{"id":200,"date":"2026-07-17T11:33:05","date_gmt":"2026-07-17T09:33:05","guid":{"rendered":"https:\/\/www.martinpavlicek.cz\/blog\/?p=200"},"modified":"2026-07-17T11:33:05","modified_gmt":"2026-07-17T09:33:05","slug":"poradna-kvantizace-binarni","status":"publish","type":"post","link":"https:\/\/www.martinpavlicek.cz\/blog\/poradna-kvantizace-binarni\/","title":{"rendered":"Po\u0159\u00e1dn\u00e1 kvantizace (bin\u00e1rn\u00ed)"},"content":{"rendered":"\n<p>V\u00fdzkumn\u00edci z PrismML ned\u00e1vno p\u0159edstavili zaj\u00edmav\u00fd technologick\u00fd posun, kter\u00fd detailn\u011b popisuj\u00ed ve sv\u00e9m whitepaperu k modelu Bonsai-27b. Zam\u011b\u0159uj\u00ed se v n\u011bm na extr\u00e9mn\u00ed \u2013 konkr\u00e9tn\u011b bin\u00e1rn\u00ed (1bitovou) \u2013 kvantizaci velk\u00fdch jazykov\u00fdch model\u016f. Zat\u00edmco standardn\u00ed modely b\u011b\u017en\u011b vy\u017eaduj\u00ed 16bitovou nebo 8bitovou p\u0159esnost, bin\u00e1rn\u00ed kvantizace o\u0159ez\u00e1v\u00e1 v\u00e1hy modelu na naprost\u00e9 minimum. Tento radik\u00e1ln\u00ed p\u0159\u00edstup drasticky sni\u017euje n\u00e1roky na opera\u010dn\u00ed pam\u011b\u0165 (VRAM) a v\u00fdpo\u010detn\u00ed v\u00fdkon, p\u0159i\u010dem\u017e c\u00edlem je zachovat co nejv\u011bt\u0161\u00ed \u010d\u00e1st p\u016fvodn\u00ed schopnosti a p\u0159esnosti modelu. V praxi to znamen\u00e1, \u017ee se masivn\u00ed modely dok\u00e1\u017eou vt\u011bsnat do zlomku p\u016fvodn\u00ed velikosti.&nbsp;<\/p>\n\n\n\n<p>Zdroj:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>https:\/\/github.com\/PrismML-Eng\/Bonsai-demo\/blob\/main\/bonsai-27b-whitepaper.pdf<\/li>\n\n\n\n<li><\/li>\n<\/ul>\n\n\n\n<p>Koment\u00e1\u0159:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Tento sm\u011br v\u00fdvoje je naprosto kl\u00ed\u010dov\u00fd. Pokud se hardwarov\u00e1 n\u00e1ro\u010dnost na provoz modelu dok\u00e1\u017ee d\u00edky t\u011bmto kompresn\u00edm metod\u00e1m sn\u00ed\u017eit v obrovsk\u00fdch n\u00e1sobc\u00edch, otev\u00edr\u00e1 n\u00e1m to cestu k datov\u00e9 nez\u00e1vislosti. Je to re\u00e1ln\u00e1 mo\u017enost, jak m\u00edt sv\u00e9 vlastn\u00ed, v\u00fdkonn\u00e9 a zcela soukrom\u00e9 osobn\u00ed AI dostupn\u00e9 na \u201ejak\u00e9mkoliv\u201c lok\u00e1ln\u00edm za\u0159\u00edzen\u00ed \u2013 a\u0165 u\u017e jde o b\u011b\u017en\u00fd notebook nebo do budoucna i mobiln\u00ed telefon \u2013 bez nutnosti neust\u00e1l\u00e9ho p\u0159ipojen\u00ed na drah\u00fd cloud.<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>V\u00fdzkumn\u00edci z PrismML ned\u00e1vno p\u0159edstavili zaj\u00edmav\u00fd technologick\u00fd posun, kter\u00fd detailn\u011b popisuj\u00ed ve sv\u00e9m whitepaperu k modelu Bonsai-27b. Zam\u011b\u0159uj\u00ed se v n\u011bm na extr\u00e9mn\u00ed \u2013 konkr\u00e9tn\u011b bin\u00e1rn\u00ed (1bitovou) \u2013 kvantizaci velk\u00fdch jazykov\u00fdch model\u016f. Zat\u00edmco standardn\u00ed modely b\u011b\u017en\u011b vy\u017eaduj\u00ed 16bitovou nebo 8bitovou p\u0159esnost, bin\u00e1rn\u00ed kvantizace o\u0159ez\u00e1v\u00e1 v\u00e1hy modelu na naprost\u00e9 minimum. Tento radik\u00e1ln\u00ed p\u0159\u00edstup drasticky sni\u017euje n\u00e1roky [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[93,92,91],"class_list":["post-200","post","type-post","status-publish","format-standard","hentry","category-nezarazene","tag-edge","tag-kvantizace","tag-prismml"],"_links":{"self":[{"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/posts\/200","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/comments?post=200"}],"version-history":[{"count":1,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/posts\/200\/revisions"}],"predecessor-version":[{"id":201,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/posts\/200\/revisions\/201"}],"wp:attachment":[{"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/media?parent=200"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/categories?post=200"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/tags?post=200"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}