{"id":26997,"date":"2026-09-27T20:02:20","date_gmt":"2026-09-27T18:02:20","guid":{"rendered":"https:\/\/www.itspy.cz\/thesis\/adaptive-layer-skipping-during-large-language-modelinference\/"},"modified":"2026-10-23T13:45:51","modified_gmt":"2026-10-23T11:45:51","slug":"adaptive-layer-skipping-during-large-language-modelinference","status":"publish","type":"thesis","link":"https:\/\/www.itspy.cz\/en\/thesis\/adaptive-layer-skipping-during-large-language-modelinference\/","title":{"rendered":"Adaptive layer skipping during large language model\ninference"},"content":{"rendered":"<p>Large language model inference demands substantial<br \/>\ncomputation, yet not all of it may be necessary. We assume<br \/>\nthat tokens vary in the amount of processing they need,<br \/>\ntherefore skipping some feed-forward network (FFN) layers<br \/>\nmay reduce computation without significantly degrading<br \/>\nmodel performance.<br \/>\nIn this work, we investigate the possibilities of adaptive<br \/>\nlayer skipping during inference, based on the current token<br \/>\nrepresentation. We focus on layer importance and the<br \/>\nrelationship between the magnitude of layer outputs and<br \/>\ntheir impact on model quality. We compare token-level<br \/>\nrouting to static layer removal.<br \/>\nWe propose two routing methods: (1) skipping layers with<br \/>\nsmall estimated output based on separately trained<br \/>\nclassifiers, and (2) using gating layers trained jointly to<br \/>\napproximate the outputs of the modified model to those of<br \/>\nthe original. Token routing allows us to omit one third of<br \/>\nFFN computation, at the cost of an increase in perplexity<br \/>\nfrom 7.3 to 15.3 on WikiText-2.<\/p>\n","protected":false},"featured_media":26757,"template":"","meta":{"_acf_changed":false,"_links_to":"","_links_to_target":""},"university":[175,177,440],"thesis-year":[466],"class_list":["post-26997","thesis","type-thesis","status-publish","has-post-thumbnail","hentry","thesis-year-466"],"acf":{"autor":"Gabriela Chut\u0148\u00e1kov\u00e1","portret":"","vedouci":"Mgr. Vladim\u00edr Bo\u017ea, PhD."},"_links":{"self":[{"href":"https:\/\/www.itspy.cz\/en\/wp-json\/wp\/v2\/thesis\/26997","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.itspy.cz\/en\/wp-json\/wp\/v2\/thesis"}],"about":[{"href":"https:\/\/www.itspy.cz\/en\/wp-json\/wp\/v2\/types\/thesis"}],"version-history":[{"count":1,"href":"https:\/\/www.itspy.cz\/en\/wp-json\/wp\/v2\/thesis\/26997\/revisions"}],"predecessor-version":[{"id":27052,"href":"https:\/\/www.itspy.cz\/en\/wp-json\/wp\/v2\/thesis\/26997\/revisions\/27052"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.itspy.cz\/en\/wp-json\/wp\/v2\/media\/26757"}],"wp:attachment":[{"href":"https:\/\/www.itspy.cz\/en\/wp-json\/wp\/v2\/media?parent=26997"}],"wp:term":[{"taxonomy":"university","embeddable":true,"href":"https:\/\/www.itspy.cz\/en\/wp-json\/wp\/v2\/university?post=26997"},{"taxonomy":"thesis-year","embeddable":true,"href":"https:\/\/www.itspy.cz\/en\/wp-json\/wp\/v2\/thesis-year?post=26997"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}