{"id":96860,"date":"2026-09-16T11:03:28","date_gmt":"2026-09-16T08:03:28","guid":{"rendered":"https:\/\/twodots.gr\/?p=96860"},"modified":"2026-09-16T11:03:29","modified_gmt":"2026-09-16T08:03:29","slug":"multimodal-speculative-decoding-parallili-ai","status":"publish","type":"post","link":"https:\/\/twodots.gr\/en\/multimodal-speculative-decoding-parallili-ai\/","title":{"rendered":"Multimodal speculative decoding: \u03c0\u03cc\u03c4\u03b5 \u03b7 \u03c0\u03b1\u03c1\u03ac\u03bb\u03bb\u03b7\u03bb\u03b7 AI \u03b3\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03cc\u03bd\u03c4\u03c9\u03c2 \u03c4\u03b1\u03c7\u03cd\u03c4\u03b5\u03c1\u03b7"},"content":{"rendered":"<div class=\"td-article-lede\">\n<p><strong>\u03a4\u03bf multimodal speculative decoding \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03ba\u03ac\u03bd\u03b5\u03b9 \u03ad\u03bd\u03b1 vision-language \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03c4\u03b1\u03c7\u03cd\u03c4\u03b5\u03c1\u03bf \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03b1\u03bb\u03bb\u03ac\u03be\u03b5\u03b9 \u03c4\u03b7 \u03b4\u03b9\u03b1\u03bd\u03bf\u03bc\u03ae \u03b5\u03be\u03cc\u03b4\u03bf\u03c5 \u03c4\u03bf\u03c5 target model.<\/strong> \u03a4\u03bf \u03ba\u03ad\u03c1\u03b4\u03bf\u03c2 \u03cc\u03bc\u03c9\u03c2 \u03b4\u03b5\u03bd \u03ba\u03c1\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03bc\u03cc\u03bd\u03bf \u03b1\u03c0\u03cc \u03c4\u03bf \u03c0\u03cc\u03c3\u03b1 draft tokens \u03b3\u03af\u03bd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c0\u03bf\u03b4\u03b5\u03ba\u03c4\u03ac. \u039a\u03c1\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03b7 \u03c4\u03b7 \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ae: vision encoding, prefill, \u03c0\u03c1\u03bf\u03b5\u03c4\u03bf\u03b9\u03bc\u03b1\u03c3\u03af\u03b1 \u03c4\u03bf\u03c5 drafter, verification \u03ba\u03b1\u03b9 serving.<\/p>\n<p>\u0397 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 <em>Is Multimodal Speculative Decoding Ready for Diffusion-Based Parallel Drafting?<\/em> \u03b2\u03c1\u03af\u03c3\u03ba\u03b5\u03b9 \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03ae end-to-end \u03b5\u03c0\u03b9\u03c4\u03ac\u03c7\u03c5\u03bd\u03c3\u03b7 \u03c3\u03b5 \u03b5\u03c5\u03bd\u03bf\u03ca\u03ba\u03ad\u03c2 \u03b4\u03b9\u03b1\u03bc\u03bf\u03c1\u03c6\u03ce\u03c3\u03b5\u03b9\u03c2, \u03b1\u03bb\u03bb\u03ac \u03ba\u03b1\u03b9 \u03b5\u03c0\u03b9\u03b2\u03c1\u03ac\u03b4\u03c5\u03bd\u03c3\u03b7 \u03cc\u03c4\u03b1\u03bd \u03b7 \u03c5\u03c8\u03b7\u03bb\u03ae \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03b1\u03c5\u03be\u03ac\u03bd\u03b5\u03b9 \u03c4\u03bf multimodal conditioning. \u0393\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7, \u03c4\u03bf \u03c3\u03c9\u03c3\u03c4\u03cc \u03b5\u03c1\u03ce\u03c4\u03b7\u03bc\u03b1 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u00ab\u03c0\u03bf\u03b9\u03b1 \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03c2 \u03ad\u03c7\u03b5\u03b9 \u03c4\u03bf \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf MAT;\u00bb, \u03b1\u03bb\u03bb\u03ac \u00ab\u03c0\u03bf\u03b9\u03b1 model\u2013task\u2013condition\u2013system \u03b4\u03b9\u03b1\u03bc\u03cc\u03c1\u03c6\u03c9\u03c3\u03b7 \u03ba\u03b5\u03c1\u03b4\u03af\u03b6\u03b5\u03b9 \u03c3\u03c4\u03bf \u03b4\u03b9\u03ba\u03cc \u03bc\u03b1\u03c2 workload;\u00bb.<\/p>\n<\/div>\n<div class=\"td-article-toc\">\n<div class=\"td-toc-title\">Contents<\/div>\n<ul>\n<li><a href=\"#speculative-decoding-l2\">\u03a4\u03b9 \u03ba\u03ac\u03bd\u03b5\u03b9 \u03c4\u03bf speculative decoding \u03ba\u03b1\u03b9 \u03b3\u03b9\u03b1\u03c4\u03af \u03b5\u03bc\u03c6\u03b1\u03bd\u03af\u03c3\u03c4\u03b7\u03ba\u03b5 \u03c4\u03bf L2<\/a><\/li>\n<li><a href=\"#multimodal-workload\">\u0393\u03b9\u03b1\u03c4\u03af \u03c4\u03bf multimodal workload \u03b5\u03af\u03bd\u03b1\u03b9 \u03b4\u03c5\u03c3\u03ba\u03bf\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf<\/a><\/li>\n<li><a href=\"#peirama-workloads\">\u03a4\u03bf \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1: \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1, \u03ad\u03be\u03b9 workloads \u03ba\u03b1\u03b9 \u03c3\u03c9\u03c3\u03c4\u03cc baseline<\/a><\/li>\n<li><a href=\"#dflash-dspark-apotelesmata\">DFlash \u03ba\u03b1\u03b9 DSpark: \u03c4\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03bc\u03b5 \u03c4\u03bf \u03c3\u03c9\u03c3\u03c4\u03cc \u03c0\u03bb\u03b1\u03af\u03c3\u03b9\u03bf<\/a><\/li>\n<li><a href=\"#megalitero-target\">\u03a0\u03cc\u03c4\u03b5 \u03c4\u03bf \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf target \u03b1\u03c0\u03bf\u03c3\u03b2\u03ad\u03bd\u03b5\u03b9 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b1 \u03c4\u03bf\u03bd drafter<\/a><\/li>\n<li><a href=\"#bottleneck-prefill\">\u03a4\u03bf bottleneck \u03bc\u03b5\u03c4\u03b1\u03c6\u03ad\u03c1\u03b5\u03c4\u03b1\u03b9 \u03c3\u03c4\u03bf multimodal prefill<\/a><\/li>\n<li><a href=\"#analysi-8k\">\u0393\u03b9\u03b1\u03c4\u03af \u03b7 \u03c5\u03c8\u03b7\u03bb\u03ae \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c6\u03ad\u03c1\u03b5\u03b9 \u03b5\u03c0\u03b9\u03b2\u03c1\u03ac\u03b4\u03c5\u03bd\u03c3\u03b7<\/a><\/li>\n<li><a href=\"#visual-context-drafter\">\u03a0\u03cc\u03c3\u03bf visual context \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac \u03bf drafter<\/a><\/li>\n<li><a href=\"#inference-ecosystem\">\u03a4\u03bf inference ecosystem \u03c0\u03c1\u03bf\u03c7\u03c9\u03c1\u03ac, \u03b1\u03bb\u03bb\u03ac \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03ac\u03bd\u03b9\u03c3\u03bf<\/a><\/li>\n<li><a href=\"#plaisio-apofasis\">\u0388\u03bd\u03b1 \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03cc \u03c0\u03bb\u03b1\u03af\u03c3\u03b9\u03bf \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7\u03c2 \u03b3\u03b9\u03b1 multimodal \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd\u03c4\u03b1<\/a><\/li>\n<li><a href=\"#teliko-symperasma\">\u03a4\u03bf \u03c3\u03c5\u03bc\u03c0\u03ad\u03c1\u03b1\u03c3\u03bc\u03b1: \u03bc\u03b5\u03c1\u03b9\u03ba\u03ae \u03b5\u03c4\u03bf\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1, \u03cc\u03c7\u03b9 \u03ba\u03b1\u03b8\u03bf\u03bb\u03b9\u03ba\u03ae \u03c3\u03c5\u03bd\u03c4\u03b1\u03b3\u03ae<\/a><\/li>\n<\/ul>\n<\/div>\n<h2 id=\"speculative-decoding-l2\">\u03a4\u03b9 \u03ba\u03ac\u03bd\u03b5\u03b9 \u03c4\u03bf speculative decoding \u03ba\u03b1\u03b9 \u03b3\u03b9\u03b1\u03c4\u03af \u03b5\u03bc\u03c6\u03b1\u03bd\u03af\u03c3\u03c4\u03b7\u03ba\u03b5 \u03c4\u03bf L2<\/h2>\n<p>\u03a3\u03c4\u03bf \u03ba\u03bb\u03b1\u03c3\u03b9\u03ba\u03cc speculative decoding, \u03ad\u03bd\u03b1\u03c2 \u03c6\u03b8\u03b7\u03bd\u03cc\u03c4\u03b5\u03c1\u03bf\u03c2 drafter \u03c0\u03c1\u03bf\u03c4\u03b5\u03af\u03bd\u03b5\u03b9 \u03bc\u03b5\u03bb\u03bb\u03bf\u03bd\u03c4\u03b9\u03ba\u03ac tokens \u03ba\u03b1\u03b9 \u03c4\u03bf \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf target model \u03c4\u03b1 \u03b5\u03c0\u03b1\u03bb\u03b7\u03b8\u03b5\u03cd\u03b5\u03b9 \u03c0\u03b1\u03c1\u03ac\u03bb\u03bb\u03b7\u03bb\u03b1. \u039c\u03b5 exact speculative sampling, \u03cc\u03c4\u03b1\u03bd \u03ad\u03bd\u03b1 token \u03b1\u03c0\u03bf\u03c1\u03c1\u03af\u03c0\u03c4\u03b5\u03c4\u03b1\u03b9 \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03b5\u03af\u03c4\u03b1\u03b9 \u03c4\u03bf \u03b1\u03c0\u03bf\u03b4\u03b5\u03ba\u03c4\u03cc prefix \u03ba\u03b1\u03b9 \u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03b5\u03af \u03b4\u03b9\u03bf\u03c1\u03b8\u03c9\u03c4\u03b9\u03ba\u03cc sampling. \u0388\u03c4\u03c3\u03b9 \u03b7 \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03c2 \u03c3\u03c4\u03bf\u03c7\u03b5\u03cd\u03b5\u03b9 \u03c3\u03b5 \u03c7\u03b1\u03bc\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf \u03c7\u03c1\u03cc\u03bd\u03bf \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae\u03c2 \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03b1\u03bd\u03c4\u03b9\u03ba\u03b1\u03b8\u03b9\u03c3\u03c4\u03ac \u03c4\u03b7\u03bd \u03ba\u03b1\u03c4\u03b1\u03bd\u03bf\u03bc\u03ae \u03c4\u03bf\u03c5 target \u03bc\u03b5 \u03b5\u03ba\u03b5\u03af\u03bd\u03b7 \u03c4\u03bf\u03c5 drafter.<\/p>\n<p>\u0397 \u03b2\u03b1\u03c3\u03b9\u03ba\u03ae \u03b4\u03c5\u03c3\u03ba\u03bf\u03bb\u03af\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03ad\u03bd\u03b1\u03c2 autoregressive drafter \u03b5\u03be\u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03b5\u03af \u03bd\u03b1 \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 \u03c4\u03b1 K \u03c5\u03c0\u03bf\u03c8\u03ae\u03c6\u03b9\u03b1 tokens \u03c3\u03b5\u03b9\u03c1\u03b9\u03b1\u03ba\u03ac. \u0397 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03b4\u03b9\u03b1\u03ba\u03c1\u03af\u03bd\u03b5\u03b9 \u03c4\u03c1\u03af\u03b1 \u03b5\u03c0\u03af\u03c0\u03b5\u03b4\u03b1 \u03c0\u03b1\u03c1\u03b1\u03bb\u03bb\u03b7\u03bb\u03b9\u03c3\u03bc\u03bf\u03cd. \u03a3\u03c4\u03bf L0, \u03ba\u03ac\u03b8\u03b5 forward pass \u03c0\u03c1\u03bf\u03c7\u03c9\u03c1\u03ac \u03ad\u03bd\u03b1 draft depth. \u03a3\u03c4\u03bf L1, \u03bc\u03af\u03b1 \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7 \u03c0\u03c1\u03bf\u03b2\u03bb\u03ad\u03c0\u03b5\u03b9 \u03c0\u03bf\u03bb\u03bb\u03ad\u03c2 \u03c0\u03c1\u03bf\u03ba\u03b1\u03b8\u03bf\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03b5\u03c2 \u03bc\u03b5\u03bb\u03bb\u03bf\u03bd\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b8\u03ad\u03c3\u03b5\u03b9\u03c2 \u03b1\u03c0\u03cc \u03ba\u03bf\u03b9\u03bd\u03cc prefix, \u03cc\u03c0\u03c9\u03c2 \u03c3\u03c5\u03bc\u03b2\u03b1\u03af\u03bd\u03b5\u03b9 \u03bc\u03b5 multi-token prediction heads. \u03a3\u03c4\u03bf L2, \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03bf \u03c4\u03bf \u03bc\u03b5\u03bb\u03bb\u03bf\u03bd\u03c4\u03b9\u03ba\u03cc block \u03b3\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b7 \u03b2\u03b1\u03c3\u03b9\u03ba\u03ae \u03b3\u03b5\u03bd\u03b5\u03c4\u03b9\u03ba\u03ae \u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03c4\u03b1\u03b9 \u03ae \u03b2\u03b5\u03bb\u03c4\u03b9\u03ce\u03bd\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03bf \u03c3\u03b5\u03b9\u03c1\u03b9\u03b1\u03ba\u03cc \u03b2\u03ac\u03b8\u03bf\u03c2.