{"id":98250,"date":"2026-10-07T09:33:55","date_gmt":"2026-10-07T06:33:55","guid":{"rendered":"https:\/\/twodots.gr\/?p=98250"},"modified":"2026-10-07T09:33:56","modified_gmt":"2026-10-07T06:33:56","slug":"minima-kv-long-context-ai-cache-compression","status":"publish","type":"post","link":"https:\/\/twodots.gr\/bg\/minima-kv-long-context-ai-cache-compression\/","title":{"rendered":"Minima-KV: \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b7 \u03bc\u03bd\u03ae\u03bc\u03b7 \u03b3\u03b9\u03b1 long-context AI \u03c7\u03c9\u03c1\u03af\u03c2 \u03b4\u03b9\u03b1\u03b3\u03c1\u03b1\u03c6\u03ae \u03c4\u03bf\u03c5 \u03c0\u03b1\u03bb\u03b9\u03bf\u03cd context"},"content":{"rendered":"<div class=\"td-article-lede\">\n<p><strong>\u041e\u0442\u0433\u043e\u0432\u043e\u0440 \u043f\u044a\u0440\u0432\u043e:<\/strong> \u03a4\u03bf Minima-KV \u03c0\u03c1\u03bf\u03c4\u03b5\u03af\u03bd\u03b5\u03b9 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03c4\u03b7\u03c2 KV cache \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03b4\u03b9\u03b1\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 \u03c4\u03b9\u03c2 \u03c0\u03b1\u03bb\u03b9\u03ad\u03c2 \u03c3\u03b5\u03bb\u03af\u03b4\u03b5\u03c2 \u03b5\u03bd\u03cc\u03c2 \u03b5\u03bd\u03b5\u03c1\u03b3\u03bf\u03cd request: \u03ba\u03c1\u03b1\u03c4\u03ac \u03c4\u03b9\u03c2 \u03c0\u03c1\u03cc\u03c3\u03c6\u03b1\u03c4\u03b5\u03c2 \u03ba\u03b1\u03b9 \u03c0\u03c1\u03bf\u03c3\u03c4\u03b1\u03c4\u03b5\u03c5\u03bc\u03ad\u03bd\u03b5\u03c2 \u03c3\u03b5 FP8 \u03ba\u03b1\u03b9 \u03bc\u03b5\u03c4\u03b1\u03c6\u03ad\u03c1\u03b5\u03b9 \u03c4\u03b9\u03c2 \u03c5\u03c0\u03cc\u03bb\u03bf\u03b9\u03c0\u03b5\u03c2 \u03c3\u03b5 TQ3. \u0397 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03cc \u03cc\u03c6\u03b5\u03bb\u03bf\u03c2 \u03bc\u03bd\u03ae\u03bc\u03b7\u03c2, \u03b1\u03bb\u03bb\u03ac \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u03b1\u03ba\u03cc\u03bc\u03b7 \u03ba\u03b1\u03b8\u03bf\u03bb\u03b9\u03ba\u03ae \u03b9\u03c3\u03bf\u03b4\u03c5\u03bd\u03b1\u03bc\u03af\u03b1 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2 \u03ae \u03c4\u03b1\u03c7\u03cd\u03c4\u03b7\u03c4\u03b1\u03c2.<\/p>\n<\/div>\n<p>\u03a4\u03bf Minima-KV \u03c3\u03c5\u03bc\u03c0\u03b9\u03ad\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd KV cache \u03c7\u03c9\u03c1\u03af\u03c2 \u03b4\u03b9\u03b1\u03b3\u03c1\u03b1\u03c6\u03ae \u03c0\u03b1\u03bb\u03b9\u03bf\u03cd context. \u03a4\u03b9 \u03ad\u03b4\u03b5\u03b9\u03be\u03b1\u03bd \u03bf\u03b9 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03bc\u03bd\u03ae\u03bc\u03b7\u03c2, \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2 \u03ba\u03b1\u03b9 throughput \u03c3\u03c4\u03bf long-context AI.<\/p>\n<p>\u038c\u03c3\u03bf \u03bc\u03b5\u03b3\u03b1\u03bb\u03ce\u03bd\u03b5\u03b9 \u03c4\u03bf context window \u03b5\u03bd\u03cc\u03c2 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf\u03c5, \u03b1\u03c5\u03be\u03ac\u03bd\u03b5\u03c4\u03b1\u03b9 \u03ba\u03b1\u03b9 \u03b7 \u03bc\u03bd\u03ae\u03bc\u03b7 \u03c0\u03bf\u03c5 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b3\u03b9\u03b1 \u03bd\u03b1 \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03b5\u03af \u03c4\u03b1 keys \u03ba\u03b1\u03b9 values \u03c4\u03c9\u03bd \u03c0\u03c1\u03bf\u03b7\u03b3\u03bf\u03cd\u03bc\u03b5\u03bd\u03c9\u03bd tokens. \u0397 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 <em>Minima-KV: Retention-Preserving KV Cache Compression with Mixed-Format Paged Attention<\/em> \u03c0\u03c1\u03bf\u03c4\u03b5\u03af\u03bd\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03c4\u03c1\u03b9\u03ce\u03bd \u03b5\u03c0\u03b9\u03c0\u03ad\u03b4\u03c9\u03bd \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae \u03c0\u03bf\u03c5 \u03c3\u03c5\u03bc\u03c0\u03b9\u03ad\u03b6\u03b5\u03b9 \u03c4\u03bf \u03c0\u03b1\u03bb\u03b9\u03cc\u03c4\u03b5\u03c1\u03bf state \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03c4\u03bf \u03b1\u03c6\u03b1\u03b9\u03c1\u03b5\u03af \u03b1\u03c0\u03cc \u03c4\u03bf \u03b5\u03bd\u03b5\u03c1\u03b3\u03cc \u03b1\u03af\u03c4\u03b7\u03bc\u03b1.<\/p>\n<p>\u0397 \u03b9\u03b4\u03ad\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03ae \u03b3\u03b9\u03b1 \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ad\u03c2 \u03c0\u03bf\u03c5 \u03b5\u03c0\u03b5\u03be\u03b5\u03c1\u03b3\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03bc\u03b5\u03b3\u03ac\u03bb\u03b1 \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03ac \u03ad\u03b3\u03b3\u03c1\u03b1\u03c6\u03b1, \u03b5\u03ba\u03c4\u03b5\u03c4\u03b1\u03bc\u03ad\u03bd\u03b5\u03c2 \u03b2\u03ac\u03c3\u03b5\u03b9\u03c2 \u03b3\u03bd\u03ce\u03c3\u03b7\u03c2, code repositories \u03ae \u03bc\u03b1\u03ba\u03c1\u03ad\u03c2 \u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03af\u03b5\u03c2 \u03b5\u03bd\u03b5\u03c1\u03b3\u03b5\u03b9\u03ce\u03bd. \u0394\u03b5\u03bd \u03b1\u03c1\u03ba\u03b5\u03af \u03cc\u03bc\u03c9\u03c2 \u03bd\u03b1 \u03b4\u03b9\u03b1\u03b2\u03ac\u03c3\u03bf\u03c5\u03bc\u03b5 \u03ad\u03bd\u03b1\u03bd \u03bb\u03cc\u03b3\u03bf \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7\u03c2 \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03c4\u03bf\u03bd \u03bc\u03b5\u03c4\u03b1\u03c4\u03c1\u03ad\u03c8\u03bf\u03c5\u03bc\u03b5 \u03b1\u03c0\u03b5\u03c5\u03b8\u03b5\u03af\u03b1\u03c2 \u03c3\u03b5 \u03c5\u03c0\u03cc\u03c3\u03c7\u03b5\u03c3\u03b7 \u03b3\u03b9\u03b1 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf\u03c5\u03c2 \u03c7\u03c1\u03ae\u03c3\u03c4\u03b5\u03c2. \u0397 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03c7\u03c9\u03c1\u03af\u03b6\u03b5\u03b9 \u03c3\u03c5\u03bd\u03b5\u03b9\u03b4\u03b7\u03c4\u03ac \u03c4\u03b1 \u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af\u03b1 \u03b3\u03b9\u03b1 \u03bc\u03bd\u03ae\u03bc\u03b7, \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1, direct decode \u03ba\u03b1\u03b9 prefix reuse \u03c3\u03b5 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ac profiles.