Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
ML on Kubernetes with Kubeflow #3 Kubeflow Pipe...
Search
Keita Watanabe
November 24, 2022
Technology
0
1.1k
ML on Kubernetes with Kubeflow #3 Kubeflow Pipelines: Part1
Kubernetes Meetup Tokyo #54
https://k8sjp.connpass.com/event/264501/
で発表したLT資料
Keita Watanabe
November 24, 2022
Tweet
Share
More Decks by Keita Watanabe
See All by Keita Watanabe
Scalable Infrastructure for Large-Scale AI Training with AWS Sagemaker Hyperpod @ Singapore AI Hour
keitaw
0
15
[AWS Summit Japan 2025] Optimizing Foundation Model Development with Amazon SageMaker HyperPod: Insights from Training the Amazon Nova Model
keitaw
0
20
Building foundation models on AWS
keitaw
0
290
[re:Invent2024 Chalktalk] Cost-effectively deploy PyTorch LLMs on AWS Inferentia using Amazon EKS
keitaw
0
86
AWS Summit New York 2024: CMP 301 Demystifying the ML software stack on Amazon EC2 accelerated instances
keitaw
0
360
re:Invent 2023 CMP319 Deploy LLMs with AWS Inferentia & Ray to optimize performance and cost
keitaw
0
36
re:Invent 2023: CMP332 De-mystifying ML software stack on Amazon EC2 accelerated instances
keitaw
1
160
AWS における LLM・GenAI 大規模学習への取り組み / Large scale GenAI・LLM training on AWS
keitaw
1
540
Amazon EC2 シリコン革命 / Amazon EC2 Silicon Innovation
keitaw
0
180
Other Decks in Technology
See All in Technology
EKS Pod Identity における推移的な session tags
z63d
1
200
小さなチーム 大きな仕事 - 個人開発でAIをフル活用する
himaratsu
0
120
Android Studio の 新しいAI機能を試してみよう / Try out the new AI features in Android Studio
yanzm
0
270
7月のガバクラ利用料が高かったので調べてみた
techniczna
2
230
ドキュメントはAIの味方!スタートアップのアジャイルを加速するADR
kawauso
3
340
Gaze-LLE: Gaze Target Estimation via Large-Scale Learned Encoders
kzykmyzw
0
310
Yahoo!ニュースにおけるソフトウェア開発
lycorptech_jp
PRO
0
330
kintone開発チームの紹介
cybozuinsideout
PRO
0
73k
第4回 関東Kaggler会 [Training LLMs with Limited VRAM]
tascj
12
1.7k
Webアクセシビリティ入門
recruitengineers
PRO
1
220
AIドリブンのソフトウェア開発 - うまいやり方とまずいやり方
okdt
PRO
9
570
人を動かすことについて考える
ichimichi
2
320
Featured
See All Featured
Measuring & Analyzing Core Web Vitals
bluesmoon
9
570
Designing Experiences People Love
moore
142
24k
Git: the NoSQL Database
bkeepers
PRO
431
65k
How STYLIGHT went responsive
nonsquared
100
5.7k
The Power of CSS Pseudo Elements
geoffreycrofte
77
5.9k
Imperfection Machines: The Place of Print at Facebook
scottboms
268
13k
Navigating Team Friction
lara
189
15k
Raft: Consensus for Rubyists
vanstee
140
7.1k
The Pragmatic Product Professional
lauravandoore
36
6.8k
Automating Front-end Workflow
addyosmani
1370
200k
Done Done
chrislema
185
16k
Building Flexible Design Systems
yeseniaperezcruz