<\/p>\n<div class=\"td-comparison td-comparison-cards td-comparison-cards--horizontal\">\n<div class=\"td-comparison-grid td-comparison-grid--three\">\n<div class=\"td-platform-card\">\n<p class=\"td-comparison-title\">L0: \u03c3\u03b5\u03b9\u03c1\u03b9\u03b1\u03ba\u03cc draft<\/p>\n<p>\u039f drafter \u03c0\u03c1\u03bf\u03c7\u03c9\u03c1\u03ac \u03bc\u03af\u03b1 \u03b8\u03ad\u03c3\u03b7 \u03b1\u03bd\u03ac forward pass. \u0397 \u03b5\u03c0\u03b1\u03bb\u03ae\u03b8\u03b5\u03c5\u03c3\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03b1\u03c1\u03ac\u03bb\u03bb\u03b7\u03bb\u03b7, \u03b1\u03bb\u03bb\u03ac \u03b7 \u03b4\u03b7\u03bc\u03b9\u03bf\u03c5\u03c1\u03b3\u03af\u03b1 \u03b2\u03ac\u03b8\u03bf\u03c5\u03c2 K \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03b1\u03b9\u03c4\u03b9\u03b1\u03ba\u03ac \u03c3\u03b5\u03b9\u03c1\u03b9\u03b1\u03ba\u03ae.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">EAGLE-3<\/span><span class=\"td-badge\">K forwards<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card\">\n<p class=\"td-comparison-title\">L1: \u03c0\u03bf\u03bb\u03bb\u03ad\u03c2 \u03b8\u03ad\u03c3\u03b5\u03b9\u03c2<\/p>\n<p>\u0388\u03bd\u03b1 \u03ba\u03bf\u03b9\u03bd\u03cc prefix \u03c4\u03c1\u03bf\u03c6\u03bf\u03b4\u03bf\u03c4\u03b5\u03af \u03c0\u03bf\u03bb\u03bb\u03ad\u03c2 \u03c0\u03c1\u03bf\u03ba\u03b1\u03b8\u03bf\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03b5\u03c2 \u03c0\u03c1\u03bf\u03b2\u03bb\u03ad\u03c8\u03b5\u03b9\u03c2, \u03c7\u03c9\u03c1\u03af\u03c2 \u03c4\u03bf block \u03bd\u03b1 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b5\u03af \u03c9\u03c2 \u03b5\u03bd\u03b9\u03b1\u03af\u03b1 \u03b5\u03be\u03b5\u03bb\u03b9\u03c3\u03c3\u03cc\u03bc\u03b5\u03bd\u03b7 \u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">MTP<\/span><span class=\"td-badge\">Multi-position<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card td-platform-card--navy\">\n<p class=\"td-comparison-title\">L2: \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03bf block<\/p>\n<p>\u03a4\u03b1 DFlash \u03ba\u03b1\u03b9 DSpark \u03c0\u03b1\u03c1\u03ac\u03b3\u03bf\u03c5\u03bd \u03c4\u03bf \u03bc\u03b5\u03bb\u03bb\u03bf\u03bd\u03c4\u03b9\u03ba\u03cc block \u03c9\u03c2 \u03ba\u03bf\u03b9\u03bd\u03ae \u03bc\u03bf\u03bd\u03ac\u03b4\u03b1, \u03bc\u03b5\u03b9\u03ce\u03bd\u03bf\u03bd\u03c4\u03b1\u03c2 \u03c4\u03bf \u03c3\u03b5\u03b9\u03c1\u03b9\u03b1\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03c4\u03bf\u03c5 \u03af\u03b4\u03b9\u03bf\u03c5 \u03c4\u03bf\u03c5 drafting.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">DFlash<\/span><span class=\"td-badge\">DSpark<\/span><\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>\u03a4\u03bf candidate tree \u03ba\u03b1\u03b9 \u03bf \u03c4\u03c1\u03cc\u03c0\u03bf\u03c2 verification \u03b5\u03af\u03bd\u03b1\u03b9 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b4\u03b9\u03b1\u03c3\u03c4\u03ac\u03c3\u03b5\u03b9\u03c2. \u0388\u03bd\u03b1 \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03ba\u03c1\u03b1\u03c4\u03ac \u03c0\u03bf\u03bb\u03bb\u03ad\u03c2 \u03b5\u03bd\u03b1\u03bb\u03bb\u03b1\u03ba\u03c4\u03b9\u03ba\u03ad\u03c2 \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03ad\u03c7\u03b5\u03b9 L2 drafter \u03ae \u03bd\u03b1 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af block-parallel drafter \u03bc\u03b5 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ae \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae acceptance. \u0397 \u03b4\u03b9\u03ac\u03ba\u03c1\u03b9\u03c3\u03b7 \u03b1\u03c0\u03bf\u03c4\u03c1\u03ad\u03c0\u03b5\u03b9 \u03c4\u03b7 \u03c3\u03cd\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7 \u03b1\u03bd\u03cc\u03bc\u03bf\u03b9\u03c9\u03bd \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ce\u03bd \u03ba\u03ac\u03c4\u03c9 \u03b1\u03c0\u03cc \u03c4\u03b7 \u03b3\u03b5\u03bd\u03b9\u03ba\u03ae \u03b5\u03c4\u03b9\u03ba\u03ad\u03c4\u03b1 \u00abparallel decoding\u00bb. \u0393\u03b9\u03b1 \u03b2\u03b1\u03b8\u03cd\u03c4\u03b5\u03c1\u03b7 \u03b5\u03b9\u03ba\u03cc\u03bd\u03b1 \u03c4\u03b7\u03c2 \u03b7\u03bc\u03b9-\u03b1\u03c5\u03c4\u03bf\u03c0\u03b1\u03bb\u03af\u03bd\u03b4\u03c1\u03bf\u03bc\u03b7\u03c2 \u03c0\u03c1\u03bf\u03c3\u03ad\u03b3\u03b3\u03b9\u03c3\u03b7\u03c2, \u03b4\u03b5\u03af\u03c4\u03b5 \u03c0\u03ce\u03c2 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b5\u03af \u03c4\u03bf <a href=\"https:\/\/twodots.gr\/dspark-speculative-decoding-ai-latency\/\">speculative decoding \u03bc\u03b5 DSpark<\/a>.<\/p>\n<h2 id=\"multimodal-workload\">\u0393\u03b9\u03b1\u03c4\u03af \u03c4\u03bf multimodal workload \u03b5\u03af\u03bd\u03b1\u03b9 \u03b4\u03c5\u03c3\u03ba\u03bf\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf<\/h2>\n<p>\u0388\u03bd\u03b1 vision-language \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03b4\u03b5\u03bd \u03be\u03b5\u03ba\u03b9\u03bd\u03ac \u03bc\u03cc\u03bd\u03bf \u03b1\u03c0\u03cc \u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf. \u03a0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03ba\u03c9\u03b4\u03b9\u03ba\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03b5\u03b9 \u03b5\u03b9\u03ba\u03cc\u03bd\u03b1 \u03ae video, \u03bd\u03b1 \u03b4\u03b7\u03bc\u03b9\u03bf\u03c5\u03c1\u03b3\u03ae\u03c3\u03b5\u03b9 multimodal representations \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03ba\u03ac\u03bd\u03b5\u03b9 prefill \u03c0\u03c1\u03b9\u03bd \u03b1\u03c1\u03c7\u03af\u03c3\u03b5\u03b9 \u03c4\u03bf decoding. \u0391\u03ba\u03cc\u03bc\u03b7 \u03ba\u03b1\u03b9 \u03b1\u03bd \u03c4\u03bf draft block \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03c4\u03b1\u03b9 \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03b1, \u03bf drafter \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c0\u03c1\u03cc\u03c3\u03b2\u03b1\u03c3\u03b7 \u03c3\u03b5 visual features \u03ae \u03bc\u03b5\u03c4\u03b1\u03c4\u03c1\u03bf\u03c0\u03ae target hidden states \u03c3\u03b5 \u03b4\u03b9\u03ba\u03cc \u03c4\u03bf\u03c5 KV cache. \u0391\u03c5\u03c4\u03ae \u03b7 \u03c0\u03c1\u03bf\u03b5\u03c4\u03bf\u03b9\u03bc\u03b1\u03c3\u03af\u03b1 \u03ba\u03bf\u03c3\u03c4\u03af\u03b6\u03b5\u03b9 \u03c7\u03c1\u03cc\u03bd\u03bf \u03ba\u03b1\u03b9 \u03bc\u03bd\u03ae\u03bc\u03b7.<\/p>\n<p>\u0397 \u03ad\u03bd\u03bd\u03bf\u03b9\u03b1 \u03c4\u03b7\u03c2 \u03c3\u03c9\u03c3\u03c4\u03ae\u03c2 \u03b1\u03c0\u03bf\u03b4\u03bf\u03c7\u03ae\u03c2 \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03b1\u03bd\u03ac domain. \u0393\u03b9\u03b1 text output, \u03b7 \u03b1\u03ba\u03c1\u03b9\u03b2\u03ae\u03c2 token equivalence \u03b5\u03af\u03bd\u03b1\u03b9 \u03c6\u03c5\u03c3\u03b9\u03ba\u03cc\u03c2 \u03c3\u03c4\u03cc\u03c7\u03bf\u03c2. \u03a3\u03b5 actions \u03b5\u03bd\u03cc\u03c2 vision-language-action agent, speech codec tokens \u03ae visual codebooks, \u03b4\u03cd\u03bf \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03af\u03b5\u03c2 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b9\u03ba\u03ac \u03ae \u03b1\u03bd\u03c4\u03b9\u03bb\u03b7\u03c0\u03c4\u03b9\u03ba\u03ac \u03b9\u03c3\u03bf\u03b4\u03cd\u03bd\u03b1\u03bc\u03b5\u03c2. \u0397 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03c4\u03bf\u03bd\u03af\u03b6\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03ba\u03ac\u03b8\u03b5 relaxed acceptance \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ae \u03b5\u03bc\u03c0\u03b5\u03b9\u03c1\u03b9\u03ba\u03ae \u03b1\u03c0\u03cc\u03b4\u03b5\u03b9\u03be\u03b7 fidelity\u00b7 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03ac\u03b4\u03b5\u03b9\u03b1 \u03bd\u03b1 \u03c0\u03b1\u03c1\u03b1\u03bb\u03b5\u03af\u03c0\u03b5\u03c4\u03b1\u03b9 target computation.<\/p>\n<p>\u0397 survey \u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03b5\u03b9 vision-language, video-language, vision-language-action, speech\/audio \u03ba\u03b1\u03b9 autoregressive visual generation. \u03a4\u03bf \u03ba\u03bf\u03b9\u03bd\u03cc \u03bc\u03bf\u03c4\u03af\u03b2\u03bf \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 conditioning, candidate construction \u03ba\u03b1\u03b9 verification \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b5\u03be\u03b5\u03c4\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03bc\u03b1\u03b6\u03af. \u0391\u03c5\u03c4\u03cc \u03b5\u03af\u03bd\u03b1\u03b9 \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03bf \u03c4\u03cc\u03c3\u03bf \u03b3\u03b9\u03b1 <a href=\"https:\/\/twodots.gr\/multimodal-ai-agents-antilipsi-drasi\/\">multimodal agents \u03c0\u03bf\u03c5 \u03bc\u03b5\u03c4\u03b1\u03c4\u03c1\u03ad\u03c0\u03bf\u03c5\u03bd \u03b1\u03bd\u03c4\u03af\u03bb\u03b7\u03c8\u03b7 \u03c3\u03b5 \u03b4\u03c1\u03ac\u03c3\u03b7<\/a> \u03cc\u03c3\u03bf \u03ba\u03b1\u03b9 \u03b3\u03b9\u03b1 <a href=\"https:\/\/twodots.gr\/infinity-parser2-document-ai-epicheiriseis\/\">Document AI \u03c3\u03b5 \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03ac \u03ad\u03b3\u03b3\u03c1\u03b1\u03c6\u03b1<\/a>, \u03cc\u03c0\u03bf\u03c5 \u03b7 \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03b5\u03b9\u03ba\u03cc\u03bd\u03b1\u03c2 \u03ba\u03b1\u03b9 \u03b7 \u03b1\u03ba\u03c1\u03af\u03b2\u03b5\u03b9\u03b1 OCR \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03ba\u03c5\u03c1\u03b9\u03b1\u03c1\u03c7\u03ae\u03c3\u03bf\u03c5\u03bd \u03c3\u03c4\u03bf \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2.