<\/p>\n<aside class=\"td-article-note\"><strong>\u0397 \u03c3\u03c9\u03c3\u03c4\u03ae \u03b1\u03bd\u03ac\u03b3\u03bd\u03c9\u03c3\u03b7:<\/strong> \u03c4\u03b1 18,3 KiB \u03b1\u03bd\u03ac live token \u03b5\u03af\u03bd\u03b1\u03b9 owner-reported aggregate \u03b3\u03b9\u03b1 \u03c4\u03bf attention KV \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf\u03c5 workload. \u0394\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ae \u03ba\u03b1\u03c4\u03b1\u03bd\u03ac\u03bb\u03c9\u03c3\u03b7 GPU, \u03c3\u03c4\u03b1\u03b8\u03b5\u03c1\u03cc\u03c2 \u03bb\u03cc\u03b3\u03bf\u03c2 \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 context \u03bf\u03cd\u03c4\u03b5 \u03b1\u03c0\u03cc\u03b4\u03b5\u03b9\u03be\u03b7 \u03cc\u03c4\u03b9 \u03b7 \u03c7\u03c9\u03c1\u03b7\u03c4\u03b9\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1 \u03ae \u03c4\u03bf throughput \u03b1\u03c5\u03be\u03ac\u03bd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c3\u03c4\u03bf\u03bd \u03af\u03b4\u03b9\u03bf \u03b2\u03b1\u03b8\u03bc\u03cc.<\/aside>\n<div class=\"td-article-toc\">\n<div class=\"td-toc-title\">\u0421\u044a\u0434\u044a\u0440\u0436\u0430\u043d\u0438\u0435<\/div>\n<ul>\n<li><a href=\"#kv-cache-emporio\">\u0393\u03b9\u03b1\u03c4\u03af \u03b7 KV cache \u03b3\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b5\u03bc\u03c0\u03cc\u03b4\u03b9\u03bf \u03c3\u03c4\u03bf long-context AI<\/a><\/li>\n<li><a href=\"#tria-epipeda\">Recent, Anchor \u03ba\u03b1\u03b9 Stale: \u03b7 \u03b9\u03b5\u03c1\u03b1\u03c1\u03c7\u03af\u03b1 \u03c4\u03bf\u03c5 Minima-KV<\/a><\/li>\n<li><a href=\"#retention-preserving\">\u03a4\u03b9 \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03cc\u03c4\u03b1\u03bd \u03b7 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03b4\u03b5\u03bd \u03b4\u03b9\u03b1\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 \u03c4\u03bf \u03c0\u03b1\u03bb\u03b9\u03cc context<\/a><\/li>\n<li><a href=\"#mixed-format-attention\">\u03a0\u03ce\u03c2 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b5\u03af \u03c4\u03bf mixed-format attention<\/a><\/li>\n<li><a href=\"#metriseis-mnimis\">\u03a4\u03b9 \u03ad\u03b4\u03b5\u03b9\u03be\u03b1\u03bd \u03bf\u03b9 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03bc\u03bd\u03ae\u03bc\u03b7\u03c2<\/a><\/li>\n<li><a href=\"#poiotita\">\u03a0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1: \u03bc\u03b9\u03ba\u03c1\u03ad\u03c2 \u03b1\u03c0\u03bf\u03ba\u03bb\u03af\u03c3\u03b5\u03b9\u03c2, \u03cc\u03c7\u03b9 \u03b3\u03b5\u03bd\u03b9\u03ba\u03cc \u00ab\u03c7\u03c9\u03c1\u03af\u03c2 \u03b1\u03c0\u03ce\u03bb\u03b5\u03b9\u03b5\u03c2\u00bb<\/a><\/li>\n<li><a href=\"#direct-decode-prefix\">Direct decode \u03ba\u03b1\u03b9 prefix reuse \u03b5\u03af\u03bd\u03b1\u03b9 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ad\u03c2<\/a><\/li>\n<li><a href=\"#epicheirimatiki-axia\">\u03a4\u03b9 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03b3\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03ac \u03c3\u03c5\u03c3\u03c4\u03ae\u03bc\u03b1\u03c4\u03b1 AI<\/a><\/li>\n<li><a href=\"#elegxos-paragogis\">\u03a0\u03ce\u03c2 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af\u03c4\u03b1\u03b9 \u03bc\u03b9\u03b1 \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ae KV compression \u03c3\u03c4\u03b7\u03bd \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae<\/a><\/li>\n<li><a href=\"#oria\">\u039f\u03b9 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03bf\u03af \u03c0\u03bf\u03c5 \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c7\u03b1\u03b8\u03bf\u03cd\u03bd<\/a><\/li>\n<li><a href=\"#ousia\">\u0397 \u03bf\u03c5\u03c3\u03af\u03b1: \u03b4\u03b9\u03b1\u03c7\u03b5\u03af\u03c1\u03b9\u03c3\u03b7 \u03c0\u03b9\u03c3\u03c4\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2, \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03bf \u03bb\u03b9\u03b3\u03cc\u03c4\u03b5\u03c1\u03b1 bits<\/a><\/li>\n<\/ul>\n<\/div>\n<h2 id=\"kv-cache-emporio\">\u0393\u03b9\u03b1\u03c4\u03af \u03b7 KV cache \u03b3\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b5\u03bc\u03c0\u03cc\u03b4\u03b9\u03bf \u03c3\u03c4\u03bf long-context AI<\/h2>\n<p>\u03a3\u03c4\u03b7\u03bd \u03b1\u03c5\u03c4\u03bf\u03c0\u03b1\u03bb\u03af\u03bd\u03b4\u03c1\u03bf\u03bc\u03b7 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae, \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03ba\u03c1\u03b1\u03c4\u03ac \u03c4\u03b1 keys \u03ba\u03b1\u03b9 values \u03c4\u03c9\u03bd \u03c0\u03c1\u03bf\u03b7\u03b3\u03bf\u03cd\u03bc\u03b5\u03bd\u03c9\u03bd tokens \u03ce\u03c3\u03c4\u03b5 \u03bd\u03b1 \u03bc\u03b7\u03bd \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03af\u03b6\u03b5\u03b9 \u03be\u03b1\u03bd\u03ac \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03bf \u03c4\u03bf \u03b9\u03c3\u03c4\u03bf\u03c1\u03b9\u03ba\u03cc \u03c3\u03b5 \u03ba\u03ac\u03b8\u03b5 \u03bd\u03ad\u03bf \u03b2\u03ae\u03bc\u03b1. \u0397 \u03b5\u03be\u03bf\u03b9\u03ba\u03bf\u03bd\u03cc\u03bc\u03b7\u03c3\u03b7 \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03b9\u03c3\u03bc\u03bf\u03cd \u03ad\u03c7\u03b5\u03b9 \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03bc\u03bd\u03ae\u03bc\u03b7\u03c2: \u03c4\u03bf \u03bc\u03ad\u03b3\u03b5\u03b8\u03bf\u03c2 \u03c4\u03b7\u03c2 KV cache \u03b1\u03c5\u03be\u03ac\u03bd\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b1\u03b6\u03af \u03bc\u03b5 \u03c4\u03bf \u03bc\u03ae\u03ba\u03bf\u03c2 \u03c4\u03bf\u03c5 context, \u03c4\u03bf \u03c0\u03bb\u03ae\u03b8\u03bf\u03c2 \u03c4\u03c9\u03bd \u03b5\u03bd\u03b5\u03c1\u03b3\u03ce\u03bd requests, \u03c4\u03b1 attention layers, \u03c4\u03b1 KV heads \u03ba\u03b1\u03b9 \u03c4\u03b7 \u03b4\u03b9\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 \u03ba\u03ac\u03b8\u03b5 head.<\/p>\n<p>\u03a3\u03c4\u03bf Qwen3.6-27B \u03c4\u03b7\u03c2 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7\u03c2, 16 \u03b1\u03c0\u03cc \u03c4\u03b1 64 language layers \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03bf\u03cd\u03bd full attention \u03ba\u03b1\u03b9 \u03c4\u03b1 \u03c5\u03c0\u03cc\u03bb\u03bf\u03b9\u03c0\u03b1 Gated DeltaNet. \u0391\u03ba\u03cc\u03bc\u03b7 \u03ba\u03b1\u03b9 \u03c3\u03b5 \u03b1\u03c5\u03c4\u03ae \u03c4\u03b7\u03bd \u03c5\u03b2\u03c1\u03b9\u03b4\u03b9\u03ba\u03ae \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae, \u03c4\u03bf attention KV payload \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c3\u03b5 64 KiB \u03b1\u03bd\u03ac live token \u03b3\u03b9\u03b1 BF16 \u03ba\u03b1\u03b9 32 KiB \u03b3\u03b9\u03b1 FP8. \u0391\u03c5\u03c4\u03cc \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 2 GiB \u03ae 1 GiB \u03b1\u03bd\u03c4\u03af\u03c3\u03c4\u03bf\u03b9\u03c7\u03b1 \u03b3\u03b9\u03b1 \u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03af\u03b1 32K tokens, \u03c0\u03c1\u03b9\u03bd \u03c0\u03c1\u03bf\u03c3\u03c4\u03b5\u03b8\u03bf\u03cd\u03bd weights, recurrent state, allocator reserve, scratch \u03ba\u03b1\u03b9 \u03ac\u03bb\u03bb\u03b5\u03c2 \u03b1\u03c0\u03b1\u03b9\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2 \u03c4\u03bf\u03c5 runtime.