328
39k
Transcript
Keita Watanabe 2022/11/24 Kubernetes Meetup Tokyo # 54 LT KubeflowͰ͡ΊΔ
ML on Kubernetes #3 Kube fl ow Pipelines: Part 1
Keita Watanabe Machine Learning Solutions Architect ▶ AWS JapanͰSelf-managedͳMachine Learning
ͷҊ݅Λ୲͍ͯ͠ΔSolutions Architect ▶ લ৬ͰDatascientist/ML Researcherͱͯ͠ ECαΠτ্ͷը૾ݕࡧػೳͷ։ൃʹैࣄ ▶ Twitter: keitaw09 ▶ Linkedin: keitawatanabe αϯσΟΤΰͰग़ձͬͨτϦ झຯ
ͪ͜ΒͷLTγϦʔζͷͰ͢ KubeflowͰ͡ΊΔML on Kubernetes • #1 Kube fl owͷ֓ཁͱηοτΞοϓ •
#2 Kube fl ow Notebooks • #3 Kube fl ow Pipelines (ࠓճʂ)
ࠓճͷςʔϚɿ Kubeflow Pipelines ʢͷಋೖʣ
Kubeflow Pipelines ʢKFPʣͱ KFPͷߏཁૉ • Pipelineͷ֬ೝɾ࣮ߦʹ༻͍ΔUI • Pipeline࣮ߦΛεέδϡʔϦϯά͢ΔEngine • ύΠϓϥΠϯͷఆٛɺϏϧυɺσϓϩΠ͕Մೳͳ
Python SDK • SDKͰͷύΠϓϥΠϯ։ൃɺ͓Αͼ࣮ߦʹؔ͢Δ Notebook αϙʔτ Componentͱͯ͠ɺ֤εςοϓΛ࣮͠ɺͦΕΒΛPipelineͱͯ͠Ұ࿈ͷॲཧʹ· ͱΊΔ͜ͱͰMLύΠϓϥΠϯΛߏங͢Δπʔϧ UI: pipeline graph view Ұ෦ൈਮ
KFP SDKͷόʔδϣϯʹ͍ͭͯ • 2022/11/24ݱࡏɺKFP SDKʹҎԼͷ̎ͭͷVersion͕ଘࡏ͢Δ v1ʢStable Statusʣ/ v2ʢBeta Statusʣ •
KFP SDK v2ɺݱࡏPre-release stageͷKFP v2ͷίΞػೳΛ KFP v1্Ͱಈ࡞Մೳͱͨ͠ͷ • ͜͜ͰKFP SDK v2Λ༻͍Δ • KFP SDK v2ͷར༻ʹKFP 1.6Ҏ߱ ͕ඞཁʢࠓճ1.8.4Λ༻ʣ https://www.kube fl ow.org/docs/components/pipelines/v1/sdk-v2/v2-compatibility/
SDKΛ༻͍ͨγϯϓϧͳPipelineͷ࣮ྫ (1) ίϯϙʔωϯτͷఆٛ ͜͜Ͱ͠ࢉΛ”python:3.9”Πϝʔδ্Ͱ ࣮ߦ͢ΔίϯϙʔωϯτΛఆٛ͢Δ (2) ύΠϓϥΠϯͷఆٛ ύΠϓϥΠϯͰ༻͍Δίϯϙʔωϯτ܈ͱ ͦͷؔΛఆٛ͢Δ (3)
ύΠϓϥΠϯͷ࣮ߦ ύΠϓϥΠϯʹೖྗΛ༩͑ɺ࣮ߦ͢Δ (1) (2) (3)
͜ͷίʔυΛ࣮ߦͯ͠ΈΔͱ… • “/var/run/secrets/kube fl ow/pipelines/token” ͕ແ͍ͱౖΒΕɺ࣮ߦʹࣦഊ͢Δ • ࣮ϚϧνϢʔβʔڥͰSDKΛ࣮ߦ͢Δࡍ ʹɺServiceAccount token͕
”KF_PIPELINES_SA_TOKEN_PATH”ʹଘࡏ͢ Δඞཁ͕͋Δ ʢ͜ͷڥมͷσϑΥϧτͷύε͕ “/var/run/secrets/kube fl ow/pipelines/token”ʣ • σϑΥϧτͷNotebookڥʹ͜ͷτʔΫϯ ͕ઃఆ͞Ε͍ͯͳ͍ˠ PodDefaultΛ༻͍Δ https://www.kube fl ow.org/docs/components/pipelines/v1/sdk/connect-api/#full-kube fl ow-subfrom-inside-clustersub
PodDefault • KubernetesʹPodʹରͯ͠ɺ ڥมVolumeͷใͳͲ Λ͋ͱ͔ΒՃ͢ΔPodPreset ͱ͍͏Ϧιʔεʢݱࡏ Alpha stageʣ͕ଘࡏ͢Δ • PodDefaultPodPreset૬ͷ
ػೳΛKube fl owଆͰ ࣮ͨ͠ͷ https://github.com/kube fl ow/kube fl ow/blob/master/components/admission-webhook/README.md
PodDefaultͷ͍ํ ৽نNotebook࡞࣌ͷ ”Con fi gurations”ΑΓ࡞ͨ͠ PodDefaultΛબ͢Δ͜ͱ͕Ͱ͖Δ
τʔΫϯ͕Ϛϯτ͞Ε͍ͯΔ͜ͱ ͕֬ೝͰ͖Δ ύΠϓϥΠϯͷ࣮ߦʹޭ͠ɺ ExperimentͱRunͷৄࡉͷϦϯΫ ͕දࣔ͞Εͨ
Run details
࣍ճ… ࣮ફతͳPipelinesߏஙͷΛ͠·͢
ࢀߟࢿྉ • https://techblog.zozo.com/entry/aip-pipelines-impl#Kube fl ow-Pipelines • https://www.kube fl ow.org/docs/components/pipelines/v1/sdk/connect-api/ #full-kube
fl ow-subfrom-inside-clustersub • https://www.kube fl ow.org/docs/components/notebooks/overview • https://github.com/aws-samples/aws-do-kube fl ow • https://www.kube fl ow.org/docs/components/pipelines/v1/sdk-v2/v2- compatibility/ • https://www.kube fl ow.org/docs/components/pipelines/v1/sdk/connect-api/ #full-kube fl ow-subfrom-inside-clustersub