<\/p>\n<aside class=\"td-article-note\">\n<p><strong>Lossless \u03b4\u03b5\u03bd \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03b4\u03c9\u03c1\u03b5\u03ac\u03bd:<\/strong> \u03b7 exact \u03b5\u03c0\u03b1\u03bb\u03ae\u03b8\u03b5\u03c5\u03c3\u03b7 \u03c0\u03c1\u03bf\u03c3\u03c4\u03b1\u03c4\u03b5\u03cd\u03b5\u03b9 \u03c4\u03b7\u03bd \u03ba\u03b1\u03c4\u03b1\u03bd\u03bf\u03bc\u03ae \u03c4\u03bf\u03c5 target, \u03b1\u03bb\u03bb\u03ac \u03ba\u03ac\u03b8\u03b5 visual token, cache construction \u03ba\u03b1\u03b9 verification pass \u03b5\u03be\u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03b5\u03af \u03bd\u03b1 \u03ba\u03b1\u03c4\u03b1\u03bd\u03b1\u03bb\u03ce\u03bd\u03b5\u03b9 \u03c7\u03c1\u03cc\u03bd\u03bf \u03ba\u03b1\u03b9 \u03bc\u03bd\u03ae\u03bc\u03b7.<\/p>\n<\/aside>\n<h2 id=\"peirama-workloads\">\u03a4\u03bf \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1: \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1, \u03ad\u03be\u03b9 workloads \u03ba\u03b1\u03b9 \u03c3\u03c9\u03c3\u03c4\u03cc baseline<\/h2>\n<p>\u0397 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03c0\u03b5\u03c1\u03b9\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 \u03bc\u03b9\u03b1 suite \u03c4\u03b5\u03c3\u03c3\u03ac\u03c1\u03c9\u03bd vision-language targets: Qwen3-VL-4B, Qwen3-VL-8B, Qwen3.6-27B \u03ba\u03b1\u03b9 Qwen3.6-35A3B. \u039a\u03b1\u03bb\u03cd\u03c0\u03c4\u03b5\u03b9 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03ba\u03bb\u03af\u03bc\u03b1\u03ba\u03b5\u03c2, dense \u03ba\u03b1\u03b9 mixture-of-experts \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ad\u03c2, \u03ba\u03b1\u03b8\u03ce\u03c2 \u03ba\u03b1\u03b9 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ad\u03c2 multimodal training \u03ae alignment. \u03a9\u03c3\u03c4\u03cc\u03c3\u03bf, \u03bf \u03ba\u03cd\u03c1\u03b9\u03bf\u03c2 \u03c0\u03af\u03bd\u03b1\u03ba\u03b1\u03c2 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03b5\u03c3\u03bc\u03ac\u03c4\u03c9\u03bd \u03b5\u03bc\u03c6\u03b1\u03bd\u03af\u03b6\u03b5\u03b9 \u03c3\u03b5\u03b9\u03c1\u03ad\u03c2 \u03b3\u03b9\u03b1 \u03c4\u03b1 \u03b4\u03cd\u03bf Qwen3-VL targets \u03ba\u03b1\u03b9 \u03c4\u03bf Qwen3.6-27B, \u03cc\u03c7\u03b9 \u03b1\u03bd\u03c4\u03af\u03c3\u03c4\u03bf\u03b9\u03c7\u03b7 \u03c3\u03b5\u03b9\u03c1\u03ac \u03b3\u03b9\u03b1 \u03c4\u03bf 35A3B. \u0386\u03c1\u03b1 \u03c4\u03b1 \u03bc\u03b5\u03c4\u03c1\u03b7\u03bc\u03ad\u03bd\u03b1 \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03ac\u03c3\u03bc\u03b1\u03c4\u03b1 \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b5\u03c0\u03b5\u03ba\u03c4\u03b5\u03af\u03bd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c5\u03c4\u03cc\u03bc\u03b1\u03c4\u03b1 \u03c3\u03b5 \u03cc\u03bb\u03b7 \u03c4\u03b7 \u03b4\u03b7\u03bb\u03c9\u03bc\u03ad\u03bd\u03b7 suite.<\/p>\n<p>\u03a4\u03bf \u03b2\u03b1\u03c3\u03b9\u03ba\u03cc evaluation set \u03b5\u03af\u03c7\u03b5 600 samples: 100 \u03b1\u03c0\u03cc \u03ba\u03b1\u03b8\u03b5\u03bc\u03af\u03b1 \u03b1\u03c0\u03cc \u03ad\u03be\u03b9 \u03ba\u03b1\u03c4\u03b7\u03b3\u03bf\u03c1\u03af\u03b5\u03c2. \u03a0\u03b5\u03c1\u03b9\u03bb\u03ac\u03bc\u03b2\u03b1\u03bd\u03b5 GQA \u03b3\u03b9\u03b1 \u03b3\u03b5\u03bd\u03b9\u03ba\u03cc VQA, Flickr30K \u03b3\u03b9\u03b1 captioning, TextVQA \u03b3\u03b9\u03b1 \u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf \u03bc\u03ad\u03c3\u03b1 \u03c3\u03b5 \u03b5\u03b9\u03ba\u03cc\u03bd\u03b5\u03c2, CharXiv \u03b3\u03b9\u03b1 charts, MMMU \u03b3\u03b9\u03b1 \u03c3\u03cd\u03bd\u03b8\u03b5\u03c4\u03bf multimodal reasoning \u03ba\u03b1\u03b9 multi-turn \u03c3\u03c5\u03bd\u03bf\u03bc\u03b9\u03bb\u03af\u03b5\u03c2 \u03b1\u03c0\u03cc ConvBench \u03ba\u03b1\u03b9 MM-MT-Bench. \u03a7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03ae\u03b8\u03b7\u03ba\u03b5 seed 42 \u03ba\u03b1\u03b9, \u03b5\u03ba\u03c4\u03cc\u03c2 \u03b1\u03c0\u03cc \u03c3\u03b7\u03bc\u03b5\u03b9\u03c9\u03bc\u03ad\u03bd\u03b5\u03c2 \u03b5\u03be\u03b1\u03b9\u03c1\u03ad\u03c3\u03b5\u03b9\u03c2, matched greedy decoding \u03c3\u03c4\u03bf SGLang \u03bc\u03b5 \u03ad\u03c9\u03c2 2.048 \u03bd\u03ad\u03b1 tokens.<\/p>\n<p>\u0397 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c4\u03bf\u03c5 draft \u03bc\u03b5\u03c4\u03c1\u03ae\u03b8\u03b7\u03ba\u03b5 \u03bc\u03b5 mean accepted tokens (MAT), \u03b4\u03b7\u03bb\u03b1\u03b4\u03ae \u03c0\u03cc\u03c3\u03b1 output tokens \u03c0\u03c1\u03bf\u03c7\u03c9\u03c1\u03bf\u03cd\u03bd \u03ba\u03b1\u03c4\u03ac \u03bc\u03ad\u03c3\u03bf \u03cc\u03c1\u03bf \u03c3\u03b5 \u03ba\u03ac\u03b8\u03b5 speculative step. \u0397 \u03b2\u03b1\u03c3\u03b9\u03ba\u03ae \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7 \u03b1\u03c0\u03bf\u03b4\u03bf\u03c4\u03b9\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2 \u03ae\u03c4\u03b1\u03bd end-to-end wall-clock speedup. \u03a0\u03b5\u03c1\u03b9\u03bb\u03b1\u03bc\u03b2\u03ac\u03bd\u03b5\u03b9 vision encoding, target \u03ba\u03b1\u03b9 drafter prefill, candidate generation, tree construction, verification \u03ba\u03b1\u03b9 sampling. \u0391\u03c5\u03c4\u03cc \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf baseline \u03b3\u03b9\u03b1 \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd: \u03af\u03b4\u03b9\u03bf hardware, \u03af\u03b4\u03b9\u03bf backend, \u03af\u03b4\u03b9\u03b1 prompts, \u03af\u03b4\u03b9\u03bf batch \u03ba\u03b1\u03b9 \u03af\u03b4\u03b9\u03b1 stopping criteria.<\/p>\n<p>\u03a5\u03c0\u03ac\u03c1\u03c7\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03b7 \u03b5\u03be\u03b1\u03af\u03c1\u03b5\u03c3\u03b7 \u03c3\u03c4\u03b7 \u03c3\u03cd\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7. \u03a4\u03b1 DFlash \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03b3\u03b9\u03b1 Qwen3-VL-4B \u03ba\u03b1\u03b9 8B \u03c3\u03b7\u03bc\u03b5\u03b9\u03ce\u03bd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c9\u03c2 HuggingFace runs, \u03b5\u03bd\u03ce \u03bf\u03b9 \u03c5\u03c0\u03cc\u03bb\u03bf\u03b9\u03c0\u03b5\u03c2 \u03c3\u03b5\u03b9\u03c1\u03ad\u03c2 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03bf\u03cd\u03bd SGLang. \u039a\u03ac\u03b8\u03b5 speedup \u03ba\u03b1\u03bd\u03bf\u03bd\u03b9\u03ba\u03bf\u03c0\u03bf\u03b9\u03b5\u03af\u03c4\u03b1\u03b9 \u03c3\u03c4\u03bf \u03b4\u03b9\u03ba\u03cc \u03c4\u03bf\u03c5 backend-matched autoregressive baseline, \u03b1\u03bb\u03bb\u03ac \u03c4\u03bf \u03b1\u03c0\u03cc\u03bb\u03c5\u03c4\u03bf latency \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03bf \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03c4\u03b1 frameworks.<\/p>\n<h2 id=\"dflash-dspark-apotelesmata\">DFlash \u03ba\u03b1\u03b9 DSpark: \u03c4\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03bc\u03b5 \u03c4\u03bf \u03c3\u03c9\u03c3\u03c4\u03cc \u03c0\u03bb\u03b1\u03af\u03c3\u03b9\u03bf<\/h2>\n<p>\u03a3\u03c4\u03bf Qwen3.6-27B \u03ba\u03b1\u03b9 \u03c3\u03c4\u03bf matched SGLang setup, \u03c4\u03bf DFlash \u03b5\u03af\u03c7\u03b5 \u03c4\u03bf \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03c3\u03b5 \u03ba\u03ac\u03b8\u03b5 benchmark: task-equal MAT 4,38 \u03ba\u03b1\u03b9 \u03bc\u03ad\u03c3\u03b7 \u03b5\u03c0\u03b9\u03c4\u03ac\u03c7\u03c5\u03bd\u03c3\u03b7 2,60\u00d7 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 autoregressive decoding. \u03a4\u03bf DSpark \u03b1\u03ba\u03bf\u03bb\u03bf\u03cd\u03b8\u03b7\u03c3\u03b5 \u03bc\u03b5 MAT 3,52 \u03ba\u03b1\u03b9 2,04\u00d7. \u03a4\u03bf MTP \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 1,64\u00d7 \u03ba\u03b1\u03b9 \u03c4\u03bf EAGLE-3 1,69\u00d7. \u03a3\u03c4\u03b7 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7 \u03b4\u03b9\u03b1\u03bc\u03cc\u03c1\u03c6\u03c9\u03c3\u03b7, \u03bb\u03bf\u03b9\u03c0\u03cc\u03bd, \u03c4\u03bf L2 \u03ad\u03b4\u03c9\u03c3\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc end-to-end \u03ba\u03ad\u03c1\u03b4\u03bf\u03c2 \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03bf \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf accepted length.<\/p>\n<div class=\"td-chart td-chart--metrics\">\n<div class=\"td-chart-head\">\n<p class=\"td-chart-title\">\u03a4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b9\u03c2 \u03b1\u03c1\u03b9\u03b8\u03bc\u03bf\u03af \u03c0\u03bf\u03c5 \u03bf\u03c1\u03af\u03b6\u03bf\u03c5\u03bd \u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03cc\u03c6\u03b5\u03bb\u03bf\u03c2<\/p>\n<p class=\"td-chart-subtitle\">\u039f\u03b9 \u03b5\u03c0\u03b9\u03c4\u03b1\u03c7\u03cd\u03bd\u03c3\u03b5\u03b9\u03c2 \u03b1\u03c6\u03bf\u03c1\u03bf\u03cd\u03bd Qwen3.6-27B \u03bc\u03b5 backend-matched SGLang baseline\u00b7 \u03bf\u03b9 \u03c7\u03c1\u03cc\u03bd\u03bf\u03b9 \u03c0\u03c1\u03bf\u03ad\u03c1\u03c7\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03b1\u03bd\u03c4\u03b9\u03c0\u03c1\u03bf\u03c3\u03c9\u03c0\u03b5\u03c5\u03c4\u03b9\u03ba\u03ae stage-level \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03c4\u03b7\u03c2 \u03af\u03b4\u03b9\u03b1\u03c2 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1\u03c2.