<\/p>\n<p>\u03a4\u03bf PagedAttention \u03bc\u03b5\u03b9\u03ce\u03bd\u03b5\u03b9 \u03c4\u03bf\u03bd \u03ba\u03b1\u03c4\u03b1\u03ba\u03b5\u03c1\u03bc\u03b1\u03c4\u03b9\u03c3\u03bc\u03cc \u03ba\u03b1\u03b9 \u03b5\u03c0\u03b9\u03c4\u03c1\u03ad\u03c0\u03b5\u03b9 \u03c3\u03b5 KV blocks \u03bd\u03b1 \u03b2\u03c1\u03af\u03c3\u03ba\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c3\u03b5 \u03bc\u03b7 \u03c3\u03c5\u03bd\u03b5\u03c7\u03cc\u03bc\u03b5\u03bd\u03b5\u03c2 \u03c6\u03c5\u03c3\u03b9\u03ba\u03ad\u03c2 \u03b8\u03ad\u03c3\u03b5\u03b9\u03c2, \u03b1\u03bb\u03bb\u03ac \u03b4\u03b5\u03bd \u03c3\u03c5\u03bc\u03c0\u03b9\u03ad\u03b6\u03b5\u03b9 \u03b1\u03c0\u03cc \u03bc\u03cc\u03bd\u03bf \u03c4\u03bf\u03c5 \u03c4\u03b1 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1 \u03ba\u03ac\u03b8\u03b5 block. \u0397 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ac \u03ad\u03c7\u03b5\u03b9 \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03ae \u03c3\u03b7\u03bc\u03b1\u03c3\u03af\u03b1: \u03ac\u03bb\u03bb\u03bf \u03b7 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b7 \u03b4\u03b9\u03b1\u03c7\u03b5\u03af\u03c1\u03b9\u03c3\u03b7 \u03c4\u03bf\u03c5 \u03b4\u03b9\u03b1\u03b8\u03ad\u03c3\u03b9\u03bc\u03bf\u03c5 \u03c7\u03ce\u03c1\u03bf\u03c5 \u03ba\u03b1\u03b9 \u03ac\u03bb\u03bb\u03bf \u03b7 \u03bc\u03b5\u03af\u03c9\u03c3\u03b7 \u03c4\u03c9\u03bd bytes \u03c0\u03bf\u03c5 \u03ba\u03b1\u03c4\u03b1\u03bb\u03b1\u03bc\u03b2\u03ac\u03bd\u03b5\u03b9 \u03b7 \u03af\u03b4\u03b9\u03b1 \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1. \u0393\u03b9\u03b1 \u03c4\u03bf \u03b3\u03b9\u03b1\u03c4\u03af \u03c4\u03bf \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf context \u03b4\u03b5\u03bd \u03b9\u03c3\u03bf\u03b4\u03c5\u03bd\u03b1\u03bc\u03b5\u03af \u03b1\u03c5\u03c4\u03bf\u03bc\u03ac\u03c4\u03c9\u03c2 \u03bc\u03b5 \u03b1\u03be\u03b9\u03cc\u03c0\u03b9\u03c3\u03c4\u03b7 \u03b1\u03bd\u03ac\u03ba\u03bb\u03b7\u03c3\u03b7, \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf \u03c3\u03c5\u03bc\u03c0\u03bb\u03ae\u03c1\u03c9\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b7 \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03c4\u03bf\u03c5 <a href=\"https:\/\/twodots.gr\/megalo-context-den-simainei-axiopisto-ai-predicatelongbench\/\">PredicateLongBench<\/a>.<\/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\">Paging<\/p>\n<p>\u03a4\u03bf PagedAttention \u03bf\u03c1\u03b3\u03b1\u03bd\u03ce\u03bd\u03b5\u03b9 \u03c4\u03b7\u03bd KV cache \u03c3\u03b5 blocks \u03ba\u03b1\u03b9 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03af\u03b6\u03b5\u03b9 \u03c4\u03b7 \u03c3\u03c0\u03b1\u03c4\u03ac\u03bb\u03b7 \u03b1\u03c0\u03cc \u03ba\u03b1\u03c4\u03b1\u03ba\u03b5\u03c1\u03bc\u03b1\u03c4\u03b9\u03c3\u03bc\u03cc \u03ba\u03b1\u03b9 \u03b4\u03b5\u03c3\u03bc\u03b5\u03cd\u03c3\u03b5\u03b9\u03c2 \u03bc\u03ad\u03b3\u03b9\u03c3\u03c4\u03bf\u03c5 \u03bc\u03ae\u03ba\u03bf\u03c5\u03c2.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">Layout<\/span><span class=\"td-badge\">Allocation<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card td-platform-card--navy\">\n<p class=\"td-comparison-title\">Compression<\/p>\n<p>\u03a4\u03bf Minima-KV \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b7 \u03c6\u03c5\u03c3\u03b9\u03ba\u03ae \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 \u03c0\u03b1\u03bb\u03b9\u03cc\u03c4\u03b5\u03c1\u03c9\u03bd \u03c3\u03b5\u03bb\u03af\u03b4\u03c9\u03bd \u03b1\u03c0\u03cc FP8 \u03c3\u03b5 TQ3 \u03ce\u03c3\u03c4\u03b5 \u03bd\u03b1 \u03bc\u03b5\u03b9\u03ce\u03c3\u03b5\u03b9 \u03c4\u03b1 bytes \u03c4\u03b7\u03c2 \u03b5\u03bd\u03b5\u03c1\u03b3\u03ae\u03c2 cache.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">FP8<\/span><span class=\"td-badge\">TQ3<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card\">\n<p class=\"td-comparison-title\">Eviction<\/p>\n<p>\u0397 \u03b4\u03b9\u03b1\u03b3\u03c1\u03b1\u03c6\u03ae \u03ba\u03c1\u03b1\u03c4\u03ac \u03bc\u03cc\u03bd\u03bf \u03b5\u03c0\u03b9\u03bb\u03b5\u03b3\u03bc\u03ad\u03bd\u03bf state. \u0395\u03bb\u03b5\u03c5\u03b8\u03b5\u03c1\u03ce\u03bd\u03b5\u03b9 \u03c7\u03ce\u03c1\u03bf, \u03b1\u03bb\u03bb\u03ac \u03bc\u03b9\u03b1 \u03c3\u03b5\u03bb\u03af\u03b4\u03b1 \u03c0\u03bf\u03c5 \u03b1\u03c0\u03bf\u03bc\u03b1\u03ba\u03c1\u03cd\u03bd\u03b8\u03b7\u03ba\u03b5 \u03b4\u03b5\u03bd \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03bd\u03b1\u03ba\u03c4\u03b7\u03b8\u03b5\u03af \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd KV cache.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">Sparse state<\/span><span class=\"td-badge\">Irreversible<\/span><\/div>\n<\/div>\n<\/div>\n<\/div>\n<h2 id=\"tria-epipeda\">Recent, Anchor \u03ba\u03b1\u03b9 Stale: \u03b7 \u03b9\u03b5\u03c1\u03b1\u03c1\u03c7\u03af\u03b1 \u03c4\u03bf\u03c5 Minima-KV<\/h2>\n<p>\u03a4\u03bf Minima-KV \u03c7\u03c9\u03c1\u03af\u03b6\u03b5\u03b9 \u03c4\u03b9\u03c2 \u03b5\u03bd\u03b5\u03c1\u03b3\u03ad\u03c2 \u03c3\u03b5\u03bb\u03af\u03b4\u03b5\u03c2 \u03c3\u03b5 \u03c4\u03c1\u03af\u03b1 tiers. \u039f\u03b9 <strong>Recent<\/strong> \u03b5\u03af\u03bd\u03b1\u03b9 \u03bf\u03b9 \u03c0\u03b9\u03bf \u03c0\u03c1\u03cc\u03c3\u03c6\u03b1\u03c4\u03b5\u03c2 \u03ba\u03b1\u03b9 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03bf\u03c5\u03bd \u03c3\u03b5 FP8 \u03b3\u03b9\u03b1 \u03bd\u03b1 \u03c0\u03c1\u03bf\u03c3\u03c4\u03b1\u03c4\u03b5\u03cd\u03b5\u03c4\u03b1\u03b9 \u03c4\u03bf \u03c4\u03c1\u03ad\u03c7\u03bf\u03bd \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03b7\u03c2 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae\u03c2. \u039f\u03b9 <strong>Anchor<\/strong> \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03b1\u03bb\u03b9\u03cc\u03c4\u03b5\u03c1\u03b5\u03c2 \u03b1\u03bb\u03bb\u03ac \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03b5\u03c2 \u03c3\u03b5\u03bb\u03af\u03b4\u03b5\u03c2, \u03cc\u03c0\u03c9\u03c2 system instructions, \u03b1\u03c1\u03c7\u03b9\u03ba\u03ac attention sinks, \u03b1\u03bd\u03b1\u03ba\u03c4\u03b7\u03bc\u03ad\u03bd\u03b1 \u03c4\u03b5\u03ba\u03bc\u03ae\u03c1\u03b9\u03b1 \u03ae identifiers \u03c0\u03bf\u03c5 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03ae\u03c3\u03bf\u03c5\u03bd \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b7 \u03c0\u03b9\u03c3\u03c4\u03cc\u03c4\u03b7\u03c4\u03b1.