Tweet
Tweet

Slide 1

Slide 1 text

打造 MLOps 平台 改善 AI 模型開發流程 Bo-Yi Wu 2023/07/19 CloudSummit 台灣雲端⼤會

Slide 2

Slide 2 text

About me • Software Engineer in Mediatek (AIDE) • Maintain Drone CI/CD Plugin (15) • Member of Gitea Platform • Member of Gin Golang Framework • Maintain GitHub Actions Plugins (18).

Slide 3

Slide 3 text

No content

Slide 4

Slide 4 text

⼤綱 • What’s MLOps? • AI Developer Pain Points • AI MLOps Solution

Slide 5

Slide 5 text

What’s MLOps

Slide 6

Slide 6 text

https://en.wikipedia.org/wiki/MLOps

Slide 7

Slide 7 text

1. Shortening Development Cycles 2. Increasing Development Velocity

Slide 8

Slide 8 text

MLOps: Manual process https://bit.ly/3L8xMRC

Slide 9

Slide 9 text

Automated ML Pipeline

Slide 10

Slide 10 text

Automated ML Pipeline

Slide 11

Slide 11 text

AI Developer Pain Points

Slide 12

Slide 12 text

AI Development Pain Points • 運算資源不⾜? (80%) • 如何⾃動化流程? (20%)

Slide 13

Slide 13 text

1. 運算資源不⾜ 1. 待顆 GPU 不⽤夠 2. 單台 GPU 不夠⽤ 3. 資料處理太慢 4. ……

Slide 14

Slide 14 text

2. 如何⾃動化流程 (20%)

Slide 15

Slide 15 text

2. 如何⾃動化流程 (20%)

Slide 16

Slide 16 text

AI MLOps Solution

Slide 17

Slide 17 text

ߗᱪҡޢ ܥ౷ҡӡ ላश੒ຊ ։ᚙ੒ຊ 運算資源不⾜? • 各個部⾨購買硬體 (如右圖) • 開發者跳到 SW/HW 使⽤ (如下圖)

Slide 18

Slide 18 text

Scaling AI and Python Workloads Effortlessly with Ray

Slide 19

Slide 19 text

No content

Slide 20

Slide 20 text

解決⽅案

Slide 21

Slide 21 text

開發者能做什麼? ฏߦԽ႔ཧ ࢿྉ႔ཧ ੔߹$PNQVUJOH'BSNࢿݯ

Slide 22

Slide 22 text

⾃動化流程管理 (Internal)

Slide 23

Slide 23 text

⾃動化流程管理 (AWS)

Slide 24

Slide 24 text

ML as a Service

Slide 25

Slide 25 text

No content