<\/p>\n<\/div>\n<div class=\"td-chart-body\">\n<div class=\"td-metric-grid\">\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">2,60\u00d7<\/span><span class=\"td-metric-label\">DFlash end-to-end speedup<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">2,04\u00d7<\/span><span class=\"td-metric-label\">DSpark end-to-end speedup<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">1,14 ms<\/span><span class=\"td-metric-label\">DFlash draft generation<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">26,45\u201327,40 ms<\/span><span class=\"td-metric-label\">Vision conditioning \u03c3\u03c4\u03b1 MTP, DSpark \u03ba\u03b1\u03b9 DFlash<\/span><\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>\u03a3\u03c4\u03b1 Qwen3-VL tests \u03c4\u03bf\u03c5 HuggingFace, \u03c4\u03bf DFlash \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 MAT 2,50 \u03ba\u03b1\u03b9 1,95\u00d7 \u03c3\u03c4\u03bf 4B, \u03ba\u03b1\u03b9 MAT 2,75 \u03bc\u03b5 2,14\u00d7 \u03c3\u03c4\u03bf 8B. \u0391\u03bd\u03c4\u03af\u03b8\u03b5\u03c4\u03b1, \u03bf\u03b9 SGLang runs \u03c4\u03bf\u03c5 EAGLE-3 \u03ad\u03bc\u03b5\u03b9\u03bd\u03b1\u03bd \u03ba\u03ac\u03c4\u03c9 \u03b1\u03c0\u03cc \u03c4\u03b1 \u03b1\u03bd\u03c4\u03af\u03c3\u03c4\u03bf\u03b9\u03c7\u03b1 autoregressive baselines: 0,71\u00d7 \u03ba\u03b1\u03b9 0,88\u00d7. \u0391\u03c5\u03c4\u03cc \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03ba\u03ac\u03b8\u03b5 DFlash deployment \u03b8\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03b1\u03c7\u03cd\u03c4\u03b5\u03c1\u03bf \u03ae \u03cc\u03c4\u03b9 \u03ba\u03ac\u03b8\u03b5 EAGLE-3 deployment \u03b8\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c1\u03b3\u03cc\u03c4\u03b5\u03c1\u03bf\u00b7 \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03c2, checkpoint, target, task \u03ba\u03b1\u03b9 backend \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03bf\u03cd\u03bd \u03b5\u03bd\u03b9\u03b1\u03af\u03b1 \u03b4\u03b9\u03b1\u03bc\u03cc\u03c1\u03c6\u03c9\u03c3\u03b7.<\/p>\n<p>\u039f \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf\u03c2 drafter \u03b4\u03b5\u03bd \u03b5\u03b3\u03b3\u03c5\u03ac\u03c4\u03b1\u03b9 \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03af\u03b1. \u03a4\u03bf DFlash \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 1,73 \u03b4\u03b9\u03c3. \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03c4\u03c1\u03bf\u03c5\u03c2, \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 0,43 \u03b4\u03b9\u03c3. \u03b3\u03b9\u03b1 \u03c4\u03bf native MTP module \u03ba\u03b1\u03b9 0,60 \u03b4\u03b9\u03c3. \u03b3\u03b9\u03b1 \u03c4\u03bf EAGLE-3, \u03b1\u03bb\u03bb\u03ac \u03c4\u03bf EAGLE-3 \u03b4\u03b5\u03bd \u03be\u03b5\u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03c4\u03bf \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf MTP module. \u0391\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae, objective \u03ba\u03b1\u03b9 checkpoint provenance \u03bc\u03b5\u03c4\u03c1\u03bf\u03cd\u03bd \u03bc\u03b1\u03b6\u03af \u03bc\u03b5 \u03c4\u03b7\u03bd \u03ba\u03bb\u03af\u03bc\u03b1\u03ba\u03b1. \u0388\u03bd\u03b1 third-party DSpark checkpoint \u03c5\u03c3\u03c4\u03ad\u03c1\u03b7\u03c3\u03b5 \u03c4\u03bf\u03c5 DFlash \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b1 \u03ad\u03be\u03b9 tasks, \u03c0\u03b1\u03c1\u03cc\u03c4\u03b9 \u03c4\u03bf DSpark \u03b5\u03af\u03c7\u03b5 \u03b9\u03c3\u03c7\u03c5\u03c1\u03ac text-only \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03c3\u03c4\u03b7\u03bd \u03b1\u03c1\u03c7\u03b9\u03ba\u03ae \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03c4\u03bf\u03c5.<\/p>\n<h2 id=\"megalitero-target\">\u03a0\u03cc\u03c4\u03b5 \u03c4\u03bf \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf target \u03b1\u03c0\u03bf\u03c3\u03b2\u03ad\u03bd\u03b5\u03b9 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b1 \u03c4\u03bf\u03bd drafter<\/h2>\n<p>\u03a3\u03c4\u03b9\u03c2 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b5\u03c2 \u03b4\u03b9\u03b1\u03bc\u03bf\u03c1\u03c6\u03ce\u03c3\u03b5\u03b9\u03c2, \u03c4\u03bf DFlash \u03b1\u03bd\u03ad\u03b2\u03b7\u03ba\u03b5 \u03b1\u03c0\u03cc 1,95\u00d7 \u03c3\u03c4\u03bf Qwen3-VL-4B \u03c3\u03b5 2,14\u00d7 \u03c3\u03c4\u03bf 8B \u03ba\u03b1\u03b9 2,60\u00d7 \u03c3\u03c4\u03bf Qwen3.6-27B. \u03a4\u03bf MAT \u03b1\u03c5\u03be\u03ae\u03b8\u03b7\u03ba\u03b5 \u03b1\u03bd\u03c4\u03af\u03c3\u03c4\u03bf\u03b9\u03c7\u03b1 \u03b1\u03c0\u03cc 2,50 \u03ba\u03b1\u03b9 2,75 \u03c3\u03b5 4,38. \u0397 \u03b5\u03c1\u03bc\u03b7\u03bd\u03b5\u03af\u03b1 \u03c4\u03c9\u03bd \u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03ad\u03c9\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b4\u03b9\u03c0\u03bb\u03ae: \u03ad\u03bd\u03b1\u03c2 \u03c1\u03b7\u03c7\u03cc\u03c2, \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 \u03c3\u03c4\u03b1\u03b8\u03b5\u03c1\u03bf\u03cd \u03ba\u03cc\u03c3\u03c4\u03bf\u03c5\u03c2 drafter \u03b1\u03c0\u03bf\u03c3\u03b2\u03ad\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b5\u03c5\u03ba\u03bf\u03bb\u03cc\u03c4\u03b5\u03c1\u03b1 \u03cc\u03c4\u03b1\u03bd \u03c4\u03bf target forward \u03b3\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b1\u03ba\u03c1\u03b9\u03b2\u03cc\u03c4\u03b5\u03c1\u03bf, \u03b5\u03bd\u03ce \u03c4\u03b1 \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b1 targets \u03b5\u03bd\u03b4\u03ad\u03c7\u03b5\u03c4\u03b1\u03b9 \u03bd\u03b1 \u03c0\u03b1\u03c1\u03ad\u03c7\u03bf\u03c5\u03bd \u03c0\u03b9\u03bf \u03c0\u03c1\u03bf\u03b2\u03bb\u03ad\u03c8\u03b9\u03bc\u03b1 features.<\/p>\n<p>\u0394\u03b5\u03bd \u03c0\u03c1\u03cc\u03ba\u03b5\u03b9\u03c4\u03b1\u03b9 \u03b3\u03b9\u03b1 \u03b3\u03b5\u03bd\u03b9\u03ba\u03cc \u03ba\u03b1\u03bd\u03cc\u03bd\u03b1 \u00ab\u03cc\u03c3\u03bf \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03c4\u03cc\u03c3\u03bf \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b1\u00bb. Model scale, checkpoint, training recipe, backend \u03ba\u03b1\u03b9 task \u03b1\u03bb\u03bb\u03ac\u03b6\u03bf\u03c5\u03bd \u03bc\u03b1\u03b6\u03af. \u0397 \u03af\u03b4\u03b9\u03b1 \u03b7 HTML \u03ad\u03ba\u03b4\u03bf\u03c3\u03b7 \u03c4\u03b7\u03c2 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1\u03c2 \u03c0\u03b5\u03c1\u03b9\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 \u03c4\u03bf Qwen3.6-27B \u03c9\u03c2 text-pretrained \u03ba\u03b1\u03b9 \u03ba\u03b1\u03c4\u03cc\u03c0\u03b9\u03bd multimodal-aligned \u03c3\u03c4\u03b7 \u03bc\u03b5\u03b8\u03bf\u03b4\u03bf\u03bb\u03bf\u03b3\u03af\u03b1, \u03b1\u03bb\u03bb\u03ac \u03b1\u03c1\u03b3\u03cc\u03c4\u03b5\u03c1\u03b1 \u03c4\u03bf \u03b1\u03c0\u03bf\u03ba\u03b1\u03bb\u03b5\u03af natively multimodal-trained \u03c3\u03c4\u03b7 \u03c3\u03c5\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03b5\u03c3\u03bc\u03ac\u03c4\u03c9\u03bd. \u0391\u03c5\u03c4\u03ae \u03b7 \u03b5\u03c3\u03c9\u03c4\u03b5\u03c1\u03b9\u03ba\u03ae \u03b1\u03c3\u03c5\u03bd\u03ad\u03c0\u03b5\u03b9\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b5\u03c0\u03b9\u03c0\u03bb\u03ad\u03bf\u03bd \u03bb\u03cc\u03b3\u03bf\u03c2 \u03bd\u03b1 \u03bc\u03b7\u03bd \u03b1\u03c0\u03bf\u03b4\u03bf\u03b8\u03b5\u03af \u03b1\u03b9\u03c4\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c3\u03c4\u03bf pretraining regime.<\/p>\n<p>\u0393\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03c5\u03c0\u03bf\u03b4\u03bf\u03bc\u03ce\u03bd, \u03c4\u03bf \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf \u03c3\u03c5\u03bc\u03c0\u03ad\u03c1\u03b1\u03c3\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03bd\u03b1 \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03af\u03b6\u03b5\u03b9 break-even \u03b1\u03bd\u03ac target \u03ba\u03b1\u03b9 workload. \u0388\u03bd\u03b1\u03c2 drafter \u03c0\u03bf\u03c5 \u03b5\u03c0\u03b9\u03b2\u03b1\u03c1\u03cd\u03bd\u03b5\u03b9 \u03ad\u03bd\u03b1 \u03bc\u03b9\u03ba\u03c1\u03cc \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03bf\u03b9\u03ba\u03bf\u03bd\u03bf\u03bc\u03b9\u03ba\u03cc\u03c2 \u03b4\u03af\u03c0\u03bb\u03b1 \u03c3\u03b5 \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf target, \u03b5\u03c6\u03cc\u03c3\u03bf\u03bd \u03c4\u03bf accepted block \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03b1\u03c1\u03ba\u03b5\u03c4\u03ac \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf \u03ba\u03b1\u03b9 \u03c4\u03bf conditioning \u03b4\u03b5\u03bd \u03b1\u03c5\u03be\u03ac\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b4\u03c5\u03c3\u03b1\u03bd\u03ac\u03bb\u03bf\u03b3\u03b1. \u0397 \u03af\u03b4\u03b9\u03b1 \u03bb\u03bf\u03b3\u03b9\u03ba\u03ae \u03b5\u03be\u03b7\u03b3\u03b5\u03af \u03b3\u03b9\u03b1\u03c4\u03af \u03ad\u03bd\u03b1 <a href=\"https:\/\/twodots.gr\/lfm25-dspark-taxytero-ai-inference\/\">\u03c4\u03b1\u03c7\u03cd\u03c4\u03b5\u03c1\u03bf AI inference stack \u03b4\u03b5\u03bd \u03b1\u03c0\u03b1\u03b9\u03c4\u03b5\u03af \u03c0\u03ac\u03bd\u03c4\u03b1 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf target model<\/a>.<\/p>\n<h2 id=\"bottleneck-prefill\">\u03a4\u03bf bottleneck \u03bc\u03b5\u03c4\u03b1\u03c6\u03ad\u03c1\u03b5\u03c4\u03b1\u03b9 \u03c3\u03c4\u03bf multimodal prefill<\/h2>\n<p>\u0397 latency decomposition \u03b5\u03af\u03bd\u03b1\u03b9 \u03af\u03c3\u03c9\u03c2 \u03c4\u03bf \u03c0\u03b9\u03bf \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03b7\u03c2 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7\u03c2. \u03a4\u03bf DFlash \u03b4\u03b1\u03c0\u03ac\u03bd\u03b7\u03c3\u03b5 1,14 ms \u03c3\u03c4\u03bf drafting \u03ba\u03b1\u03b9 \u03c4\u03bf DSpark 1,78 ms. \u03a4\u03b7\u03bd \u03af\u03b4\u03b9\u03b1 \u03c3\u03c4\u03b9\u03b3\u03bc\u03ae, MTP, DSpark \u03ba\u03b1\u03b9 DFlash \u03b4\u03b1\u03c0\u03ac\u03bd\u03b7\u03c3\u03b1\u03bd 26,45\u201327,40 ms \u03c3\u03c4\u03bf vision-conditioning stage. \u03a4\u03bf L2 \u03bc\u03b5\u03b9\u03ce\u03bd\u03b5\u03b9 \u03b4\u03c1\u03b1\u03c3\u03c4\u03b9\u03ba\u03ac \u03c4\u03bf \u03c3\u03b5\u03b9\u03c1\u03b9\u03b1\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03b4\u03b7\u03bc\u03b9\u03bf\u03c5\u03c1\u03b3\u03af\u03b1\u03c2 \u03c4\u03bf\u03c5 draft, \u03b1\u03bb\u03bb\u03ac \u03b4\u03b5\u03bd \u03b5\u03be\u03b1\u03bb\u03b5\u03af\u03c6\u03b5\u03b9 \u03c4\u03bf \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03c0\u03c1\u03bf\u03b5\u03c4\u03bf\u03b9\u03bc\u03b1\u03c3\u03af\u03b1\u03c2 \u03c4\u03bf\u03c5 multimodal condition.