<\/p>\n<p>\u0421\u0430\u0439\u0442\u044a\u0442 <strong>Stale<\/strong> \u03c3\u03b5\u03bb\u03af\u03b4\u03b5\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03b1\u03bb\u03b9\u03cc\u03c4\u03b5\u03c1\u03b5\u03c2 \u03ba\u03b1\u03b9 \u03b5\u03ba\u03c4\u03cc\u03c2 \u03c4\u03bf\u03c5 \u03c0\u03c1\u03bf\u03c3\u03c4\u03b1\u03c4\u03b5\u03c5\u03bc\u03ad\u03bd\u03bf\u03c5 budget. \u039a\u03c9\u03b4\u03b9\u03ba\u03bf\u03c0\u03bf\u03b9\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9 \u03c3\u03b5 TQ3 \u03bc\u03b5 packed \u03c4\u03c1\u03b9\u03ce\u03bd bit \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7, scales, norm corrections \u03ba\u03b1\u03b9 metadata. \u03a0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03bf\u03c5\u03bd \u03cc\u03bc\u03c9\u03c2 addressable \u03bc\u03ad\u03c3\u03b1 \u03c3\u03c4\u03bf \u03b5\u03bd\u03b5\u03c1\u03b3\u03cc request. \u0391\u03bd \u03c4\u03bf controller \u03ba\u03c1\u03af\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03bc\u03b9\u03b1 Stale \u03c3\u03b5\u03bb\u03af\u03b4\u03b1 \u03ad\u03b3\u03b9\u03bd\u03b5 \u03be\u03b1\u03bd\u03ac \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03ae, \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c4\u03b7\u03bd \u03b1\u03bd\u03b1\u03ba\u03b1\u03c4\u03b1\u03c3\u03ba\u03b5\u03c5\u03ac\u03c3\u03b5\u03b9 \u03c3\u03b5 \u03bd\u03ad\u03b1 FP8 \u03c3\u03b5\u03bb\u03af\u03b4\u03b1 \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03c4\u03b7\u03bd \u03c0\u03c1\u03bf\u03b1\u03b3\u03ac\u03b3\u03b5\u03b9 \u03c3\u03c4\u03bf Anchor tier.<\/p>\n<p>\u03a4\u03bf paper \u03c0\u03b5\u03c1\u03b9\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 attention scoring \u03b3\u03b9\u03b1 \u03c4\u03ad\u03c4\u03bf\u03b9\u03b5\u03c2 \u03b1\u03c0\u03bf\u03c6\u03ac\u03c3\u03b5\u03b9\u03c2, \u03b1\u03bb\u03bb\u03ac \u03bf \u03bc\u03b7\u03c7\u03b1\u03bd\u03b9\u03c3\u03bc\u03cc\u03c2 \u03ae\u03c4\u03b1\u03bd \u03b1\u03c0\u03b5\u03bd\u03b5\u03c1\u03b3\u03bf\u03c0\u03bf\u03b9\u03b7\u03bc\u03ad\u03bd\u03bf\u03c2 \u03c3\u03c4\u03b1 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b7\u03bc\u03ad\u03bd\u03b1 Qwen3.6 profiles. \u0395\u03c0\u03bf\u03bc\u03ad\u03bd\u03c9\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03bf\u03c5 \u03c3\u03c7\u03b5\u03b4\u03b9\u03b1\u03c3\u03bc\u03bf\u03cd \u03c4\u03bf\u03c5 controller \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03bc\u03b5\u03c4\u03c1\u03b7\u03bc\u03ad\u03bd\u03bf \u03c0\u03bb\u03b5\u03bf\u03bd\u03ad\u03ba\u03c4\u03b7\u03bc\u03b1 \u03c4\u03c9\u03bd \u03b4\u03b7\u03bc\u03bf\u03c3\u03b9\u03b5\u03c5\u03bc\u03ad\u03bd\u03c9\u03bd \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03b5\u03c3\u03bc\u03ac\u03c4\u03c9\u03bd.<\/p>\n<div class=\"td-chart td-chart--metrics\">\n<div class=\"td-chart-head\">\n<p class=\"td-chart-kicker\">\u03a4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b1 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ac \u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af\u03b1<\/p>\n<p class=\"td-chart-title\">\u039f\u03b9 \u03b1\u03c1\u03b9\u03b8\u03bc\u03bf\u03af \u03b4\u03b5\u03bd \u03c0\u03c1\u03bf\u03ad\u03c1\u03c7\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03ad\u03bd\u03b1 \u03b5\u03bd\u03b9\u03b1\u03af\u03bf benchmark<\/p>\n<p class=\"td-chart-intro\">\u039a\u03ac\u03b8\u03b5 \u03c4\u03b9\u03bc\u03ae \u03b1\u03bd\u03ae\u03ba\u03b5\u03b9 \u03c3\u03c4\u03bf \u03b4\u03b9\u03ba\u03cc \u03c4\u03b7\u03c2 configuration-bound profile \u03ba\u03b1\u03b9 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b4\u03b9\u03b1\u03b2\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bc\u03cc\u03bd\u03bf \u03bc\u03ad\u03c3\u03b1 \u03c3\u03b5 \u03b1\u03c5\u03c4\u03cc \u03c4\u03bf scope.<\/p>\n<\/div>\n<div class=\"td-metric-grid\">\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">18,3 KiB\/token<\/span><span class=\"td-metric-label\">Owner-reported deployment aggregate \u03b3\u03b9\u03b1 attention KV \u03b1\u03bd\u03ac live token<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">3,50\u00d7<\/span><span class=\"td-metric-label\">\u03a5\u03c0\u03bf\u03bb\u03bf\u03b3\u03b9\u03c3\u03bc\u03ad\u03bd\u03b7 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03c4\u03bf\u03c5 aggregate \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 BF16 attention KV<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">3,625\u00d7<\/span><span class=\"td-metric-label\">Active-KV \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03c3\u03b5 \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03cc direct-decode canary<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">0,9821\u00d7<\/span><span class=\"td-metric-label\">Throughput ratio \u03c4\u03bf\u03c5 \u03af\u03b4\u03b9\u03bf\u03c5 \u03bc\u03bf\u03bd\u03b1\u03b4\u03b9\u03ba\u03bf\u03cd direct pair \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 control<\/span><\/div>\n<\/div>\n<\/div>\n<h2 id=\"retention-preserving\">\u03a4\u03b9 \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03cc\u03c4\u03b1\u03bd \u03b7 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03b4\u03b5\u03bd \u03b4\u03b9\u03b1\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 \u03c4\u03bf \u03c0\u03b1\u03bb\u03b9\u03cc context<\/h2>\n<p>\u0397 \u03ad\u03bd\u03bd\u03bf\u03b9\u03b1 <em>retention-preserving<\/em> \u03b5\u03af\u03bd\u03b1\u03b9 \u03bf \u03c0\u03c5\u03c1\u03ae\u03bd\u03b1\u03c2 \u03c4\u03b7\u03c2 \u03c0\u03c1\u03cc\u03c4\u03b1\u03c3\u03b7\u03c2. Sparse retention \u03ba\u03b1\u03b9 eviction \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03ba\u03c1\u03b1\u03c4\u03ae\u03c3\u03bf\u03c5\u03bd \u03bc\u03cc\u03bd\u03bf \u03c4\u03b9\u03c2 \u03b8\u03ad\u03c3\u03b5\u03b9\u03c2 \u03c0\u03bf\u03c5 \u03b8\u03b5\u03c9\u03c1\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9 \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03ad\u03c2 \u03c3\u03b5 \u03bc\u03b9\u03b1 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b7 \u03c3\u03c4\u03b9\u03b3\u03bc\u03ae. \u0391\u03bd \u03cc\u03bc\u03c9\u03c2 \u03b7 \u03c0\u03c1\u03bf\u03c3\u03bf\u03c7\u03ae \u03bc\u03b5\u03c4\u03b1\u03ba\u03b9\u03bd\u03b7\u03b8\u03b5\u03af \u03b1\u03c1\u03b3\u03cc\u03c4\u03b5\u03c1\u03b1 \u03c3\u03b5 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c7\u03ae \u03c0\u03bf\u03c5 \u03b5\u03af\u03c7\u03b5 \u03b4\u03b9\u03b1\u03b3\u03c1\u03b1\u03c6\u03b5\u03af, \u03b7 \u03b1\u03bd\u03c4\u03af\u03c3\u03c4\u03bf\u03b9\u03c7\u03b7 KV \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1 \u03b4\u03b5\u03bd \u03c5\u03c0\u03ac\u03c1\u03c7\u03b5\u03b9 \u03c0\u03bb\u03ad\u03bf\u03bd.<\/p>\n<p>\u03a4\u03bf Minima-KV \u03b5\u03c0\u03b9\u03bb\u03ad\u03b3\u03b5\u03b9 \u03c0\u03b9\u03bf \u03c3\u03c5\u03bd\u03c4\u03b7\u03c1\u03b7\u03c4\u03b9\u03ba\u03cc \u03c3\u03c5\u03bc\u03b2\u03b9\u03b2\u03b1\u03c3\u03bc\u03cc: \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 \u03c3\u03b5\u03bb\u03af\u03b4\u03b1 \u03b5\u03bd\u03cc\u03c2 live request \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03b5\u03af \u03b1\u03ba\u03c1\u03b9\u03b2\u03ae \u03ae \u03c0\u03c1\u03bf\u03c3\u03b5\u03b3\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7. \u0391\u03c5\u03c4\u03cc \u03c4\u03b1\u03b9\u03c1\u03b9\u03ac\u03b6\u03b5\u03b9 \u03c3\u03b5 workloads \u03cc\u03c0\u03bf\u03c5 \u03bc\u03b9\u03b1 \u03c0\u03b1\u03bb\u03b9\u03cc\u03c4\u03b5\u03c1\u03b7 \u03bf\u03b4\u03b7\u03b3\u03af\u03b1, \u03c1\u03ae\u03c4\u03c1\u03b1, \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7 \u03ae \u03c0\u03b7\u03b3\u03ae \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b3\u03af\u03bd\u03b5\u03b9 \u03be\u03b1\u03bd\u03ac \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03ae \u03c7\u03c9\u03c1\u03af\u03c2 \u03c0\u03c1\u03bf\u03b5\u03b9\u03b4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7. \u0394\u03b5\u03bd \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03b7 TQ3 \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b9\u03c3\u03bf\u03b4\u03cd\u03bd\u03b1\u03bc\u03b7 \u03bc\u03b5 \u03c4\u03b7\u03bd \u03c0\u03c5\u03ba\u03bd\u03ae cache\u00b7 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03b1\u03c0\u03bf\u03c6\u03b5\u03cd\u03b3\u03b5\u03c4\u03b1\u03b9 \u03b7 \u03bc\u03b7 \u03b1\u03bd\u03b1\u03c3\u03c4\u03c1\u03ad\u03c8\u03b9\u03bc\u03b7 \u03b1\u03c6\u03b1\u03af\u03c1\u03b5\u03c3\u03b7 \u03c4\u03b7\u03c2 \u03c3\u03b5\u03bb\u03af\u03b4\u03b1\u03c2.<\/p>\n<p>\u0397 \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae \u03b5\u03be\u03b1\u03c1\u03c4\u03ac\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf failure cost. \u03a3\u03b5 \u03bc\u03b9\u03b1 \u03c7\u03b1\u03bc\u03b7\u03bb\u03bf\u03cd \u03c1\u03af\u03c3\u03ba\u03bf\u03c5 \u03c0\u03b5\u03c1\u03af\u03bb\u03b7\u03c8\u03b7 \u03af\u03c3\u03c9\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03bf\u03b4\u03b5\u03ba\u03c4\u03cc \u03bd\u03b1 \u03b1\u03c6\u03b1\u03b9\u03c1\u03b5\u03b8\u03bf\u03cd\u03bd \u03bb\u03b5\u03c0\u03c4\u03bf\u03bc\u03ad\u03c1\u03b5\u03b9\u03b5\u03c2. \u03a3\u03b5 \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03ac \u03ad\u03b3\u03b3\u03c1\u03b1\u03c6\u03b1, code analysis \u03ae \u03c0\u03bf\u03bb\u03cd\u03b2\u03b7\u03bc\u03b1 AI agent, \u03b7 \u03b4\u03c5\u03bd\u03b1\u03c4\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b5\u03c0\u03b9\u03c3\u03c4\u03c1\u03bf\u03c6\u03ae\u03c2 \u03c3\u03b5 \u03c0\u03b1\u03bb\u03b9\u03cc\u03c4\u03b5\u03c1\u03bf evidence \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03b9\u03bf \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03ae \u03b1\u03c0\u03cc \u03c4\u03b7 \u03bc\u03ad\u03b3\u03b9\u03c3\u03c4\u03b7 \u03b5\u03be\u03bf\u03b9\u03ba\u03bf\u03bd\u03cc\u03bc\u03b7\u03c3\u03b7. \u0397 \u03c3\u03c7\u03b5\u03c4\u03b9\u03ba\u03ae \u03c3\u03c5\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7 \u03b3\u03b9\u03b1 \u03bc\u03bd\u03ae\u03bc\u03b7 \u03bc\u03b5 provenance \u03b1\u03bd\u03b1\u03c0\u03c4\u03cd\u03c3\u03c3\u03b5\u03c4\u03b1\u03b9 \u03ba\u03b1\u03b9 \u03c3\u03c4\u03bf <a href=\"https:\/\/twodots.gr\/echo-mnimi-ai-agent-apodeixeis-provenance\/\">ECHO \u03b3\u03b9\u03b1 \u03bc\u03bd\u03ae\u03bc\u03b7 AI agents<\/a>.<\/p>\n<h2 id=\"mixed-format-attention\">\u03a0\u03ce\u03c2 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b5\u03af \u03c4\u03bf mixed-format attention<\/h2>\n<p>\u0397 \u03b1\u03c0\u03bb\u03ae \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03b4\u03b5\u03bd \u03b1\u03c1\u03ba\u03b5\u03af \u03b1\u03bd \u03ba\u03ac\u03b8\u03b5 decode \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03c0\u03c1\u03ce\u03c4\u03b1 \u03bd\u03b1 \u03b1\u03c0\u03bf\u03c3\u03c5\u03bc\u03c0\u03b9\u03ad\u03c3\u03b5\u03b9 \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03b7 \u03c4\u03b7\u03bd cache \u03c3\u03b5 \u03ad\u03bd\u03b1 \u03c0\u03c5\u03ba\u03bd\u03cc \u03b1\u03bd\u03c4\u03af\u03b3\u03c1\u03b1\u03c6\u03bf. \u03a4\u03bf Minima-KV \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af format-specific kernels: \u03ad\u03bd\u03b1 branch \u03b4\u03b9\u03b1\u03b2\u03ac\u03b6\u03b5\u03b9 FP8 \u03c3\u03b5\u03bb\u03af\u03b4\u03b5\u03c2 \u03ba\u03b1\u03b9 \u03ad\u03bd\u03b1 \u03ac\u03bb\u03bb\u03bf packed TQ3 \u03c3\u03b5\u03bb\u03af\u03b4\u03b5\u03c2 \u03b1\u03c0\u03b5\u03c5\u03b8\u03b5\u03af\u03b1\u03c2 \u03ba\u03b1\u03b8\u03ce\u03c2 \u03c6\u03bf\u03c1\u03c4\u03ce\u03bd\u03bf\u03bd\u03c4\u03b1\u03b9.