{"id":890,"date":"2023-12-22T05:06:34","date_gmt":"2023-12-22T05:06:34","guid":{"rendered":"https:\/\/tastycounter.net\/index.php\/2023\/12\/22\/apple-phat-trien-phuong-phap-dot-pha-de-chay-llm-tren-iphone\/"},"modified":"2023-12-22T05:06:34","modified_gmt":"2023-12-22T05:06:34","slug":"apple-phat-trien-phuong-phap-dot-pha-de-chay-llm-tren-iphone","status":"publish","type":"post","link":"https:\/\/tastycounter.net\/index.php\/2023\/12\/22\/apple-phat-trien-phuong-phap-dot-pha-de-chay-llm-tren-iphone\/","title":{"rendered":"Apple ph\u00e1t tri\u1ec3n ph\u01b0\u01a1ng ph\u00e1p \u0111\u1ed9t ph\u00e1 \u0111\u1ec3 ch\u1ea1y LLM tr\u00ean iPhone"},"content":{"rendered":"<\/p>\n
\n

Vi\u1ec7c s\u1edf h\u1eefu Apple GPT ngay trong t\u1ea7m tay c\u00f3 th\u1ec3 tr\u1edf th\u00e0nh hi\u1ec7n th\u1ef1c s\u1edbm h\u01a1n b\u1ea1n ngh\u0129. C\u00e1c nh\u00e0 nghi\u00ean c\u1ee9u AI c\u1ee7a Apple cho bi\u1ebft h\u1ecd \u0111\u00e3 \u0111\u1ea1t \u0111\u01b0\u1ee3c b\u01b0\u1edbc \u0111\u1ed9t ph\u00e1 quan tr\u1ecdng trong vi\u1ec7c tri\u1ec3n khai c\u00e1c m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn (LLM) tr\u00ean iPhone v\u00e0 nh\u1eefng thi\u1ebft b\u1ecb Apple kh\u00e1c c\u00f3 b\u1ed9 nh\u1edb h\u1ea1n ch\u1ebf b\u1eb1ng c\u00e1ch ph\u00e1t minh ra m\u1ed9t k\u1ef9 thu\u1eadt s\u1eed d\u1ee5ng b\u1ed9 nh\u1edb flash c\u1ea3i ti\u1ebfn.<\/p>\n

\"Siri
Siri tr\u00ean iPhone<\/figcaption><\/figure>\n

LLM v\u00e0 h\u1ea1n ch\u1ebf v\u1ec1 b\u1ed9 nh\u1edb<\/h2>\n

C\u00e1c chatbot d\u1ef1a tr\u00ean LLM nh\u01b0 ChatGPT<\/a> v\u00e0 Claude c\u1ef1c k\u1ef3 t\u1ed1n d\u1eef li\u1ec7u v\u00e0 b\u1ed9 nh\u1edb, th\u01b0\u1eddng c\u1ea7n l\u01b0\u1ee3ng b\u1ed9 nh\u1edb l\u1edbn \u0111\u1ec3 ho\u1ea1t \u0111\u1ed9ng, \u0111\u00e2y l\u00e0 m\u1ed9t th\u00e1ch th\u1ee9c \u0111\u1ed1i v\u1edbi c\u00e1c thi\u1ebft b\u1ecb nh\u01b0 iPhone c\u00f3 dung l\u01b0\u1ee3ng b\u1ed9 nh\u1edb h\u1ea1n ch\u1ebf. \u0110\u1ec3 gi\u1ea3i quy\u1ebft v\u1ea5n \u0111\u1ec1 n\u00e0y, c\u00e1c nh\u00e0 nghi\u00ean c\u1ee9u c\u1ee7a Apple \u0111\u00e3 ph\u00e1t tri\u1ec3n m\u1ed9t k\u1ef9 thu\u1eadt m\u1edbi s\u1eed d\u1ee5ng b\u1ed9 nh\u1edb flash \u2013 ch\u00ednh b\u1ed9 nh\u1edb l\u01b0u tr\u1eef \u1ee9ng d\u1ee5ng v\u00e0 \u1ea3nh c\u1ee7a b\u1ea1n \u2013 \u0111\u1ec3 l\u01b0u tr\u1eef d\u1eef li\u1ec7u c\u1ee7a m\u00f4 h\u00ecnh AI.<\/p>\n