<\/p>\n<p>\u03a4\u03bf DFlash \u03c0\u03c1\u03bf\u03b2\u03ac\u03bb\u03bb\u03b5\u03b9 target hidden states \u03c3\u03c4\u03bf KV cache \u03c4\u03bf\u03c5 drafter. \u0391\u03c5\u03c4\u03cc \u03b2\u03b5\u03bb\u03c4\u03b9\u03ce\u03bd\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b5\u03c5\u03b8\u03c5\u03b3\u03c1\u03ac\u03bc\u03bc\u03b9\u03c3\u03b7 \u03bc\u03b5 \u03c4\u03bf target, \u03cc\u03bc\u03c9\u03c2 \u03c0\u03c1\u03bf\u03c3\u03b8\u03ad\u03c4\u03b5\u03b9 prefill-like \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03c0\u03bf\u03c5 \u03c3\u03c5\u03bd\u03b4\u03ad\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c4\u03bf \u03bc\u03ae\u03ba\u03bf\u03c2 \u03c4\u03b7\u03c2 visual sequence. \u039c\u03b5 \u03b1\u03c0\u03bb\u03ac \u03bb\u03cc\u03b3\u03b9\u03b1, \u03c4\u03bf \u03b5\u03c0\u03cc\u03bc\u03b5\u03bd\u03bf draft token \u03b3\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03c0\u03bf\u03bb\u03cd \u03c6\u03b8\u03b7\u03bd\u03cc, \u03b1\u03bb\u03bb\u03ac \u03c4\u03bf \u03b5\u03c0\u03cc\u03bc\u03b5\u03bd\u03bf visual token \u03cc\u03c7\u03b9.<\/p>\n<p>\u0397 \u03c3\u03c9\u03c3\u03c4\u03ae \u03ba\u03b1\u03c4\u03b5\u03cd\u03b8\u03c5\u03bd\u03c3\u03b7 optimization \u03b5\u03c0\u03bf\u03bc\u03ad\u03bd\u03c9\u03c2 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03bc\u03cc\u03bd\u03bf \u03ad\u03bd\u03b1\u03c2 \u03b1\u03ba\u03cc\u03bc\u03b7 \u03c4\u03b1\u03c7\u03cd\u03c4\u03b5\u03c1\u03bf\u03c2 draft forward pass. \u0395\u03af\u03bd\u03b1\u03b9 compressed visual tokens, \u03b5\u03c0\u03b1\u03bd\u03b1\u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 target features, shared caches \u03ba\u03b1\u03b9 \u03b5\u03bb\u03b1\u03c6\u03c1\u03cd\u03c4\u03b5\u03c1\u03bf\u03b9 adapters. \u03a3\u03b5 OCR, catalog search \u03ae <a href=\"https:\/\/twodots.gr\/ai-agents-etairika-eggrafa-domi-anagnosi-proza-syggrafi\/\">AI agents \u03b3\u03b9\u03b1 \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03ac \u03ad\u03b3\u03b3\u03c1\u03b1\u03c6\u03b1<\/a>, \u03b7 \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03c3\u03b5\u03bb\u03af\u03b4\u03b1\u03c2 \u03ba\u03b1\u03b9 \u03b7 \u03b4\u03b9\u03b1\u03c4\u03ae\u03c1\u03b7\u03c3\u03b7 grounding \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03c0\u03b7\u03c1\u03b5\u03ac\u03b6\u03bf\u03c5\u03bd \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf \u03c4\u03bf \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc latency \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03b5 \u03b4\u03cd\u03bf L2 algorithms.<\/p>\n<h2 id=\"analysi-8k\">\u0393\u03b9\u03b1\u03c4\u03af \u03b7 \u03c5\u03c8\u03b7\u03bb\u03ae \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c6\u03ad\u03c1\u03b5\u03b9 \u03b5\u03c0\u03b9\u03b2\u03c1\u03ac\u03b4\u03c5\u03bd\u03c3\u03b7<\/h2>\n<p>\u03a3\u03c4\u03bf HR-Bench, \u03c4\u03bf MAT \u03c0\u03b1\u03c1\u03ad\u03bc\u03b5\u03b9\u03bd\u03b5 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 2,5\u20133,2 \u03c3\u03b5 4K \u03ba\u03b1\u03b9 8K, \u03b1\u03bb\u03bb\u03ac \u03c4\u03bf speedup \u03ba\u03b1\u03c4\u03ad\u03c1\u03c1\u03b5\u03c5\u03c3\u03b5. \u03a3\u03c4\u03b1 4K \u03bf\u03b9 \u03c4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b9\u03c2 \u03c4\u03b9\u03bc\u03ad\u03c2 \u03ae\u03c4\u03b1\u03bd \u03bf\u03c1\u03b9\u03b1\u03ba\u03ad\u03c2, \u03b1\u03c0\u03cc 0,96\u00d7 \u03ad\u03c9\u03c2 1,11\u00d7. \u03a3\u03c4\u03b1 8K \u03cc\u03bb\u03b5\u03c2 \u03bf\u03b9 \u03b4\u03b9\u03b1\u03bc\u03bf\u03c1\u03c6\u03ce\u03c3\u03b5\u03b9\u03c2 \u03b2\u03c1\u03ad\u03b8\u03b7\u03ba\u03b1\u03bd \u03c3\u03c4\u03bf \u03ae \u03ba\u03ac\u03c4\u03c9 \u03b1\u03c0\u03cc \u03c4\u03bf autoregressive baseline: 0,90\u00d7 \u03ba\u03b1\u03b9 0,83\u00d7 \u03c3\u03c4\u03bf Qwen3-VL-4B, 0,99\u00d7 \u03ba\u03b1\u03b9 0,85\u00d7 \u03c3\u03c4\u03bf 8B.<\/p>\n<p>\u03a4\u03bf \u03c0\u03b9\u03bf \u03b1\u03c0\u03bf\u03ba\u03b1\u03bb\u03c5\u03c0\u03c4\u03b9\u03ba\u03cc \u03c0\u03b1\u03c1\u03ac\u03b4\u03b5\u03b9\u03b3\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf Qwen3-VL-8B \u03bc\u03b5 sampling. \u03a4\u03bf MAT \u03b1\u03c5\u03be\u03ae\u03b8\u03b7\u03ba\u03b5 \u03b1\u03c0\u03cc 2,88 \u03c3\u03b5 3,22 \u03cc\u03c4\u03b1\u03bd \u03b7 \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03c0\u03ae\u03b3\u03b5 \u03b1\u03c0\u03cc 4K \u03c3\u03b5 8K, \u03cc\u03bc\u03c9\u03c2 \u03c4\u03bf speedup \u03ad\u03c0\u03b5\u03c3\u03b5 \u03b1\u03c0\u03cc 1,11\u00d7 \u03c3\u03b5 0,85\u00d7. \u039f drafter \u03c0\u03c1\u03cc\u03c4\u03b5\u03b9\u03bd\u03b5 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b1 \u03b1\u03c0\u03bf\u03b4\u03b5\u03ba\u03c4\u03ac blocks, \u03b1\u03bb\u03bb\u03ac \u03c4\u03bf \u03c0\u03c1\u03cc\u03c3\u03b8\u03b5\u03c4\u03bf visual prefill \u03ba\u03b1\u03b9 \u03b7 cache construction \u03ba\u03cc\u03c3\u03c4\u03b9\u03c3\u03b1\u03bd \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf \u03b1\u03c0\u03cc \u03cc,\u03c4\u03b9 \u03ba\u03ad\u03c1\u03b4\u03b9\u03c3\u03b5 \u03c4\u03bf decoding.<\/p>\n<aside class=\"td-article-note\">\n<p><strong>\u03a4\u03bf HR-Bench stress test \u03b5\u03af\u03c7\u03b5 \u03ba\u03bf\u03b9\u03bd\u03cc manifest 30 \u03b4\u03b5\u03b9\u03b3\u03bc\u03ac\u03c4\u03c9\u03bd \u03ba\u03b1\u03b9 \u03cc\u03c1\u03b9\u03bf 512 generated tokens.<\/strong> \u0395\u03af\u03bd\u03b1\u03b9 \u03ba\u03b1\u03b8\u03b1\u03c1\u03ae \u03ad\u03bd\u03b4\u03b5\u03b9\u03be\u03b7 failure mode, \u03cc\u03c7\u03b9 \u03c0\u03bb\u03ae\u03c1\u03b7\u03c2 \u03c0\u03c1\u03cc\u03b2\u03bb\u03b5\u03c8\u03b7 \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 \u03ba\u03ac\u03bc\u03b5\u03c1\u03b1, document corpus \u03ae batch profile.<\/p>\n<\/aside>\n<p>\u0393\u03b9\u03b1 production KPI, \u03c4\u03bf MAT \u03b5\u03af\u03bd\u03b1\u03b9 \u03b5\u03bd\u03b4\u03b9\u03ac\u03bc\u03b5\u03c3\u03bf metric \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1. \u0397 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bc\u03b5\u03c4\u03c1\u03ac p50 \u03ba\u03b1\u03b9 p95 end-to-end latency, throughput \u03c3\u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc concurrency, GPU memory, \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03b1\u03bd\u03ac \u03bf\u03bb\u03bf\u03ba\u03bb\u03b7\u03c1\u03c9\u03bc\u03ad\u03bd\u03bf \u03b1\u03af\u03c4\u03b7\u03bc\u03b1 \u03ba\u03b1\u03b9 quality regressions. \u0388\u03bd\u03b1 \u03c5\u03c8\u03b7\u03bb\u03cc acceptance metric \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c3\u03c5\u03bd\u03c5\u03c0\u03ac\u03c1\u03c7\u03b5\u03b9 \u03bc\u03b5 \u03c7\u03b5\u03b9\u03c1\u03cc\u03c4\u03b5\u03c1\u03b7 \u03b5\u03bc\u03c0\u03b5\u03b9\u03c1\u03af\u03b1 \u03c7\u03c1\u03ae\u03c3\u03c4\u03b7.<\/p>\n<h2 id=\"visual-context-drafter\">\u03a0\u03cc\u03c3\u03bf visual context \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac \u03bf drafter<\/h2>\n<p>\u0397 ablation \u03bc\u03b5 DFlash \u03ad\u03b4\u03c9\u03c3\u03b5 \u03bc\u03b9\u03b1 \u03b5\u03bd\u03b4\u03b9\u03b1\u03c6\u03ad\u03c1\u03bf\u03c5\u03c3\u03b1 \u03ad\u03bd\u03b4\u03b5\u03b9\u03be\u03b7. \u038c\u03c4\u03b1\u03bd \u03b1\u03c6\u03b1\u03b9\u03c1\u03ad\u03b8\u03b7\u03ba\u03b1\u03bd \u03c4\u03b1 visual-token K\/V entries \u03b1\u03c0\u03cc \u03c4\u03bf context \u03c4\u03bf\u03c5 drafter, \u03c4\u03bf \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc MAT \u03bc\u03b5\u03b9\u03ce\u03b8\u03b7\u03ba\u03b5 4,2% \u03c3\u03c4\u03bf Qwen3-VL-4B \u03ba\u03b1\u03b9 3,9% \u03c3\u03c4\u03bf 8B. \u038c\u03c4\u03b1\u03bd \u03b1\u03c6\u03b1\u03b9\u03c1\u03ad\u03b8\u03b7\u03ba\u03b5 \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03bf \u03c4\u03bf \u03b1\u03c1\u03c7\u03b9\u03ba\u03cc prefill context, \u03b7 \u03bc\u03b5\u03af\u03c9\u03c3\u03b7 \u03ae\u03c4\u03b1\u03bd 14,1% \u03ba\u03b1\u03b9 11,0% \u03b1\u03bd\u03c4\u03af\u03c3\u03c4\u03bf\u03b9\u03c7\u03b1. \u03a4\u03bf target \u03c3\u03c5\u03bd\u03ad\u03c7\u03b9\u03c3\u03b5 \u03c0\u03ac\u03bd\u03c4\u03b1 \u03bd\u03b1 \u03b2\u03bb\u03ad\u03c0\u03b5\u03b9 \u03c4\u03bf \u03c0\u03bb\u03ae\u03c1\u03b5\u03c2 multimodal input.<\/p>\n<p>\u03a3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03cc \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03c9\u03bd \u03b1\u03c0\u03bf\u03b4\u03b5\u03ba\u03c4\u03ce\u03bd tokens \u03c6\u03b1\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03ad\u03c4\u03c3\u03b9 \u03bd\u03b1 \u03c0\u03c1\u03bf\u03b2\u03bb\u03ad\u03c0\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc local continuation statistics \u03ba\u03b1\u03b9 \u03c0\u03c1\u03cc\u03c3\u03c6\u03b1\u03c4\u03b1 contextual states. \u0397 \u03b5\u03c5\u03b1\u03b9\u03c3\u03b8\u03b7\u03c3\u03af\u03b1 \u03cc\u03bc\u03c9\u03c2 \u03ae\u03c4\u03b1\u03bd task-specific: \u03c4\u03bf TextVQA \u03b5\u03c0\u03b7\u03c1\u03b5\u03ac\u03c3\u03c4\u03b7\u03ba\u03b5 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf, \u03b5\u03b9\u03b4\u03b9\u03ba\u03ac \u03c3\u03c4\u03bf 8B \u03cc\u03c0\u03bf\u03c5 \u03c4\u03bf MAT \u03ad\u03c0\u03b5\u03c3\u03b5 \u03b1\u03c0\u03cc 2,750 \u03c3\u03b5 1,953 \u03c7\u03c9\u03c1\u03af\u03c2 prefill context. \u03a4\u03bf MMMU \u03ad\u03bc\u03b5\u03b9\u03bd\u03b5 \u03c0\u03b9\u03bf \u03c3\u03c4\u03b1\u03b8\u03b5\u03c1\u03cc \u03ba\u03b1\u03b9 \u03c4\u03bf GQA \u03b4\u03b5\u03bd \u03b5\u03bc\u03c6\u03ac\u03bd\u03b9\u03c3\u03b5 \u03c3\u03c5\u03bd\u03b5\u03c0\u03ae \u03bc\u03bf\u03bd\u03bf\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae \u03c0\u03c4\u03ce\u03c3\u03b7.<\/p>\n<p>\u0391\u03c5\u03c4\u03cc \u03b1\u03bd\u03bf\u03af\u03b3\u03b5\u03b9 \u03b4\u03c1\u03cc\u03bc\u03bf \u03b3\u03b9\u03b1 selective conditioning. \u0388\u03bd\u03b1\u03c2 router \u03b8\u03b1 \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03c3\u03b5 \u03bd\u03b1 \u03b5\u03bd\u03b5\u03c1\u03b3\u03bf\u03c0\u03bf\u03b9\u03b5\u03af \u03c0\u03bb\u03ae\u03c1\u03b7 visual \u03c0\u03c1\u03cc\u03c3\u03b2\u03b1\u03c3\u03b7 \u03bc\u03cc\u03bd\u03bf \u03cc\u03c4\u03b1\u03bd \u03b7 \u03b1\u03b2\u03b5\u03b2\u03b1\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1, \u03c4\u03bf OCR \u03ae \u03c4\u03bf task \u03c4\u03bf \u03b1\u03c0\u03b1\u03b9\u03c4\u03b5\u03af. \u03a0\u03c1\u03cc\u03ba\u03b5\u03b9\u03c4\u03b1\u03b9 \u03b3\u03b9\u03b1 \u03c5\u03c0\u03bf\u03c3\u03c7\u03cc\u03bc\u03b5\u03bd\u03b7 \u03ba\u03b1\u03c4\u03b5\u03cd\u03b8\u03c5\u03bd\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03b3\u03b9\u03b1 \u03ae\u03b4\u03b7 \u03b1\u03c0\u03bf\u03b4\u03b5\u03b4\u03b5\u03b9\u03b3\u03bc\u03ad\u03bd\u03bf production recipe: \u03ba\u03ac\u03b8\u03b5 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03ae \u03b1\u03c6\u03b1\u03af\u03c1\u03b5\u03c3\u03b7 context \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b5\u03c0\u03b1\u03bd\u03b5\u03bb\u03ad\u03b3\u03c7\u03b5\u03c4\u03b1\u03b9 \u03c9\u03c2 \u03c0\u03c1\u03bf\u03c2 grounding \u03ba\u03b1\u03b9 task quality.