<\/p>\n<p>\u039a\u03ac\u03b8\u03b5 branch \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03af\u03b6\u03b5\u03b9 \u03bc\u03b5\u03c1\u03b9\u03ba\u03ae \u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 attention \u2014 \u03bc\u03ad\u03b3\u03b9\u03c3\u03c4\u03bf score, \u03ac\u03b8\u03c1\u03bf\u03b9\u03c3\u03bc\u03b1 \u03b5\u03ba\u03b8\u03b5\u03c4\u03b9\u03ba\u03ce\u03bd \u03ba\u03b1\u03b9 \u03ba\u03b1\u03bd\u03bf\u03bd\u03b9\u03ba\u03bf\u03c0\u03bf\u03b9\u03b7\u03bc\u03ad\u03bd\u03bf output. \u039f\u03b9 \u03ba\u03b1\u03c4\u03b1\u03c3\u03c4\u03ac\u03c3\u03b5\u03b9\u03c2 \u03c3\u03c5\u03b3\u03c7\u03c9\u03bd\u03b5\u03cd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03ad\u03c0\u03b5\u03b9\u03c4\u03b1 \u03bc\u03b5 \u03ba\u03bf\u03b9\u03bd\u03ae online-softmax \u03ba\u03b1\u03bd\u03bf\u03bd\u03b9\u03ba\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7, \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03ce\u03bd\u03c4\u03b1\u03c2 \u03c4\u03b7 \u03bb\u03bf\u03b3\u03b9\u03ba\u03ae \u03c3\u03b5\u03b9\u03c1\u03ac \u03c4\u03c9\u03bd \u03c3\u03b5\u03bb\u03af\u03b4\u03c9\u03bd \u03ba\u03b1\u03b9 \u03c4\u03bf causal mask. \u03a3\u03c4\u03bf direct canary, \u03b7 dense shadow cache \u03b5\u03af\u03c7\u03b5 \u03b1\u03c0\u03bf\u03b4\u03b5\u03c3\u03bc\u03b5\u03c5\u03c4\u03b5\u03af, \u03bf\u03c0\u03cc\u03c4\u03b5 \u03b7 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03b4\u03b5\u03bd \u03ae\u03c4\u03b1\u03bd \u03b1\u03c0\u03bb\u03ce\u03c2 \u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03cc \u03c4\u03ad\u03c7\u03bd\u03b1\u03c3\u03bc\u03b1 \u03c0\u03af\u03c3\u03c9 \u03b1\u03c0\u03cc \u03ad\u03bd\u03b1 \u03c0\u03bb\u03ae\u03c1\u03b5\u03c2 \u03c0\u03c5\u03ba\u03bd\u03cc \u03b1\u03bd\u03c4\u03af\u03b3\u03c1\u03b1\u03c6\u03bf.<\/p>\n<p>\u0397 \u03c5\u03bb\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af TileLang kernels \u03c0\u03bf\u03c5 \u03bc\u03b5\u03c4\u03b1\u03b3\u03bb\u03c9\u03c4\u03c4\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c3\u03b5 CUDA \u03ba\u03b1\u03b9 \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 route counters \u03ba\u03b1\u03b9 fallbacks. \u0391\u03c5\u03c4\u03ae \u03b7 \u03c0\u03b1\u03c1\u03b1\u03c4\u03b7\u03c1\u03b7\u03c3\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03bf\u03c5\u03c3\u03b9\u03ce\u03b4\u03b7\u03c2: \u03ad\u03bd\u03b1 prompt \u03c0\u03bf\u03c5 \u03ad\u03bc\u03b5\u03b9\u03bd\u03b5 \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03bf \u03c3\u03b5 FP8 \u03b4\u03b5\u03bd \u03b8\u03b1 \u03ad\u03c0\u03c1\u03b5\u03c0\u03b5 \u03bd\u03b1 \u03c0\u03b1\u03c1\u03bf\u03c5\u03c3\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c9\u03c2 \u03b1\u03c0\u03cc\u03b4\u03b5\u03b9\u03be\u03b7 \u03cc\u03c4\u03b9 \u03b5\u03ba\u03c4\u03b5\u03bb\u03ad\u03c3\u03c4\u03b7\u03ba\u03b5 \u03c4\u03bf TQ3 path. \u03a0\u03b1\u03c1\u03cc\u03bc\u03bf\u03b9\u03b1, \u03ad\u03bd\u03b1 \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03bf kernel microbenchmark \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ae \u03bc\u03ad\u03c3\u03b1 \u03c3\u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf, \u03cc\u03c0\u03c9\u03c2 \u03b5\u03be\u03b7\u03b3\u03b5\u03af \u03b7 \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03b3\u03b9\u03b1 \u03c4\u03bf <a href=\"https:\/\/twodots.gr\/llm4llm-gpu-kernel-dokimi-pragmatiko-montelo\/\">LLM4LLM \u03ba\u03b1\u03b9 \u03c4\u03b1 end-to-end kernel tests<\/a>.<\/p>\n<h2 id=\"metriseis-mnimis\">\u03a4\u03b9 \u03ad\u03b4\u03b5\u03b9\u03be\u03b1\u03bd \u03bf\u03b9 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03bc\u03bd\u03ae\u03bc\u03b7\u03c2<\/h2>\n<p>\u03a4\u03bf deployment accounting \u03b1\u03bd\u03b1\u03c6\u03ad\u03c1\u03b5\u03b9 18,3 KiB attention KV \u03b1\u03bd\u03ac live token. \u039c\u03b5 \u03b2\u03ac\u03c3\u03b7 \u03c4\u03b1 64 KiB \u03c4\u03bf\u03c5 BF16 \u03ba\u03b1\u03b9 \u03c4\u03b1 32 KiB \u03c4\u03bf\u03c5 FP8 \u03b3\u03b9\u03b1 \u03c4\u03b7 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7 \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae, \u03b7 \u03c4\u03b9\u03bc\u03ae \u03b1\u03bd\u03c4\u03b9\u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af \u03c3\u03b5 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 3,497\u00d7 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf footprint \u03b1\u03c0\u03cc BF16 \u03ba\u03b1\u03b9 1,749\u00d7 \u03b1\u03c0\u03cc FP8. \u03a3\u03c4\u03b7\u03bd \u03b1\u03bd\u03b1\u03bb\u03c5\u03c4\u03b9\u03ba\u03ae \u03c0\u03c1\u03bf\u03b2\u03bf\u03bb\u03ae \u03b3\u03b9\u03b1 \u03ad\u03bd\u03b1 \u03b5\u03ba\u03b1\u03c4\u03bf\u03bc\u03bc\u03cd\u03c1\u03b9\u03bf live tokens, \u03bf\u03b9 \u03c4\u03b9\u03bc\u03ad\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 61,04 GiB \u03b3\u03b9\u03b1 BF16, 30,52 GiB \u03b3\u03b9\u03b1 FP8 \u03ba\u03b1\u03b9 17,45 GiB \u03b3\u03b9\u03b1 \u03c4\u03bf \u03c3\u03b5\u03bd\u03ac\u03c1\u03b9\u03bf \u03c4\u03c9\u03bd 18,3 KiB \u03b1\u03bd\u03ac token.<\/p>\n<p>\u039f\u03b9 \u03c0\u03c1\u03bf\u03b2\u03bf\u03bb\u03ad\u03c2 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03b5\u03c2 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03c3\u03b5 \u03ba\u03ac\u03b8\u03b5 \u03bc\u03ae\u03ba\u03bf\u03c2 context. \u0397 \u03b4\u03b9\u03b1\u03b8\u03ad\u03c3\u03b9\u03bc\u03b7 \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03b1\u03c6\u03ae \u03b4\u03b5\u03bd \u03b1\u03bd\u03b1\u03bb\u03cd\u03b5\u03b9 \u03c0\u03bb\u03ae\u03c1\u03c9\u03c2 \u03c4\u03bf aggregate \u03c3\u03b5 Recent, Anchor \u03ba\u03b1\u03b9 Stale bytes \u03bf\u03cd\u03c4\u03b5 \u03b1\u03c0\u03bf\u03bc\u03bf\u03bd\u03ce\u03bd\u03b5\u03b9 metadata, scratch, occupancy \u03ba\u03b1\u03b9 context-dependent \u03b5\u03c0\u03b9\u03b4\u03c1\u03ac\u03c3\u03b5\u03b9\u03c2. \u0393\u03b9\u03b1 \u03c4\u03bf\u03bd \u03af\u03b4\u03b9\u03bf \u03bb\u03cc\u03b3\u03bf, \u03bf \u03b9\u03c3\u03c4\u03bf\u03c1\u03b9\u03ba\u03cc\u03c2 \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03b9\u03c3\u03bc\u03cc\u03c2 53 resident contexts \u03c4\u03c9\u03bd 32K \u03bc\u03b5 FP8 \u03ba\u03b1\u03b9 84 \u03bc\u03b5 \u03b1\u03ba\u03c1\u03b9\u03b2\u03ad\u03c2 \u03c3\u03b5\u03bd\u03ac\u03c1\u03b9\u03bf 3,50\u00d7 \u03b5\u03af\u03bd\u03b1\u03b9 memory ceiling, \u03cc\u03c7\u03b9 SLO \u03b3\u03b9\u03b1 84 \u03c4\u03b1\u03c5\u03c4\u03cc\u03c7\u03c1\u03bf\u03bd\u03b1 decoding requests.