L\u01b0u tr\u1eef AI tr\u00ean b\u1ed9 nh\u1edb Flash<\/h2>\n

Trong m\u1ed9t b\u00e0i nghi\u00ean c\u1ee9u m\u1edbi c\u00f3 ti\u00eau \u0111\u1ec1 “LLM in a flash: Efficient Large Language Model Inference with Limited Memory” <\/em>(T\u1ea1m d\u1ecbch l\u00e0: LLM tr\u00ean flash: Suy lu\u1eadn m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn hi\u1ec7u qu\u1ea3 v\u1edbi b\u1ed9 nh\u1edb h\u1ea1n ch\u1ebf”, c\u00e1c t\u00e1c gi\u1ea3 l\u01b0u \u00fd r\u1eb1ng b\u1ed9 nh\u1edb flash c\u00f3 nhi\u1ec1u trong c\u00e1c thi\u1ebft b\u1ecb di \u0111\u1ed9ng h\u01a1n RAM th\u01b0\u1eddng \u0111\u01b0\u1ee3c s\u1eed d\u1ee5ng \u0111\u1ec3 ch\u1ea1y LLM. Ph\u01b0\u01a1ng ph\u00e1p c\u1ee7a h\u1ecd \u0111\u00e3 kh\u00e9o l\u00e9o v\u01b0\u1ee3t qua gi\u1edbi h\u1ea1n b\u1eb1ng c\u00e1ch s\u1eed d\u1ee5ng hai k\u1ef9 thu\u1eadt ch\u00ednh gi\u00fap gi\u1ea3m thi\u1ec3u vi\u1ec7c truy\u1ec1n d\u1eef li\u1ec7u v\u00e0 t\u1ed1i \u0111a h\u00f3a th\u00f4ng l\u01b0\u1ee3ng b\u1ed9 nh\u1edb flash:<\/p>\n

<\/ins><\/div>\n
    \n
  1. Windowing<\/strong>: H\u00e3y coi \u0111\u00e2y nh\u01b0 m\u1ed9t ph\u01b0\u01a1ng ph\u00e1p t\u00e1i ch\u1ebf. Thay v\u00ec load d\u1eef li\u1ec7u m\u1edbi m\u1ed7i l\u1ea7n, m\u00f4 h\u00ecnh AI s\u1ebd s\u1eed d\u1ee5ng l\u1ea1i m\u1ed9t s\u1ed1 d\u1eef li\u1ec7u \u0111\u00e3 \u0111\u01b0\u1ee3c x\u1eed l\u00fd. \u0110i\u1ec1u n\u00e0y l\u00e0m gi\u1ea3m nhu c\u1ea7u t\u00ecm n\u1ea1p b\u1ed9 nh\u1edb li\u00ean t\u1ee5c, gi\u00fap qu\u00e1 tr\u00ecnh nhanh v\u00e0 m\u01b0\u1ee3t m\u00e0 h\u01a1n.<\/li>\n
  2. Row-Column Bundling<\/strong>: K\u1ef9 thu\u1eadt n\u00e0y gi\u1ed1ng nh\u01b0 \u0111\u1ecdc m\u1ed9t cu\u1ed1n s\u00e1ch theo t\u1eebng \u0111o\u1ea1n l\u1edbn h\u01a1n thay v\u00ec \u0111\u1ecdc t\u1eebng t\u1eeb m\u1ed9t. B\u1eb1ng c\u00e1ch nh\u00f3m d\u1eef li\u1ec7u hi\u1ec7u qu\u1ea3 h\u01a1n, d\u1eef li\u1ec7u c\u00f3 th\u1ec3 \u0111\u01b0\u1ee3c \u0111\u1ecdc nhanh h\u01a1n t\u1eeb b\u1ed9 nh\u1edb flash, t\u0103ng t\u1ed1c kh\u1ea3 n\u0103ng hi\u1ec3u v\u00e0 t\u1ea1o ng\u00f4n ng\u1eef c\u1ee7a AI.<\/li>\n<\/ol>\n

