of machine learning (ML) models on hardware accelerators like GPUs has led to substantial progress in recent years. Of critical importance is achieving high FLOP utilization during training. However, efficiently scheduling a diverse range of machine learning operators onto the specialized compute cores present in these hardware accelerators remains a challenge. For example, tensor core units, which are specialized to only compute matrix-matrix multiplication (GEMMs), compose up to 94% of the available FLOPs on Nvidia H100 GPUs. This specialization presents a challenge for modern ML models, which often interleave GEMMs with non-linearities (e.g., exponentiation), reduction operations (e.g., row-wise average), or non-GEMM operations (e.g., FFT).This thesis investigates how to schedule and codesign core machine learning operations onto specialized GEMM units, with a focus on Nvidia tensor cores (though the findings have been generalized to other hardware accelerators). We present two case studies on scheduling two core ML operations, attention and state space models (SSMs), followed by a study on codesigning a new architecture, Monarch Mixer, to utilize tensor cores from the ground-up. First, we focus on scheduling the attention operation, the core primitive at the heart of the Transformer architecture. We find that a simple algebraic reformulation of the softmax operation (combination of exponentiation and row-wise reduction) allows a tiling-based algorithm to achieve high FLOP utilization onto tensor cores. Second, we focus on SSMs, a promising next-generation architecture with better asymptotic complexity than Transformers. SSMs rely on a non-GEMM operation at their core, an FFT convolution. We show that an alternate algebraic formulation of the FFT, called a Monarch decomposition, allows the FFT convolution to be rewritten as a series of tensor core operations. Finally, we conduct a study on codesigning a new architecture, Monarch Mixer, to utilize tensor cores from the ground-up. We show that a generalized Monarch decomposition allows for a high-quality, asymptotically-efficient, and GEMM-friendly ML architecture. Overall, this thesis demonstrates that algebraic reformulation of ML operations is a powerful tool for scheduling diverse ML operations onto GEMMs.
■590 ▼aSchool code: 0212.
■650 4▼aSparsity
■650 4▼aDecomposition
■650 4▼aBenchmarks
■650 4▼aComputer science
■690 ▼a0984
■690 ▼a0800
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g86-12B.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17357124▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