<\/p>\n<div class=\"td-decision-band\">\n<p class=\"td-decision-kicker\">\u039a\u03b1\u03bd\u03cc\u03bd\u03b1\u03c2 \u03c7\u03c9\u03c1\u03b7\u03c4\u03b9\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2<\/p>\n<p class=\"td-decision-title\">\u039c\u03b7\u03bd \u03b5\u03be\u03b9\u03c3\u03ce\u03bd\u03b5\u03c4\u03b5 \u03c4\u03bf\u03bd \u03bb\u03cc\u03b3\u03bf \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7\u03c2 \u03bc\u03b5 \u03af\u03c3\u03bf \u03c0\u03bf\u03bb\u03bb\u03b1\u03c0\u03bb\u03b1\u03c3\u03b9\u03b1\u03c3\u03bc\u03cc \u03c7\u03c1\u03b7\u03c3\u03c4\u03ce\u03bd<\/p>\n<p>\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 \u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ac live KV bytes, \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc HBM, admission rate, queue time, TTFT, inter-token latency \u03ba\u03b1\u03b9 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1. \u03a4\u03b1 weights, \u03c4\u03bf recurrent state, \u03c4\u03bf scratch, \u03bf scheduler \u03ba\u03b1\u03b9 \u03c4\u03bf \u03c0\u03c1\u03bf\u03c3\u03c6\u03b5\u03c1\u03cc\u03bc\u03b5\u03bd\u03bf \u03c6\u03bf\u03c1\u03c4\u03af\u03bf \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03bf\u03c5\u03bd \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03bf\u03b9 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03bf\u03af.<\/p>\n<\/div>\n<h2 id=\"poiotita\">\u03a0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1: \u03bc\u03b9\u03ba\u03c1\u03ad\u03c2 \u03b1\u03c0\u03bf\u03ba\u03bb\u03af\u03c3\u03b5\u03b9\u03c2, \u03cc\u03c7\u03b9 \u03b3\u03b5\u03bd\u03b9\u03ba\u03cc \u00ab\u03c7\u03c9\u03c1\u03af\u03c2 \u03b1\u03c0\u03ce\u03bb\u03b5\u03b9\u03b5\u03c2\u00bb<\/h2>\n<p>\u03a4\u03bf quality profile \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af dynamic retiering, \u03ba\u03ac\u03bd\u03b5\u03b9 materialization \u03c4\u03c9\u03bd TQ3 \u03c3\u03b5\u03bb\u03af\u03b4\u03c9\u03bd \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc \u03c4\u03bf attention \u03ba\u03b1\u03b9 \u03ad\u03c7\u03b5\u03b9 \u03b1\u03c0\u03b5\u03bd\u03b5\u03c1\u03b3\u03bf\u03c0\u03bf\u03b9\u03b7\u03bc\u03ad\u03bd\u03bf attention scoring. \u03a3\u03c4\u03b9\u03c2 \u03bf\u03ba\u03c4\u03ce RULER needle-in-a-haystack \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b5\u03c2, \u03bc\u03b5 32 \u03c0\u03b1\u03c1\u03b1\u03b4\u03b5\u03af\u03b3\u03bc\u03b1\u03c4\u03b1 \u03b1\u03bd\u03ac task \u03ba\u03b1\u03b9 \u03bc\u03ae\u03ba\u03bf\u03c2, \u03c4\u03bf Minima-KV \u03b9\u03c3\u03bf\u03c6\u03ac\u03c1\u03b9\u03c3\u03b5 \u03c4\u03bf dense control \u03c3\u03c4\u03b1 16K. \u03a3\u03c4\u03b1 8K \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03c3\u03b5 \u03c4\u03bf task-macro score \u03ba\u03b1\u03c4\u03ac 0,20 \u03c0\u03bf\u03c3\u03bf\u03c3\u03c4\u03b9\u03b1\u03af\u03b5\u03c2 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b5\u03c2, \u03b5\u03bd\u03ce \u03c3\u03c4\u03b1 4K \u03c5\u03c0\u03bf\u03c7\u03ce\u03c1\u03b7\u03c3\u03b5 \u03ba\u03b1\u03c4\u03ac 0,90 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b5\u03c2.<\/p>\n<p>\u03a3\u03c4\u03bf LongBench v2, \u03c0\u03ac\u03bd\u03c9 \u03c3\u03c4\u03bf \u03af\u03b4\u03b9\u03bf \u03c3\u03cd\u03bd\u03bf\u03bb\u03bf 503 \u03b5\u03c1\u03c9\u03c4\u03ae\u03c3\u03b5\u03c9\u03bd, \u03bf\u03b9 \u03b1\u03c0\u03bf\u03ba\u03bb\u03af\u03c3\u03b5\u03b9\u03c2 \u03ae\u03c4\u03b1\u03bd \u22120,80 \u03c0\u03bf\u03c3\u03bf\u03c3\u03c4\u03b9\u03b1\u03af\u03b5\u03c2 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b5\u03c2 \u03c3\u03c4\u03b1 16K, \u22120,60 \u03c3\u03c4\u03b1 32K \u03ba\u03b1\u03b9 \u22120,40 \u03c3\u03c4\u03b1 64K. \u0394\u03b5\u03bd \u03b4\u03af\u03bd\u03bf\u03bd\u03c4\u03b1\u03b9 paired-bootstrap confidence intervals \u03bf\u03cd\u03c4\u03b5 \u03c0\u03c1\u03bf\u03ba\u03b1\u03b8\u03bf\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03bf non-inferiority margin. \u03a4\u03bf \u03c3\u03c9\u03c3\u03c4\u03cc \u03c3\u03c5\u03bc\u03c0\u03ad\u03c1\u03b1\u03c3\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03bf\u03b9 \u03b1\u03c0\u03bf\u03ba\u03bb\u03af\u03c3\u03b5\u03b9\u03c2 \u03ad\u03bc\u03b5\u03b9\u03bd\u03b1\u03bd \u03bc\u03b9\u03ba\u03c1\u03ad\u03c2 \u03c3\u03c4\u03b1 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b1 tests, \u03cc\u03c7\u03b9 \u03cc\u03c4\u03b9 \u03b7 \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ae \u03b5\u03af\u03bd\u03b1\u03b9 \u03b3\u03b5\u03bd\u03b9\u03ba\u03ac lossless.<\/p>\n<p>\u039c\u03b9\u03b1 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae\u03c2 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b4\u03b9\u03ba\u03cc \u03c4\u03b7\u03c2 evaluation \u03c3\u03b5 \u03b3\u03bb\u03ce\u03c3\u03c3\u03b5\u03c2, \u03ad\u03b3\u03b3\u03c1\u03b1\u03c6\u03b1 \u03ba\u03b1\u03b9 failure modes \u03c0\u03bf\u03c5 \u03b1\u03bd\u03c4\u03b9\u03c3\u03c4\u03bf\u03b9\u03c7\u03bf\u03cd\u03bd \u03c3\u03c4\u03b7\u03bd \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ae \u03c4\u03b7\u03c2. \u0393\u03b9\u03b1 \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03ac \u03c3\u03c5\u03c3\u03c4\u03ae\u03bc\u03b1\u03c4\u03b1 \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7\u03c2, \u03b3\u03b9\u03b1 \u03c0\u03b1\u03c1\u03ac\u03b4\u03b5\u03b9\u03b3\u03bc\u03b1, \u03b7 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c0\u03b5\u03c1\u03b9\u03bb\u03b1\u03bc\u03b2\u03ac\u03bd\u03b5\u03b9 retrieval, grounded answer quality \u03ba\u03b1\u03b9 behavior \u03c3\u03b5 \u03b1\u03c0\u03bf\u03bc\u03b1\u03ba\u03c1\u03c5\u03c3\u03bc\u03ad\u03bd\u03b1 \u03c3\u03b7\u03bc\u03b5\u03af\u03b1 \u03c4\u03bf\u03c5 context \u2014 \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03bf \u03ad\u03bd\u03b1 \u03bc\u03ad\u03c3\u03bf benchmark score. \u03a4\u03bf \u03af\u03b4\u03b9\u03bf \u03ba\u03c1\u03b9\u03c4\u03ae\u03c1\u03b9\u03bf workload-specific validation \u03b5\u03bc\u03c6\u03b1\u03bd\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b7\u03bd \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03b3\u03b9\u03b1 <a href=\"https:\/\/twodots.gr\/semantic-compression-trees-rag-context\/\">Semantic Compression Trees \u03c3\u03c4\u03bf RAG<\/a>.