    Theo nghi\u00ean c\u1ee9u, s\u1ef1 k\u1ebft h\u1ee3p c\u1ee7a c\u00e1c ph\u01b0\u01a1ng ph\u00e1p n\u00e0y cho ph\u00e9p c\u00e1c m\u00f4 h\u00ecnh AI ch\u1ea1y \u0111\u01b0\u1ee3c dung l\u01b0\u1ee3ng g\u1ea5p \u0111\u00f4i b\u1ed9 nh\u1edb kh\u1ea3 d\u1ee5ng c\u1ee7a iPhone. \u0110i\u1ec1u n\u00e0y c\u00f3 ngh\u0129a l\u00e0 t\u1ed1c \u0111\u1ed9 t\u0103ng g\u1ea5p 4-5 l\u1ea7n tr\u00ean b\u1ed9 x\u1eed l\u00fd (CPU) ti\u00eau chu\u1ea9n v\u00e0 nhanh h\u01a1n \u1ea5n t\u01b0\u1ee3ng 20-25 l\u1ea7n tr\u00ean b\u1ed9 x\u1eed l\u00fd \u0111\u1ed3 h\u1ecda (GPU). C\u00e1c t\u00e1c gi\u1ea3 vi\u1ebft: \u201cB\u01b0\u1edbc \u0111\u1ed9t ph\u00e1 n\u00e0y \u0111\u1eb7c bi\u1ec7t quan tr\u1ecdng \u0111\u1ec3 tri\u1ec3n khai LLM ti\u00ean ti\u1ebfn trong m\u00f4i tr\u01b0\u1eddng h\u1ea1n ch\u1ebf v\u1ec1 t\u00e0i nguy\u00ean, t\u1eeb \u0111\u00f3 m\u1edf r\u1ed9ng kh\u1ea3 n\u0103ng \u1ee9ng d\u1ee5ng v\u00e0 m\u1ee9c \u0111\u1ed9 ti\u1ebfp c\u1eadn c\u1ee7a ch\u00fang\u201d<\/em>.<\/p>\n

    AI nhanh h\u01a1n tr\u00ean iPhone<\/h2>\n

    B\u01b0\u1edbc \u0111\u1ed9t ph\u00e1 v\u1ec1 hi\u1ec7u qu\u1ea3 c\u1ee7a AI m\u1edf ra nh\u1eefng kh\u1ea3 n\u0103ng m\u1edbi cho iPhone trong t\u01b0\u01a1ng lai, ch\u1eb3ng h\u1ea1n nh\u01b0 kh\u1ea3 n\u0103ng Siri ti\u00ean ti\u1ebfn h\u01a1n, d\u1ecbch ng\u00f4n ng\u1eef theo th\u1eddi gian th\u1ef1c v\u00e0 c\u00e1c t\u00ednh n\u0103ng ph\u1ee9c t\u1ea1p do AI \u0111i\u1ec1u khi\u1ec3n trong ch\u1ee5p \u1ea3nh v\u00e0 th\u1ef1c t\u1ebf t\u0103ng c\u01b0\u1eddng. C\u00f4ng ngh\u1ec7 n\u00e0y c\u0169ng t\u1ea1o ti\u1ec1n \u0111\u1ec1 cho iPhone ch\u1ea1y c\u00e1c tr\u1ee3 l\u00fd AI v\u00e0 chatbot ph\u1ee9c t\u1ea1p tr\u00ean thi\u1ebft b\u1ecb, \u0111i\u1ec1u m\u00e0 Apple \u0111\u01b0\u1ee3c cho l\u00e0 \u0111ang nghi\u00ean c\u1ee9u.<\/p>\n

