Upgrade to Pro — share decks privately, control downloads, hide ads and more …

8-bit Quantization of Transformer Model

Sponsored · Your Podcast. Everywhere. Effortlessly. Share. Educate. Inspire. Entertain. You do you. We'll handle the rest. →
Avatar for Scatter Lab Inc. Scatter Lab Inc.
April 29, 2020
Research
0
2.3k

8-bit Quantization of Transformer Model

Avatar for Scatter Lab Inc.

Scatter Lab Inc.

April 29, 2020
Tweet

More Decks by Scatter Lab Inc.

See All by Scatter Lab Inc.
zeta introduction
Avatar for Scatter Lab Inc. scatterlab
0
1.8k
SimCLR: A Simple Framework for Contrastive Learning of Visual Representations
Avatar for Scatter Lab Inc. scatterlab
0
4.3k
Adversarial Filters of Dataset Biases
Avatar for Scatter Lab Inc. scatterlab
0
2.3k
Sparse, Dense, and Attentional Representations for Text Retrieval
Avatar for Scatter Lab Inc. scatterlab
0
2.3k
Weight Poisoning Attacks on Pre-trained Models
Avatar for Scatter Lab Inc. scatterlab
0
2.2k
Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval
Avatar for Scatter Lab Inc. scatterlab
0
2.5k
Beyond Accuracy: Behavioral Testing of NLP Models with CheckList
Avatar for Scatter Lab Inc. scatterlab
0
2.3k
Open-Retrieval Conversational Question Answering
Avatar for Scatter Lab Inc. scatterlab
0
2.3k
What Can Neural Networks Reason About?
Avatar for Scatter Lab Inc. scatterlab
0
2.3k

Other Decks in Research

See All in Research
データサイエンティストの業務変化
Avatar for The Japan DataScientist Society datascientistsociety
PRO
0
330
「車1割削減、渋滞半減、公共交通2倍」を 熊本から岡山へ＠RACDA設立30周年記念都市交通フォーラム2026
Avatar for Traffic Brain trafficbrain
1
840
A History of Approximate Nearest Neighbor Search from an Applications Perspective
Avatar for Yusuke Matsui matsui_528
1
220
SkySense V2: A Unified Foundation Model for Multi-modal Remote Sensing
Avatar for SatAI.challenge satai
3
710
ドメイン知識がない領域での自然言語処理の始め方
Avatar for Hayahide Yamagishi hargon24
1
270
From Data Meshes to Data Spaces
Avatar for Posedio posedio
PRO
0
540
Unified Audio Source Separation (Defense Slides)
Avatar for Kohei Saijo kohei_1979
1
540
HU Berlin: Industrial-Strength Natural Language Processing with spaCy and Prodigy
Avatar for Ines Montani inesmontani
PRO
0
300
「なんとなく」の顧客理解から脱却する ──顧客の解像度を武器にするインサイトマネジメント
Avatar for 田島佳穂 tajima_kaho
10
7.2k
[IBIS 2025] 深層基盤モデルのための強化学習 驚きから理論にもとづく納得へ
Avatar for Akifumi Wachi akifumi_wachi
20
9.8k
第二言語習得研究における 明示的・暗示的知識の再検討：この分類は何に役に立つか，何に役に立たないか
Avatar for Yu Tamura tam07pb915
0
2.3k
オーストリア流 都市の公共交通サービス水準評価@公共交通オープンデータ最前線2026
Avatar for Traffic Brain trafficbrain
0
110

Featured

See All Featured
A Guide to Academic Writing Using Generative AI - A Workshop
Avatar for Kenji Saito ks91
PRO
0
250
VelocityConf: Rendering Performance Case Studies
Avatar for Addy Osmani addyosmani
333
24k
Scaling GitHub
Avatar for Zach Holman holman
464
140k
B2B Lead Gen: Tactics, Traps & Triumph
Avatar for Sophie Logan marketingsoph
0
94
The Anti-SEO Checklist Checklist. Pubcon Cyber Week
Avatar for Ryan Jones ryanjones
0
110
We Are The Robots
Avatar for Honza Javorek honzajavorek
0
210
HDC tutorial
Avatar for Michiel Stock michielstock
1
590
Helping Users Find Their Own Way: Creating Modern Search Experiences
Avatar for Dan Newman danielanewman
31
3.1k
SEO in 2025: How to Prepare for the Future of Search
Avatar for Michael King ipullrank
3
3.4k
Bootstrapping a Software Product
Avatar for Garrett Dimon garrettdimon
PRO
307
120k
What's in a price? How to price your products and services
Avatar for Michael Herold michaelherold
247
13k
Discover your Explorer Soul
Avatar for Emna emna__ayadi
2
1.1k