<\/p>\n<h2 id=\"inference-ecosystem\">\u03a4\u03bf inference ecosystem \u03c0\u03c1\u03bf\u03c7\u03c9\u03c1\u03ac, \u03b1\u03bb\u03bb\u03ac \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03ac\u03bd\u03b9\u03c3\u03bf<\/h2>\n<p>\u0397 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 DFlash checkpoints \u03b3\u03b9\u03b1 Qwen3, Qwen3.5\/3.6 \u03ba\u03b1\u03b9 Gemma, \u03bc\u03b5 \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ad\u03c2 \u03bc\u03ad\u03c3\u03c9 MLX, SGLang \u03ba\u03b1\u03b9 vLLM. \u03a4\u03bf DeepSpec \u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03b5\u03b9 training \u03ba\u03b1\u03b9 evaluation \u03b3\u03b9\u03b1 DFlash \u03ba\u03b1\u03b9 DSpark, \u03c4\u03bf AngelSpec \u03c0\u03b5\u03c1\u03b9\u03bb\u03b1\u03bc\u03b2\u03ac\u03bd\u03b5\u03b9 DFly, DFlash, DFlare \u03ba\u03b1\u03b9 DSpark, \u03b5\u03bd\u03ce Speculators \u03ba\u03b1\u03b9 SpecForge \u03c3\u03c5\u03bd\u03b4\u03ad\u03bf\u03c5\u03bd \u03c4\u03b7\u03bd \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7 \u03bc\u03b5 vLLM \u03ba\u03b1\u03b9 SGLang \u03b1\u03bd\u03c4\u03af\u03c3\u03c4\u03bf\u03b9\u03c7\u03b1.<\/p>\n<p>\u03a4\u03b1 vLLM \u03ba\u03b1\u03b9 SGLang \u03c0\u03b1\u03c1\u03bf\u03c5\u03c3\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c9\u03c2 \u03bf\u03b9 \u03c0\u03bb\u03b7\u03c1\u03ad\u03c3\u03c4\u03b5\u03c1\u03b5\u03c2 production-oriented \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ad\u03c2 \u03b3\u03b9\u03b1 block-parallel speculative decoding. \u0397 \u03b5\u03c0\u03af\u03c3\u03b7\u03bc\u03b7 \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03af\u03c9\u03c3\u03b7 \u03c4\u03bf\u03c5 vLLM Speculators \u03c0\u03b5\u03c1\u03b9\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 DFlash \u03ba\u03b1\u03b9 DSpark \u03c9\u03c2 \u03b5\u03bd\u03b5\u03c1\u03b3\u03ac \u03b5\u03be\u03b5\u03bb\u03b9\u03c3\u03c3\u03cc\u03bc\u03b5\u03bd\u03b5\u03c2 \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ad\u03c2, \u03b5\u03bd\u03ce \u03c4\u03bf dynamic speculative decoding \u03bc\u03b5\u03b9\u03ce\u03bd\u03b5\u03b9 \u03c4\u03bf K \u03cc\u03c4\u03b1\u03bd batch size \u03ae confidence \u03ba\u03ac\u03bd\u03bf\u03c5\u03bd \u03c4\u03bf verification \u03b1\u03ba\u03c1\u03b9\u03b2\u03cc. \u0391\u03c5\u03c4\u03cc \u03c3\u03c5\u03bc\u03c0\u03bb\u03b7\u03c1\u03ce\u03bd\u03b5\u03b9 \u03c4\u03b7 \u03b3\u03b5\u03bd\u03b9\u03ba\u03cc\u03c4\u03b5\u03c1\u03b7 \u03bc\u03b5\u03c4\u03ac\u03b2\u03b1\u03c3\u03b7 \u03c4\u03c9\u03bd <a href=\"https:\/\/twodots.gr\/vllm-transformers-native-speed-ai-inference\/\">vLLM \u03ba\u03b1\u03b9 Transformers \u03c3\u03b5 native-speed AI inference<\/a>.<\/p>\n<p>\u0397 \u03c9\u03c1\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c4\u03bf\u03c5 algorithm \u03ba\u03b1\u03b9 \u03b7 \u03c9\u03c1\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c4\u03bf\u03c5 stack \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf \u03af\u03b4\u03b9\u03bf. \u0388\u03bd\u03b1 paper checkpoint \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03bc\u03b7 \u03bc\u03b5\u03c4\u03b1\u03c6\u03ad\u03c1\u03b5\u03c4\u03b1\u03b9 \u03b5\u03cd\u03ba\u03bf\u03bb\u03b1 \u03c3\u03c4\u03bf backend, quantization, batching, observability \u03ae hardware \u03c0\u03bf\u03c5 \u03b1\u03c0\u03b1\u03b9\u03c4\u03b5\u03af \u03b7 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7. \u03a4\u03bf deployment plan \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c3\u03c5\u03bc\u03b2\u03b1\u03c4\u03cc\u03c4\u03b7\u03c4\u03b1 target\u2013drafter, \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf cache behavior, fallback \u03ba\u03b1\u03b9 \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03b1\u03b3\u03ce\u03b3\u03b9\u03bc\u03b1 backend-matched tests. \u03a3\u03b5 \u03c5\u03c8\u03b7\u03bb\u03cc concurrency, \u03c4\u03bf \u03b3\u03b9\u03bd\u03cc\u03bc\u03b5\u03bd\u03bf batch size \u00d7 K \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03ba\u03ac\u03bd\u03b5\u03b9 \u03c4\u03bf speculative verification \u03b1\u03ba\u03c1\u03b9\u03b2\u03cc\u03c4\u03b5\u03c1\u03bf, \u03b3\u03b9\u2019 \u03b1\u03c5\u03c4\u03cc \u03bf scheduler \u03b5\u03af\u03bd\u03b1\u03b9 \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03b7\u03c2 \u03bb\u03cd\u03c3\u03b7\u03c2 \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03bb\u03b5\u03c0\u03c4\u03bf\u03bc\u03ad\u03c1\u03b5\u03b9\u03b1.<\/p>\n<h2 id=\"plaisio-apofasis\">\u0388\u03bd\u03b1 \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03cc \u03c0\u03bb\u03b1\u03af\u03c3\u03b9\u03bf \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7\u03c2 \u03b3\u03b9\u03b1 multimodal \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd\u03c4\u03b1<\/h2>\n<p>\u03a0\u03c1\u03ce\u03c4\u03b1 \u03bf\u03c1\u03af\u03c3\u03c4\u03b5 \u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc workload: \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03b5\u03b9\u03ba\u03cc\u03bd\u03b1\u03c2, \u03b1\u03c1\u03b9\u03b8\u03bc\u03cc visual tokens, \u03bc\u03ae\u03ba\u03bf\u03c2 output, concurrency, task mix \u03ba\u03b1\u03b9 quality contract. \u0388\u03c0\u03b5\u03b9\u03c4\u03b1 \u03b4\u03b7\u03bc\u03b9\u03bf\u03c5\u03c1\u03b3\u03ae\u03c3\u03c4\u03b5 autoregressive baseline \u03c3\u03c4\u03bf \u03af\u03b4\u03b9\u03bf hardware \u03ba\u03b1\u03b9 backend. Latency \u03b1\u03c0\u03cc \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03cc framework \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03ba\u03b1\u03c4\u03ac\u03bb\u03bb\u03b7\u03bb\u03bf \u03c3\u03b7\u03bc\u03b5\u03af\u03bf \u03b1\u03bd\u03b1\u03c6\u03bf\u03c1\u03ac\u03c2.<\/p>\n<p>\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 \u03b1\u03bd\u03ac stage: vision encoding, multimodal prefill, drafter conditioning, draft generation, tree construction, verification \u03ba\u03b1\u03b9 sampling. \u0391\u03bd \u03c4\u03bf conditioning \u03ba\u03c5\u03c1\u03b9\u03b1\u03c1\u03c7\u03b5\u03af, \u03ad\u03bd\u03b1\u03c2 \u03b3\u03c1\u03b7\u03b3\u03bf\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf\u03c2 L2 drafter \u03bc\u03cc\u03bd\u03bf\u03c2 \u03c4\u03bf\u03c5 \u03b4\u03b5\u03bd \u03b1\u03c1\u03ba\u03b5\u03af. \u0394\u03bf\u03ba\u03b9\u03bc\u03ac\u03c3\u03c4\u03b5 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf visual representation, feature reuse \u03ae routing, \u03b1\u03bb\u03bb\u03ac \u03b5\u03bb\u03ad\u03b3\u03be\u03c4\u03b5 \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ac OCR, grounding \u03ba\u03b1\u03b9 answer quality.<\/p>\n<div class=\"td-step-list\">\n<p class=\"td-step-list-title\">\u0395\u03c0\u03c4\u03ac \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf\u03b9 \u03c0\u03c1\u03b9\u03bd \u03b2\u03ac\u03bb\u03b5\u03c4\u03b5 L2 drafting \u03c3\u03b5 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae<\/p>\n<ol>\n<li><span class=\"td-step-kicker\">Test 1<\/span><strong>\u03a0\u03b1\u03b3\u03ce\u03c3\u03c4\u03b5 \u03c4\u03bf autoregressive baseline<\/strong>\n<p>\u03a7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03c4\u03b5 \u03af\u03b4\u03b9\u03bf target, backend, hardware, prompts, batch, stopping criteria \u03ba\u03b1\u03b9 output cap, \u03ce\u03c3\u03c4\u03b5 \u03c4\u03bf speedup \u03bd\u03b1 \u03ad\u03c7\u03b5\u03b9 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03bd\u03cc\u03b7\u03bc\u03b1.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Check 2<\/span><strong>\u03a7\u03b1\u03c1\u03c4\u03bf\u03b3\u03c1\u03b1\u03c6\u03ae\u03c3\u03c4\u03b5 \u03c4\u03bf multimodal condition<\/strong>\n<p>\u039a\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c8\u03c4\u03b5 \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7, visual-token length, vision encoder, prefill \u03ba\u03b1\u03b9 \u03c4\u03c5\u03c7\u03cc\u03bd target-feature-to-KV \u03bc\u03b5\u03c4\u03b1\u03c4\u03c1\u03bf\u03c0\u03ae \u03c4\u03bf\u03c5 drafter.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Check 3<\/span><strong>\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 MAT \u03ba\u03b1\u03b9 end-to-end \u03c7\u03c1\u03cc\u03bd\u03bf<\/strong>\n<p>\u039a\u03c1\u03b1\u03c4\u03ae\u03c3\u03c4\u03b5 \u03c4\u03bf accepted length \u03c9\u03c2 \u03b4\u03b9\u03b1\u03b3\u03bd\u03c9\u03c3\u03c4\u03b9\u03ba\u03cc metric, \u03b1\u03bb\u03bb\u03ac \u03b1\u03c0\u03bf\u03c6\u03b1\u03c3\u03af\u03c3\u03c4\u03b5 \u03bc\u03b5 p50\/p95 latency, throughput, \u03bc\u03bd\u03ae\u03bc\u03b7 \u03ba\u03b1\u03b9 \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03b1\u03bd\u03ac \u03bf\u03bb\u03bf\u03ba\u03bb\u03b7\u03c1\u03c9\u03bc\u03ad\u03bd\u03bf \u03b1\u03af\u03c4\u03b7\u03bc\u03b1.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Check 4<\/span><strong>\u0394\u03b9\u03b1\u03c7\u03c9\u03c1\u03af\u03c3\u03c4\u03b5 tasks \u03ba\u03b1\u03b9 \u03b1\u03bd\u03b1\u03bb\u03cd\u03c3\u03b5\u03b9\u03c2<\/strong>\n<p>\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 \u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ac OCR, captioning, visual reasoning \u03ba\u03b1\u03b9 multi-turn flows\u00b7 \u03ad\u03bd\u03b1 \u03bc\u03ad\u03c3\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03ba\u03c1\u03cd\u03b2\u03b5\u03b9 \u03b1\u03c1\u03bd\u03b7\u03c4\u03b9\u03ba\u03cc speedup \u03c3\u03b5 8K input.