    C\u00f4ng vi\u1ec7c c\u1ee7a Apple v\u1ec1 Generative AI<\/a> cu\u1ed1i c\u00f9ng c\u00f3 th\u1ec3 \u0111\u01b0\u1ee3c t\u00edch h\u1ee3p v\u00e0o tr\u1ee3 l\u00fd gi\u1ecdng n\u00f3i \u200cSiri\u200c c\u1ee7a h\u1ecd. Apple v\u00e0o th\u00e1ng 2 n\u0103m 2023 \u0111\u00e3 t\u1ed5 ch\u1ee9c h\u1ed9i ngh\u1ecb th\u01b0\u1ee3ng \u0111\u1ec9nh v\u1ec1 AI v\u00e0 gi\u1edbi thi\u1ec7u t\u00f3m t\u1eaft cho nh\u00e2n vi\u00ean v\u1ec1 c\u00f4ng vi\u1ec7c tr\u00ean m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn c\u1ee7a m\u00ecnh. Theo Bloomberg, Apple \u0111ang h\u01b0\u1edbng t\u1edbi m\u1ed9t phi\u00ean b\u1ea3n Siri th\u00f4ng minh h\u01a1n \u0111\u01b0\u1ee3c t\u00edch h\u1ee3p s\u00e2u v\u1edbi AI. Apple \u0111ang c\u00f3 k\u1ebf ho\u1ea1ch c\u1eadp nh\u1eadt c\u00e1ch \u200cSiri\u200c t\u01b0\u01a1ng t\u00e1c v\u1edbi \u1ee9ng d\u1ee5ng Messages, cho ph\u00e9p ng\u01b0\u1eddi d\u00f9ng \u0111\u01b0a ra c\u00e1c c\u00e2u h\u1ecfi ph\u1ee9c t\u1ea1p v\u00e0 t\u1ef1 \u0111\u1ed9ng ho\u00e0n th\u00e0nh c\u00e2u hi\u1ec7u qu\u1ea3 h\u01a1n. Ngo\u00e0i ra, Apple c\u00f2n \u0111\u01b0\u1ee3c \u0111\u1ed3n \u0111\u1ea1i l\u00e0 \u0111ang c\u00f3 k\u1ebf ho\u1ea1ch b\u1ed5 sung AI v\u00e0o c\u00e0ng nhi\u1ec1u \u1ee9ng d\u1ee5ng c\u1ee7a Apple c\u00e0ng t\u1ed1t.<\/p>\n

    <\/ins><\/div>\n

    Apple GPT<\/h2>\n

    Apple \u0111\u01b0\u1ee3c cho l\u00e0 \u0111ang ph\u00e1t tri\u1ec3n m\u00f4 h\u00ecnh Generative AI c\u1ee7a ri\u00eang m\u00ecnh c\u00f3 t\u00ean l\u00e0 “Ajax”. \u0110\u01b0\u1ee3c thi\u1ebft k\u1ebf \u0111\u1ec3 c\u1ea1nh tranh v\u1edbi GPT-3 v\u00e0 GPT-4 c\u1ee7a OpenAI, Ajax ho\u1ea1t \u0111\u1ed9ng tr\u00ean 200 t\u1ef7 tham s\u1ed1, cho th\u1ea5y m\u1ee9c \u0111\u1ed9 ph\u1ee9c t\u1ea1p, c\u0169ng nh\u01b0 kh\u1ea3 n\u0103ng hi\u1ec3u v\u00e0 t\u1ea1o ng\u00f4n ng\u1eef cao. \u0110\u01b0\u1ee3c bi\u1ebft \u0111\u1ebfn v\u1edbi t\u00ean g\u1ecdi n\u1ed9i b\u1ed9 l\u00e0 “Apple GPT”, Ajax \u0111\u1eb7t m\u1ee5c ti\u00eau th\u1ed1ng nh\u1ea5t ho\u1ea1t \u0111\u1ed9ng ph\u00e1t tri\u1ec3n Machine Learning tr\u00ean to\u00e0n Apple, \u0111\u1ec1 xu\u1ea5t m\u1ed9t chi\u1ebfn l\u01b0\u1ee3c l\u1edbn \u0111\u1ec3 t\u00edch h\u1ee3p AI s\u00e2u h\u01a1n v\u00e0o h\u1ec7 sinh th\u00e1i c\u1ee7a Apple.<\/p>\n