Transcript

  1. #JU2VBOUJ[BUJPOPG5SBOTGPSNFS.PEFM ੿਌੤ .BDIJOF-FBSOJOH4PGUXBSF&OHJOFFS 1JOHQPOH

  2. ݾର ݾର  &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM  "CTUSBDU  *OUSPEVDUJPO  .FUIPE

     3FTVMU

  3. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS /FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM

  4. "CTUSBDU &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM • *$.- *OUFM • ߓ੄ࣘب UISPVHIQVUೱ࢚ਸ঳ਵݶࢲ੿ب੄#-&6TDPSFBDDVSBDZ݅ڄয૗ • *OUFMDQVী୭੸ച

    • 5FOTPS'MPX • '1*OU 4

  5. *OUSPEVDUJPO &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM • ୭न੄*OUFM$16ٜ਷WFDUPSJ[FEOFVSBMOFUXPSLJOTUSVDUJPO 7//* ٜਸನೣ • ѐ੄CJUܳ'." 'VTFE.VMUJQMZBOE"EE 0QFSBUJPOਸجܻחѪਸ$ZDMF۽ࣻ೯

    • .BJO$POUSJCVUJPO • '1*/5RVBOUJ[BUJPOਸ޷݅੄405"#-&64DPSFೞۅ݅ਵ۽੉ܖযն • 1FSGPSNBODF0QUJNJ[BUJPO • .BU.VM • 2VBOUJ[FE.BU.VM(SBQI0QUJNJ[BUJPO • *OQVU1JQFMJOF0QUJNJ[BUJPO • 1BSBMMFM&YFDVUJPO 5

  6. .PEFM%FTDSJQUJPO &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM • ೨ब'1*/5۽QSFDJTJPOਸ൞ࢤೞ؊ۄب୽࠙൤ࢿמ੉ੜաৢஹನք౟ܳ2VBOUJ[F • 4PGUNBY -BZFS/PSNBMJ[BUJPOҗэ਷҃਋ח*/5۽աఋղӝী൨ٚч੉݆ই
 ֫਷BDD൞ࢤ੉৘࢚ؽ • .FBO

    7BSJBODF &YQ١੄҅࢑੉*/5۽աఋղӝী൨ٝ 6

  7. /BJWF2VBOUJ[BUJPO &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM • ӝઓ݆੉ॳ੉؍-JOFBS2VBOUJ[BUJPOߑधࢎਊ • .BY .JOਸ҅࢑೧ঠೞ޲۽-JOFBSTDBOਸ೙ਃ۽ೣ • 2VBOUJ[BUJPO0WFSIFBE0 /

     7

  8. /BJWF2VBOUJ[BUJPO &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 8

  9. /BJWF2VBOUJ[BUJPO &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 9 .JHBD[ 4CJUJOGFSFODFXJUIUFOTPSSU  63-IUUQPOEFNBOEHQVUFDIDPOGDPNHUDQSFTFOUBUJPOTCJUJOGFSFODFXJUIUFOTPSSUQEG

  10. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM ,-%JWFSHFODFGPSPQUJNBMUISFTIPME • 2VBOUJ[BUJPO਷যରೖNBQQJOHਸযڌѱੜೞוջоޙઁ • '1UFOTPSEJTUSJCVUJPO_*/5UFOTPSEJTUSJCVUJPO • ߈ࠂ೧оݶࢲ*/5߸ജਸਤೠ0QUJNBM.JO .BYܳ଺ח׮ •

    0QUJNBM౸ױਸ,-%JWFSHFODF۽҅࢑ • 7BMJEBUJPO%BUBTFUѐ੄ޙ੢઺ѐ੿بSBOEPNTBNQMJOH • .JO .BY5ISFTIPMEفѐܳ੿೧ঠೞחؘ Ӓߑߨਸࣁо૑੿ب۽ա־যࠆ 10 4ZNNFUSJD $POKVHBUF Ӓրٮ۽҅࢑