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Test 5<\/span><strong>\u0395\u03bb\u03ad\u03b3\u03be\u03c4\u03b5 checkpoint \u03ba\u03b1\u03b9 backend<\/strong>\n<p>\u0395\u03c0\u03b9\u03b2\u03b5\u03b2\u03b1\u03b9\u03ce\u03c3\u03c4\u03b5 target\u2013drafter compatibility, tokenizer, quantization, cache behavior \u03ba\u03b1\u03b9 \u03b1\u03bd \u03c4\u03bf serving engine \u03c5\u03c0\u03bf\u03c3\u03c4\u03b7\u03c1\u03af\u03b6\u03b5\u03b9 \u03c4\u03b7 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7 \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Test 6<\/span><strong>\u03a0\u03c1\u03bf\u03c3\u03b1\u03c1\u03bc\u03cc\u03c3\u03c4\u03b5 K \u03ba\u03b1\u03b9 concurrency<\/strong>\n<p>\u039c\u03b5\u03b9\u03ce\u03c3\u03c4\u03b5 verify length \u03cc\u03c4\u03b1\u03bd confidence \u03ae batch size \u03c0\u03b5\u03c1\u03bd\u03bf\u03cd\u03bd \u03c4\u03bf break-even, \u03b1\u03bd\u03c4\u03af \u03bd\u03b1 \u03b5\u03c0\u03b9\u03b2\u03ac\u03bb\u03bb\u03b5\u03c4\u03b5 \u03c3\u03c4\u03b1\u03b8\u03b5\u03c1\u03cc \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf block \u03c3\u03b5 \u03ba\u03ac\u03b8\u03b5 \u03b1\u03af\u03c4\u03b7\u03bc\u03b1.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0388\u03bb\u03b5\u03b3\u03c7\u03bf\u03c2 7<\/span><strong>\u039a\u03c1\u03b1\u03c4\u03ae\u03c3\u03c4\u03b5 target-only fallback<\/strong>\n<p>\u038c\u03c4\u03b1\u03bd \u03c4\u03bf \u03c0\u03c1\u03bf\u03b2\u03bb\u03b5\u03c0\u03cc\u03bc\u03b5\u03bd\u03bf \u03cc\u03c6\u03b5\u03bb\u03bf\u03c2 \u03b3\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c1\u03bd\u03b7\u03c4\u03b9\u03ba\u03cc \u03ae \u03b1\u03c0\u03bf\u03c4\u03c5\u03b3\u03c7\u03ac\u03bd\u03b5\u03b9 quality check, \u03b5\u03c0\u03b9\u03c3\u03c4\u03c1\u03ad\u03c8\u03c4\u03b5 \u03b4\u03c5\u03bd\u03b1\u03bc\u03b9\u03ba\u03ac \u03c3\u03b5 autoregressive decoding \u03ba\u03b1\u03b9 \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c8\u03c4\u03b5 \u03c4\u03bf\u03bd \u03bb\u03cc\u03b3\u03bf.<\/p>\n<\/li>\n<\/ol>\n<\/div>\n<p>\u03a4\u03bf fallback \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c3\u03c5\u03bd\u03b4\u03ad\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 observability \u03ba\u03b1\u03b9 capacity planning. \u038c\u03c0\u03c9\u03c2 \u03ba\u03b1\u03b9 \u03c3\u03c4\u03bf <a href=\"https:\/\/twodots.gr\/multi-agent-ai-concurrency-control\/\">concurrency control \u03c0\u03bf\u03bb\u03bb\u03ce\u03bd AI agents<\/a>, \u03b7 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b7 \u03c4\u03bf\u03c0\u03b9\u03ba\u03ae \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b2\u03bb\u03ac\u03c8\u03b5\u03b9 \u03c4\u03bf \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03cc\u03c4\u03b1\u03bd \u03b1\u03c5\u03be\u03ac\u03bd\u03b5\u03b9 \u03c4\u03b1\u03c5\u03c4\u03cc\u03c7\u03c1\u03bf\u03bd\u03b1 verification load, \u03bf\u03c5\u03c1\u03ad\u03c2 \u03ba\u03b1\u03b9 memory pressure.<\/p>\n<h2 id=\"teliko-symperasma\">\u03a4\u03bf \u03c3\u03c5\u03bc\u03c0\u03ad\u03c1\u03b1\u03c3\u03bc\u03b1: \u03bc\u03b5\u03c1\u03b9\u03ba\u03ae \u03b5\u03c4\u03bf\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1, \u03cc\u03c7\u03b9 \u03ba\u03b1\u03b8\u03bf\u03bb\u03b9\u03ba\u03ae \u03c3\u03c5\u03bd\u03c4\u03b1\u03b3\u03ae<\/h2>\n<p>\u0397 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03b4\u03b5\u03bd \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03bf\u03cd\u03c4\u03b5 \u03ba\u03b1\u03b8\u03bf\u03bb\u03b9\u03ba\u03ae \u03b1\u03c0\u03bf\u03c4\u03c5\u03c7\u03af\u03b1 \u03bf\u03cd\u03c4\u03b5 \u03ba\u03b1\u03b8\u03bf\u03bb\u03b9\u03ba\u03ae \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03af\u03b1. \u03a4\u03bf L2 drafting \u03b5\u03af\u03bd\u03b1\u03b9 \u03b5\u03c6\u03b9\u03ba\u03c4\u03cc \u03ba\u03b1\u03b9 \u03ad\u03b4\u03c9\u03c3\u03b5 \u03ad\u03c9\u03c2 2,60\u00d7 end-to-end speedup \u03c3\u03c4\u03b7 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 \u03c4\u03bf\u03c5 Qwen3.6-27B. \u03a3\u03c4\u03b1 Qwen3-VL targets, \u03cc\u03bc\u03c9\u03c2, \u03c5\u03c8\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03b7 \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03c3\u03b5 \u03bd\u03b1 \u03bc\u03b7\u03b4\u03b5\u03bd\u03af\u03c3\u03b5\u03b9 \u03ae \u03bd\u03b1 \u03b1\u03bd\u03c4\u03b9\u03c3\u03c4\u03c1\u03ad\u03c8\u03b5\u03b9 \u03c4\u03bf \u03cc\u03c6\u03b5\u03bb\u03bf\u03c2 \u03c0\u03b1\u03c1\u03cc\u03c4\u03b9 \u03c4\u03bf MAT \u03c0\u03b1\u03c1\u03ad\u03bc\u03b5\u03bd\u03b5 \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf.<\/p>\n<div class=\"td-decision-band\">\n<div class=\"td-decision-band-content\">\n<p class=\"td-decision-band-kicker\">\u03a4\u03bf \u03c3\u03c9\u03c3\u03c4\u03cc \u03cc\u03c1\u03b9\u03bf \u03c4\u03bf\u03c5 \u03b9\u03c3\u03c7\u03c5\u03c1\u03b9\u03c3\u03bc\u03bf\u03cd<\/p>\n<p class=\"td-decision-band-title\">\u0397 \u03b5\u03c4\u03bf\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b1\u03bd\u03ae\u03ba\u03b5\u03b9 \u03c3\u03c4\u03b7 \u03b4\u03b9\u03b1\u03bc\u03cc\u03c1\u03c6\u03c9\u03c3\u03b7 model\u2013task\u2013condition\u2013system<\/p>\n<p>\u03a4\u03bf DFlash \u03ae \u03c4\u03bf DSpark \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03cc \u03bc\u03cc\u03bd\u03bf \u03c4\u03bf\u03c5 \u00ab\u03b3\u03c1\u03b7\u03b3\u03bf\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf\u00bb. \u03a4\u03bf claim \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 backend-matched baseline, \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03b5\u03b9\u03ba\u03cc\u03bd\u03b1\u03c2, production concurrency \u03ba\u03b1\u03b9 \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7 \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03bf\u03c5 \u03c4\u03bf\u03c5 request path.<\/p>\n<\/div>\n<\/div>\n<p>\u0397 \u03bf\u03c5\u03c3\u03af\u03b1 \u03b3\u03b9\u03b1 \u03ad\u03bd\u03b1\u03bd decision maker \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03bb\u03ae: \u03b1\u03b3\u03bf\u03c1\u03ac\u03c3\u03c4\u03b5 end-to-end \u03c7\u03c1\u03cc\u03bd\u03bf, \u03cc\u03c7\u03b9 \u03ad\u03bd\u03b1 \u03b5\u03bd\u03c4\u03c5\u03c0\u03c9\u03c3\u03b9\u03b1\u03ba\u03cc acceptance metric. \u0395\u03be\u03b5\u03c4\u03ac\u03c3\u03c4\u03b5 \u03c0\u03bf\u03b9\u03bf\u03c2 \u03c0\u03bb\u03b7\u03c1\u03ce\u03bd\u03b5\u03b9 \u03c4\u03bf multimodal prefill, \u03b1\u03bd \u03bf drafter \u03b1\u03be\u03b9\u03bf\u03c0\u03bf\u03b9\u03b5\u03af \u03c0\u03c1\u03ac\u03b3\u03bc\u03b1\u03c4\u03b9 \u03c4\u03bf visual context, \u03b1\u03bd \u03c4\u03bf backend \u03c5\u03c0\u03bf\u03c3\u03c4\u03b7\u03c1\u03af\u03b6\u03b5\u03b9 \u03c4\u03bf\u03bd \u03bc\u03b7\u03c7\u03b1\u03bd\u03b9\u03c3\u03bc\u03cc \u03ba\u03b1\u03b9 \u03b1\u03bd \u03c4\u03bf \u03ba\u03ad\u03c1\u03b4\u03bf\u03c2 \u03b5\u03c0\u03b9\u03b2\u03b9\u03ce\u03bd\u03b5\u03b9 \u03c3\u03c4\u03b7 \u03b4\u03b9\u03ba\u03ae \u03c3\u03b1\u03c2 \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7, batch \u03ba\u03b1\u03b9 task distribution.<\/p>\n<p>\u0388\u03bd\u03b1 pilot \u03bc\u03b5 \u03c3\u03b1\u03c6\u03ad\u03c2 baseline, stage telemetry \u03ba\u03b1\u03b9 target-only fallback \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03c0\u03bf\u03b4\u03b5\u03af\u03be\u03b5\u03b9 \u03b1\u03bd \u03c4\u03bf speculative decoding \u03b1\u03be\u03af\u03b6\u03b5\u03b9 \u03c3\u03c4\u03bf \u03b4\u03b9\u03ba\u03cc \u03c3\u03b1\u03c2 \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd. \u03a7\u03c9\u03c1\u03af\u03c2 \u03b1\u03c5\u03c4\u03ac, \u03c4\u03bf 2,60\u00d7 \u03b5\u03bd\u03cc\u03c2 paper \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03b5\u03cd\u03c1\u03b7\u03bc\u03b1 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7\u03c2 \u03b4\u03b9\u03b1\u03bc\u03cc\u03c1\u03c6\u03c9\u03c3\u03b7\u03c2 \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03c5\u03c0\u03cc\u03c3\u03c7\u03b5\u03c3\u03b7 \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 multimodal \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ae.<\/p>\n<section class=\"td-service-cta\">\n<div class=\"td-service-cta-content\">\n<p class=\"td-service-cta-kicker\">\u0391\u03c0\u03cc \u03c4\u03bf benchmark \u03c3\u03c4\u03bf \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03b9\u03ba\u03cc workload<\/p>\n<p class=\"td-service-cta-title\">\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 \u03c4\u03bf multimodal AI pipeline \u03c0\u03c1\u03b9\u03bd \u03b5\u03c0\u03b5\u03bd\u03b4\u03cd\u03c3\u03b5\u03c4\u03b5 \u03c3\u03c4\u03b7\u03bd \u03b5\u03c0\u03b9\u03c4\u03ac\u03c7\u03c5\u03bd\u03c3\u03b7<\/p>\n<p>\u0397 TWO DOTS \u03c3\u03c7\u03b5\u03b4\u03b9\u03ac\u03b6\u03b5\u03b9 <a href=\"https:\/\/twodots.gr\/aftomatismoi-epicheiriseon-ai\/\">business automation with AI<\/a> \u03bc\u03b5 backend-matched baselines, latency \u03b1\u03bd\u03ac stage, observability, capacity controls \u03ba\u03b1\u03b9 \u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c2 fallback, \u03ce\u03c3\u03c4\u03b5 \u03b7 \u03b2\u03b5\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03bd\u03b1 \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03c4\u03b1\u03b9 \u03c3\u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc workload.<\/p>\n<div class=\"td-service-cta-actions\"><a class=\"td-service-cta-button\" href=\"https:\/\/twodots.gr\/aftomatismoi-epicheiriseon-ai\/\">\u03a3\u03c5\u03b6\u03b7\u03c4\u03ae\u03c3\u03c4\u03b5 \u03c4\u03bf AI workload \u03c3\u03b1\u03c2<\/a><\/div>\n<\/div>\n<\/section>\n<section id=\"sychnes-erotiseis\" class=\"td-faq-section\">\n<div class=\"td-faq\">\n<p class=\"td-faq-heading\">Frequently Asked Questions (FAQs)<\/p>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03b9 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf multimodal speculative decoding;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0395\u03af\u03bd\u03b1\u03b9 draft-and-verify \u03b5\u03c0\u03b9\u03c4\u03ac\u03c7\u03c5\u03bd\u03c3\u03b7 \u03b3\u03b9\u03b1 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03c0\u03bf\u03c5 \u03b4\u03ad\u03c7\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b5\u03b9\u03ba\u03cc\u03bd\u03b1, video, audio \u03ae \u03ac\u03bb\u03bb\u03bf multimodal context. \u0388\u03bd\u03b1\u03c2 \u03c6\u03b8\u03b7\u03bd\u03cc\u03c4\u03b5\u03c1\u03bf\u03c2 \u03bc\u03b7\u03c7\u03b1\u03bd\u03b9\u03c3\u03bc\u03cc\u03c2 \u03c0\u03c1\u03bf\u03c4\u03b5\u03af\u03bd\u03b5\u03b9 tokens \u03ba\u03b1\u03b9 \u03c4\u03bf target model \u03c4\u03b1 \u03b5\u03c0\u03b1\u03bb\u03b7\u03b8\u03b5\u03cd\u03b5\u03b9.