    Theo c\u00e1c b\u00e1o c\u00e1o m\u1edbi nh\u1ea5t, Ajax \u0111\u01b0\u1ee3c coi l\u00e0 c\u00f3 kh\u1ea3 n\u0103ng cao h\u01a1n ChatGPT 3.5 th\u1ebf h\u1ec7 tr\u01b0\u1edbc. Tuy nhi\u00ean, ng\u01b0\u1eddi ta c\u0169ng g\u1ee3i \u00fd r\u1eb1ng c\u00e1c m\u00f4 h\u00ecnh m\u1edbi h\u01a1n c\u1ee7a OpenAI c\u00f3 th\u1ec3 \u0111\u00e3 v\u01b0\u1ee3t qu\u00e1 kh\u1ea3 n\u0103ng c\u1ee7a Ajax k\u1ec3 t\u1eeb th\u00e1ng 9 n\u0103m 2023.<\/p>\n

    C\u1ea3 The Information v\u00e0 nh\u00e0 ph\u00e2n t\u00edch Jeff Pu \u0111\u1ec1u kh\u1eb3ng \u0111\u1ecbnh r\u1eb1ng Apple s\u1ebd c\u00f3 m\u1ed9t s\u1ed1 lo\u1ea1i t\u00ednh n\u0103ng Generative AI c\u00f3 s\u1eb5n tr\u00ean \u200ciPhone\u200c v\u00e0 iPad v\u00e0o kho\u1ea3ng cu\u1ed1i n\u0103m 2024, \u0111\u00f3 l\u00e0 th\u1eddi \u0111i\u1ec3m iOS 18 ra m\u1eaft. Pu cho bi\u1ebft v\u00e0o th\u00e1ng 10 r\u1eb1ng Apple \u0111ang x\u00e2y d\u1ef1ng v\u00e0i tr\u0103m m\u00e1y ch\u1ee7 AI v\u00e0o n\u0103m 2023 v\u00e0 nhi\u1ec1u h\u01a1n n\u1eefa v\u00e0o n\u0103m 2024. Apple \u0111\u01b0\u1ee3c cho l\u00e0 s\u1ebd cung c\u1ea5p s\u1ef1 k\u1ebft h\u1ee3p gi\u1eefa AI d\u1ef1a tr\u00ean \u0111\u00e1m m\u00e2y v\u00e0 AI v\u1edbi kh\u1ea3 n\u0103ng x\u1eed l\u00fd tr\u00ean thi\u1ebft b\u1ecb.<\/p>\n<\/div>\n","protected":false},"excerpt":{"rendered":"

    Vi\u1ec7c s\u1edf h\u1eefu Apple GPT ngay trong t\u1ea7m tay c\u00f3 th\u1ec3 tr\u1edf th\u00e0nh hi\u1ec7n th\u1ef1c s\u1edbm h\u01a1n b\u1ea1n ngh\u0129. C\u00e1c nh\u00e0 nghi\u00ean c\u1ee9u AI c\u1ee7a Apple cho bi\u1ebft h\u1ecd \u0111\u00e3 \u0111\u1ea1t \u0111\u01b0\u1ee3c b\u01b0\u1edbc \u0111\u1ed9t ph\u00e1 quan tr\u1ecdng trong vi\u1ec7c tri\u1ec3n khai c\u00e1c m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn (LLM) tr\u00ean iPhone v\u00e0 nh\u1eefng thi\u1ebft b\u1ecb Apple […]<\/p>\n","protected":false},"author":1,"featured_media":891,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-890","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-khong-phan-loai"],"_links":{"self":[{"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/posts\/890","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/comments?post=890"}],"version-history":[{"count":0,"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/posts\/890\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/media\/891"}],"wp:attachment":[{"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/media?parent=890"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/categories?post=890"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/tags?post=890"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}