  11. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM ,-%JWFSHFODFGPSPQUJNBMUISFTIPME 11

  12. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM ,-%JWFSHFODFGPSPQUJNBMUISFTIPME • /BJWF2VBOUJ[BUJPO਷45015PLFOਸյࣻহ঻ӝٸޙী/" • Ӕؘࣘب࢚0GGTFU੉;FSPоغחಞ੉޷ࣁೞѱࣘبо؊ࡅܰ׮ • Ӓېࢲ4ZNNFUSJDࢎਊೣ 12

  13. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM ,-%JWFSHFODFGPSPQUJNBMUISFTIPME 13 .JHBD[ 4CJUJOGFSFODFXJUIUFOTPSSU  63-IUUQPOEFNBOEHQVUFDIDPOGDPNHUDQSFTFOUBUJPOTCJUJOGFSFODFXJUIUFOTPSSUQEG

  14. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 1FSGPSNBODF0QUJNJ[BUJPO • 7//**OTUSVDUJPOࢎਊ • 0QFSBUJPO୨іࣻ઴੉ӝ • 3FPSEFS0QFSBUJPO • .,-۽0QUJNJ[BUJPO૓೯

    • 1BSBMMFMJ[FCBUDIJOHFYFDVUJPO 14

  15. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 1FSGPSNBODF0QUJNJ[BUJPO2VBOUJ[FE.BU.VMT • "79 ࠺౟੄'."ো࢑$ZDMFীоמ • '1ѐ */5ѐ • $BTDBEF-BLF$16ࠗఠ*/57//*ܳ؊୭੸ച

    • 7//*ࢎਊೠ*/5.BU.VM਷"79ࢎਊೠ'1.BU.VMࠁ׮Yࡅܴ • 7//*ࢎਊೠ*/5.BU.VM਷"79ࢎਊೠ*/5.BU.VMࠁ׮Yࡅܴ 15

  16. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 1FSGPSNBODF0QUJNJ[BUJPO2VBOUJ[FE.BU.VMT 16

  17. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 1FSGPSNBODF0QUJNJ[BUJPO2VBOUJ[FE.BU.VMT • ೞ૑݅5FOTPS'MPXחJOUFHFS.BU.VMਊਵ۽(&..-081ۄחPQFOTPVSDFܳࢎਊೣ • (&..-081ח*/57//*ܳࢎਊೞ૑ঋҊ */5ো࢑द߸ജҗ੿ژೠ೙ਃೣ • Ӓېࢲ.,-#-"4ೣٜࣻ۽૒੽2VBOUJ[BUJPO4UFQ੘ࢿ •

    /PO;FSP0GGTFUبহছ 17

  18. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 1FSGPSNBODF0QUJNJ[BUJPO2VBOUJ[FE.BU.VMT 18 • HPPHMFHFNNMPXQ • UFOTPSGMPXUFOTPSGMPXীࢲחই૒ࢎਊೞחѪਸഛੋ • UFOTPSGMPXDPSFLFSOFMTRVBOUJ[FE@NBUNVM@PQTDD •

    UFOTPSGMPXMJUFLFSFOFMTDQV@CBDLFOE@HFNN@HFNNMPXQI

  19. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 19 UFOTPSGMPXDPSFLFSOFMTRVBOUJ[FE@NBUNVM@PQTDD

  20. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 20 UFOTPSGMPXDPSFLFSOFMTRVBOUJ[FE@NBUNVM@PQTDD

  21. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 21 UFOTPSGMPXDPSFLFSOFMTEFRVBOUJ[F@PQDD

  22. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 22 UFOTPSGMPXDPSFLFSOFMTEFRVBOUJ[F@PQDD

  23. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 1FSGPSNBODF0QUJNJ[BUJPO2VBOUJ[FE.BU.VMT 23 • /PO;FSP0GGTFUੌ҃਋HFNN@TVT੄ো࢑

  24. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 1FSGPSNBODF0QUJNJ[BUJPO(BUIFS/% 24 • (BUIFS/%חݫݽܻ*0о઺ਃೠো࢑ੋ݅ఀ */5۽઴੉؊ۄبো࢑ࣘبীח੉ٙ੉হ਺ • ೞ૑݅'1ࠁ׮*/5੉EBUBTJ[Fоߓ पઁ۽חY ੘ਵ޲۽*0੄ೱ࢚ਸӝ؀ೡࣻ੓਺