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03b9 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03bf\u03c5\u03bd \u03c4\u03b1 \u03b5\u03c0\u03af\u03c0\u03b5\u03b4\u03b1 L0, L1 \u03ba\u03b1\u03b9 L2;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a4\u03bf L0 \u03c0\u03c1\u03bf\u03c7\u03c9\u03c1\u03ac \u03c4\u03bf draft depth \u03c3\u03b5\u03b9\u03c1\u03b9\u03b1\u03ba\u03ac, \u03c4\u03bf L1 \u03c0\u03c1\u03bf\u03b2\u03bb\u03ad\u03c0\u03b5\u03b9 \u03c0\u03bf\u03bb\u03bb\u03ad\u03c2 \u03c0\u03c1\u03bf\u03ba\u03b1\u03b8\u03bf\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03b5\u03c2 \u03b8\u03ad\u03c3\u03b5\u03b9\u03c2 \u03b1\u03c0\u03cc \u03ba\u03bf\u03b9\u03bd\u03cc prefix \u03ba\u03b1\u03b9 \u03c4\u03bf L2 \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 \u03ae \u03b2\u03b5\u03bb\u03c4\u03b9\u03ce\u03bd\u03b5\u03b9 \u03ad\u03bd\u03b1 \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03bf \u03bc\u03b5\u03bb\u03bb\u03bf\u03bd\u03c4\u03b9\u03ba\u03cc block \u03c9\u03c2 \u03ba\u03bf\u03b9\u03bd\u03ae \u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03bf\u03b9\u03bf \u03ae\u03c4\u03b1\u03bd \u03c4\u03bf \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03bc\u03b5\u03c4\u03c1\u03b7\u03bc\u03ad\u03bd\u03bf speedup;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a3\u03c4\u03bf matched SGLang setup \u03c4\u03bf\u03c5 Qwen3.6-27B, \u03c4\u03bf DFlash \u03b5\u03af\u03c7\u03b5 task-equal MAT 4,38 \u03ba\u03b1\u03b9 \u03bc\u03ad\u03c3\u03b7 end-to-end \u03b5\u03c0\u03b9\u03c4\u03ac\u03c7\u03c5\u03bd\u03c3\u03b7 2,60\u00d7 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 autoregressive decoding.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0393\u03b9\u03b1\u03c4\u03af \u03b7 \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 8K \u03bc\u03b5\u03af\u03c9\u03c3\u03b5 \u03c4\u03b7\u03bd \u03c4\u03b1\u03c7\u03cd\u03c4\u03b7\u03c4\u03b1;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0395\u03c0\u03b5\u03b9\u03b4\u03ae \u03c4\u03bf \u03c0\u03c1\u03cc\u03c3\u03b8\u03b5\u03c4\u03bf visual prefill \u03ba\u03b1\u03b9 \u03b7 \u03ba\u03b1\u03c4\u03b1\u03c3\u03ba\u03b5\u03c5\u03ae drafter-side cache \u03ba\u03cc\u03c3\u03c4\u03b9\u03c3\u03b1\u03bd \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf \u03b1\u03c0\u03cc \u03cc\u03c3\u03bf \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03c3\u03b1\u03bd \u03bd\u03b1 \u03b1\u03c0\u03bf\u03c3\u03b2\u03ad\u03c3\u03bf\u03c5\u03bd \u03c4\u03b1 \u03b1\u03c0\u03bf\u03b4\u03b5\u03ba\u03c4\u03ac draft tokens.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0391\u03c1\u03ba\u03b5\u03af \u03c4\u03bf MAT \u03b3\u03b9\u03b1 \u03bd\u03b1 \u03b5\u03c0\u03b9\u03bb\u03ad\u03be\u03bf\u03c5\u03bc\u03b5 \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u038c\u03c7\u03b9. \u03a4\u03bf MAT \u03b5\u03af\u03bd\u03b1\u03b9 \u03b4\u03b9\u03b1\u03b3\u03bd\u03c9\u03c3\u03c4\u03b9\u03ba\u03cc metric \u03b3\u03b9\u03b1 accepted tokens \u03ba\u03b1\u03b9 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c3\u03c5\u03bd\u03b4\u03c5\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 end-to-end latency, throughput, GPU memory, quality \u03ba\u03b1\u03b9 backend-matched baseline.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bf drafter \u03cc\u03bb\u03bf \u03c4\u03bf visual context;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u038c\u03c7\u03b9 \u03c0\u03ac\u03bd\u03c4\u03b1. \u03a3\u03c4\u03b9\u03c2 ablations \u03b7 \u03b1\u03c6\u03b1\u03af\u03c1\u03b5\u03c3\u03b7 visual K\/V \u03b5\u03af\u03c7\u03b5 \u03bc\u03b9\u03ba\u03c1\u03ae \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ae \u03b5\u03c0\u03af\u03b4\u03c1\u03b1\u03c3\u03b7 \u03c3\u03c4\u03bf MAT, \u03b1\u03bb\u03bb\u03ac \u03c4\u03bf TextVQA \u03ae\u03c4\u03b1\u03bd \u03c3\u03b1\u03c6\u03ce\u03c2 \u03c0\u03b9\u03bf \u03b5\u03c5\u03b1\u03af\u03c3\u03b8\u03b7\u03c4\u03bf, \u03ac\u03c1\u03b1 \u03b1\u03c0\u03b1\u03b9\u03c4\u03b5\u03af\u03c4\u03b1\u03b9 task-specific \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf\u03c2.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0395\u03af\u03bd\u03b1\u03b9 \u03c4\u03b1 vLLM \u03ba\u03b1\u03b9 SGLang \u03ad\u03c4\u03bf\u03b9\u03bc\u03b1 \u03b3\u03b9\u03b1 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a0\u03c1\u03bf\u03c3\u03c6\u03ad\u03c1\u03bf\u03c5\u03bd \u03c4\u03b9\u03c2 \u03c0\u03bb\u03b7\u03c1\u03ad\u03c3\u03c4\u03b5\u03c1\u03b5\u03c2 production-oriented \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ad\u03c2 \u03c0\u03bf\u03c5 \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 \u03b7 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1, \u03b1\u03bb\u03bb\u03ac \u03b7 \u03ba\u03ac\u03bb\u03c5\u03c8\u03b7 DFlash \u03ba\u03b1\u03b9 DSpark \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03ac\u03bd\u03b9\u03c3\u03b7 \u03b1\u03bd\u03ac model, backend \u03ba\u03b1\u03b9 hardware.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03bf\u03b9\u03bf \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf \u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c2 business takeaway;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0394\u03bf\u03ba\u03b9\u03bc\u03ac\u03c3\u03c4\u03b5 \u03c4\u03b7 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7 \u03b4\u03b9\u03b1\u03bc\u03cc\u03c1\u03c6\u03c9\u03c3\u03b7 model\u2013task\u2013condition\u2013system \u03c3\u03c4\u03bf \u03b4\u03b9\u03ba\u03cc \u03c3\u03b1\u03c2 workload \u03ba\u03b1\u03b9 \u03ba\u03c1\u03b1\u03c4\u03ae\u03c3\u03c4\u03b5 \u03b4\u03c5\u03bd\u03b1\u03bc\u03b9\u03ba\u03cc target-only fallback \u03cc\u03c4\u03b1\u03bd \u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03ba\u03ad\u03c1\u03b4\u03bf\u03c2 \u03b3\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c1\u03bd\u03b7\u03c4\u03b9\u03ba\u03cc.<\/p>\n<\/div>\n<\/details>\n<\/div>\n<\/section>\n<div class=\"td-source-list\">\n<p id=\"piges\" class=\"td-source-list-title\">Sources<\/p>\n<ul>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2608.20743\" target=\"_blank\" rel=\"noopener\">Li et al. \u2014 Is Multimodal Speculative Decoding Ready for Diffusion-Based Parallel Drafting?, abstract \u03ba\u03b1\u03b9 metadata, arXiv v1<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/html\/2608.20743v1\" target=\"_blank\" rel=\"noopener\">Li et al. \u2014 \u03a0\u03bb\u03ae\u03c1\u03b5\u03c2 HTML paper \u03bc\u03b5 taxonomy, \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1, ablations \u03ba\u03b1\u03b9 \u03cc\u03c1\u03b9\u03b1 \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03af\u03c9\u03c3\u03b7\u03c2<\/a><\/li>\n<li><a href=\"https:\/\/docs.vllm.ai\/projects\/speculators\/en\/latest\/user_guide\/algorithms\/dflash\/\" target=\"_blank\" rel=\"noopener\">vLLM Speculators \u2014 \u0395\u03c0\u03af\u03c3\u03b7\u03bc\u03b7 \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03af\u03c9\u03c3\u03b7 \u03c4\u03bf\u03c5 DFlash \u03ba\u03b1\u03b9 \u03c4\u03c9\u03bd \u03b4\u03b9\u03b1\u03b8\u03ad\u03c3\u03b9\u03bc\u03c9\u03bd integrations<\/a><\/li>\n<li><a href=\"https:\/\/docs.vllm.ai\/en\/latest\/features\/speculative_decoding\/dynamic_speculative_decoding\/\" target=\"_blank\" rel=\"noopener\">vLLM \u2014 Dynamic speculative decoding \u03b3\u03b9\u03b1 \u03c0\u03c1\u03bf\u03c3\u03b1\u03c1\u03bc\u03bf\u03b3\u03ae verify length \u03c3\u03b5 batch size \u03ba\u03b1\u03b9 confidence<\/a><\/li>\n<\/ul>\n<\/div>","protected":false},"excerpt":{"rendered":"<p>\u03a4\u03bf multimodal speculative decoding \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03c0\u03b9\u03c4\u03b1\u03c7\u03cd\u03bd\u03b5\u03b9 \u03c4\u03b7\u03bd \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae tokens, \u03b1\u03bb\u03bb\u03ac visual prefill, \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03b5\u03b9\u03ba\u03cc\u03bd\u03b1\u03c2 \u03ba\u03b1\u03b9 serving stack \u03ba\u03b1\u03b8\u03bf\u03c1\u03af\u03b6\u03bf\u03c5\u03bd \u03c4\u03bf \u03c4\u03b5\u03bb\u03b9\u03ba\u03cc \u03ba\u03ad\u03c1\u03b4\u03bf\u03c2.<\/p>","protected":false},"author":1,"featured_media":97203,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_gspb_post_css":"","content-type":"","footnotes":""},"categories":[199],"tags":[7336,20304,18467,19545,19806],"class_list":["post-96860","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-techniti-noimosyni","tag-ai-inference","tag-dflash","tag-multimodal-ai","tag-speculative-decoding","tag-vision-language-models"],"blocksy_meta":{"styles_descriptor":{"styles":{"desktop":"","tablet":"","mobile":""},"google_fonts":[],"version":8}},"_links":{"self":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts\/96860","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/comments?post=96860"}],"version-history":[{"count":0,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts\/96860\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/media\/97203"}],"wp:attachment":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/media?parent=96860"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/categories?post=96860"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/tags?post=96860"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}