    • (FOFSBUJPO-PPQীࢲ੹ױ҅੄Ѿҗীࢲ%FRVBOUJ[FEܳউೞҊ(BUIFS/%ܳ߄۽ࣻ೯ೡ ࣻ੓ਵ޲۽Y੄ࢿמೱ࢚੉੓঻਺

  25. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 1FSGPSNBODF0QUJNJ[BUJPO*OQVU1JQFMJOF 25 • 1BE5PLFO੉ҭ੢൤௾0WFSIFBE੉ӝٸޙী೧׼ష௾ਸহগࠁ۰חदب • 5PSDI੄1BDLFE4FRVFODFܳࢤп೮חؘ ӒѤইצ٠ • ੉֤ޙীࢲחৈ۞ߓ஖ܳTPSUJOH೧ࢲQBEUPLFO੉୭ࣗ۽ٜযоب۾ೣ

    • ੿ب੄ࢿמೱ࢚ਸ঳਺ • ੉ࠗ࠙਷5PSDI੄1BDLFE4FRVFODFэ਷ো࢑җэ਷ѐ֛ਸҳഅ೧ࢲॳݶ੿۳ೡ೙ਃبহ ੉഻ঁࡅܳѪэ׮

  26. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 1FSGPSNBODF0QUJNJ[BUJPO(SBQI0QUJNJ[BUJPO 26 • ੉੹ীࢲࣿ೮٠,-%JWFSHFODFܳ੉ਊ೧ UISFTEIPMETܳ଺חߑध਷NJO NBYܳো࢑ೞח दрਸহগળ׮$POTU0QFSBUJPOਵ۽߄Պ • 3FRVBOUJ[F৬3FRVBOUJ[BUJPO3BOHF

    0QFSBUJPOਸ(SBQIীࢲহঞҊ */5ীࢲ '-0"5۽߄۽߄Բب۾ো࢑ਸ߸҃೮׮ ӝઓ ো࢑਷*/5*/5'-0"5 

  27. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 1FSGPSNBODF0QUJNJ[BUJPO(SBQI0QUJNJ[BUJPO 27

  28. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 1FSGPSNBODF0QUJNJ[BUJPO1BSBMMFM#BUDIJOH 28 • &YFDVUJPOUJNF਷CBUDIউ੄TFOUFODFMFOHUIী੄ઓ੸੉׮ • -POHFSTFOUFODFח$16഻ܳঁബਯ੸ਵ۽ॳחѪਸҙ଴ೡࣻ੓঻Ҋ  • 4FSJBMFYFDVUJPOदীח഻ঁ࠺ബਯ੸ਵ۽ॳחѪਸࠅࣻ੓঻׮

    • ੉ѦQBSBMMFMFYFDVUJPOೞݶYࢿמೱ࢚੉ઓ੤ೣ • *NQMFNFOUBUJPO • '*'02VFVFܳҙܻೞח1BSFOU5FOTPS'MPX4FTTJPO੘ࢿ • ౠ੿$16௏য৬MPDBMNFNPSZীBGGJOJUJ[FEػ /6." DIJMEQSPDFTTGPSL • $IJMEQSPDFTTח2VFVFীࢲ"TZODISPOPVTೞѱো࢑૓೯

  29. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 1FSGPSNBODF0QUJNJ[BUJPO1BSBMMFM#BUDIJOH 29

  30. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 1FSGPSNBODF0QUJNJ[BUJPO$PODMVTJPO 30

  31. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 1FSGPSNBODF0QUJNJ[BUJPO$PODMVTJPO 31

  32. &GGJDJFOU#JU2VBOUJ[BUJPOPG5SBOTGPSNFS/FVSBM.BDIJOF-BOHVBHF5SBOTMBUJPO.PEFM 1FSGPSNBODF0QUJNJ[BUJPO$PODMVTJPO 32

  33. хࢎ೤פ׮✌ ୶о૕ޙژחҾӘೠ੼੉੓׮ݶ঱ઁٚইېোۅ୊۽োۅ઱ࣁਃ ੿਌੤ .BDIJOF-FBSOJOH4PGUXBSF&OHJOFFS 1JOHQPOH  &NBJMVLKBF!TDBUUFSMBCDPLS 'BDFCPPL!KFPOHVLKBF -JOLFEJO!KFPOHVLKBF