<\/p>\n<h2 id=\"direct-decode-prefix\">Direct decode \u03ba\u03b1\u03b9 prefix reuse \u03b5\u03af\u03bd\u03b1\u03b9 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ad\u03c2<\/h2>\n<p>\u03a4\u03bf direct canary \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b5 \u03b4\u03cd\u03bf \u03b5\u03bd\u03b5\u03c1\u03b3\u03ac requests \u03bc\u03b5 prompt 59.008 tokens \u03c4\u03bf \u03ba\u03b1\u03b8\u03ad\u03bd\u03b1, CUDA graph replay, \u03c3\u03c4\u03b1\u03c4\u03b9\u03ba\u03cc \u03bc\u03b5\u03af\u03b3\u03bc\u03b1 FP8\/TQ3, \u03b1\u03c0\u03b5\u03bd\u03b5\u03c1\u03b3\u03bf\u03c0\u03bf\u03b9\u03b7\u03bc\u03ad\u03bd\u03bf dynamic retiering \u03ba\u03b1\u03b9 \u03c7\u03c9\u03c1\u03af\u03c2 dense shadow. \u039a\u03b1\u03c4\u03ad\u03b3\u03c1\u03b1\u03c8\u03b5 29,270 tokens \u03b1\u03bd\u03ac \u03b4\u03b5\u03c5\u03c4\u03b5\u03c1\u03cc\u03bb\u03b5\u03c0\u03c4\u03bf \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 29,804 \u03c4\u03bf\u03c5 control, throughput ratio 0,9821 \u03ba\u03b1\u03b9 active-KV \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 3,625\u00d7. \u039a\u03b1\u03b9 \u03c4\u03b1 16 full-attention layers \u03b1\u03ba\u03bf\u03bb\u03bf\u03cd\u03b8\u03b7\u03c3\u03b1\u03bd \u03c4\u03bf direct route, \u03bc\u03b5 \u03bc\u03b7\u03b4\u03b5\u03bd\u03b9\u03ba\u03ac fallbacks \u03c3\u03b5 32 request-layer instances.<\/p>\n<p>\u03a0\u03c1\u03cc\u03ba\u03b5\u03b9\u03c4\u03b1\u03b9 \u03b3\u03b9\u03b1 \u03ad\u03bd\u03b1 \u03bc\u03cc\u03bd\u03bf \u03b6\u03b5\u03cd\u03b3\u03bf\u03c2 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03c9\u03bd. \u0394\u03b5\u03bd \u03c5\u03c0\u03ac\u03c1\u03c7\u03bf\u03c5\u03bd \u03b5\u03c0\u03b1\u03bd\u03b1\u03bb\u03ae\u03c8\u03b5\u03b9\u03c2, confidence intervals, TTFT \u03ae percentiles inter-token latency, \u03b5\u03bd\u03ce \u03b7 \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03b1\u03c6\u03ae \u03b4\u03b5\u03bd \u03b4\u03b5\u03c3\u03bc\u03b5\u03cd\u03b5\u03b9 \u03c4\u03bf\u03bd KV dtype \u03c4\u03bf\u03c5 dense control. \u0395\u03c0\u03bf\u03bc\u03ad\u03bd\u03c9\u03c2 \u03c4\u03bf 0,9821 \u03b4\u03b5\u03bd \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c0\u03b1\u03c1\u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b5\u03af \u03c9\u03c2 \u03b3\u03b5\u03bd\u03b9\u03ba\u03ae \u03c3\u03cd\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7 BF16 \u03ae FP8.<\/p>\n<p>\u03a3\u03b5 \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03cc warm-prefix coordinate, \u03bc\u03b5 blocks 832 tokens \u03ba\u03b1\u03b9 N=32, \u03b1\u03bd\u03b1\u03c6\u03ad\u03c1\u03bf\u03bd\u03c4\u03b1\u03b9 32\/32 \u03b1\u03ba\u03c1\u03b9\u03b2\u03ae passkey outputs, \u03bc\u03b7\u03b4\u03b5\u03bd\u03b9\u03ba\u03ac fallbacks, 4,342\u00d7 physical KV compression \u03ba\u03b1\u03b9 throughput ratio 1,021\u00d7 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 dense FP8 automatic prefix caching. \u0391\u03c5\u03c4\u03ac \u03c4\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03b9\u03ce\u03bd\u03bf\u03c5\u03bd prefix mechanics \u03c3\u03c4\u03b7 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7 \u03c1\u03cd\u03b8\u03bc\u03b9\u03c3\u03b7\u00b7 \u03b4\u03b5\u03bd \u03c3\u03c5\u03b3\u03c7\u03c9\u03bd\u03b5\u03cd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c4\u03bf quality profile \u03ae \u03c4\u03bf direct canary. \u0393\u03b9\u03b1 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ae \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ae \u03b5\u03c0\u03b1\u03bd\u03b1\u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7\u03c2 \u03ba\u03bf\u03b9\u03bd\u03ce\u03bd \u03c4\u03bc\u03b7\u03bc\u03ac\u03c4\u03c9\u03bd context, \u03b4\u03b5\u03af\u03c4\u03b5 \u03ba\u03b1\u03b9 \u03c4\u03bf <a href=\"https:\/\/twodots.gr\/kvboost-taxytera-llm-koino-periexomeno-oxi-stin-arxi\/\">KVBoost<\/a>.<\/p>\n<h2 id=\"epicheirimatiki-axia\">\u03a4\u03b9 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03b3\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03ac \u03c3\u03c5\u03c3\u03c4\u03ae\u03bc\u03b1\u03c4\u03b1 AI<\/h2>\n<p>\u0397 \u03c0\u03b9\u03b8\u03b1\u03bd\u03ae \u03b1\u03be\u03af\u03b1 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03ba\u03ac\u03b8\u03b5 \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7 \u03b8\u03b1 \u03b3\u03af\u03bd\u03b5\u03b9 \u03b1\u03c5\u03c4\u03bf\u03bc\u03ac\u03c4\u03c9\u03c2 \u03b3\u03c1\u03b7\u03b3\u03bf\u03c1\u03cc\u03c4\u03b5\u03c1\u03b7. \u039c\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03b7 KV cache \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03c0\u03b9\u03c4\u03c1\u03ad\u03c8\u03b5\u03b9 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b1 resident contexts, \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b1 batches \u03ae \u03bb\u03b9\u03b3\u03cc\u03c4\u03b5\u03c1\u03b1 admission blocks \u03cc\u03c4\u03b1\u03bd \u03c4\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03b2\u03c1\u03af\u03c3\u03ba\u03b5\u03c4\u03b1\u03b9 \u03c5\u03c0\u03cc \u03c5\u03c8\u03b7\u03bb\u03cc \u03c6\u03bf\u03c1\u03c4\u03af\u03bf. \u0397 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03bc\u03b5\u03c4\u03ac \u03c4\u03bf prefill \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03c4\u03b1\u03c5\u03c4\u03cc\u03c7\u03c1\u03bf\u03bd\u03b1 \u03bd\u03b1 \u03c0\u03c1\u03bf\u03c3\u03b8\u03ad\u03c3\u03b5\u03b9 conversion cost, \u03ac\u03c1\u03b1 \u03c4\u03bf service-observed \u03cc\u03c6\u03b5\u03bb\u03bf\u03c2 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bc\u03b5\u03c4\u03c1\u03b7\u03b8\u03b5\u03af \u03c3\u03c4\u03b7\u03bd \u03bf\u03c5\u03c1\u03ac \u03ba\u03b1\u03b9 \u03c3\u03c4\u03bf\u03bd scheduler.<\/p>\n<p>\u0393\u03b9\u03b1 e-commerce, customer support \u03ae knowledge workflows, \u03c4\u03bf \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03bf \u03b5\u03c1\u03ce\u03c4\u03b7\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03bd \u03c4\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03b5\u03af \u03c4\u03b9\u03c2 \u03c0\u03b1\u03bb\u03b9\u03ad\u03c2 \u03bf\u03b4\u03b7\u03b3\u03af\u03b5\u03c2 \u03ba\u03b1\u03b9 \u03c0\u03b7\u03b3\u03ad\u03c2 \u03c0\u03bf\u03c5 \u03b5\u03c0\u03b7\u03c1\u03b5\u03ac\u03b6\u03bf\u03c5\u03bd \u03c4\u03b7\u03bd \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7, \u03bc\u03ad\u03c3\u03b1 \u03c3\u03c4\u03bf latency \u03ba\u03b1\u03b9 \u03c4\u03bf \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03c0\u03bf\u03c5 \u03b1\u03bd\u03c4\u03ad\u03c7\u03b5\u03b9 \u03b7 \u03c5\u03c0\u03b7\u03c1\u03b5\u03c3\u03af\u03b1. \u0397 \u03bc\u03bd\u03ae\u03bc\u03b7 GPU \u03b5\u03af\u03bd\u03b1\u03b9 \u03bc\u03cc\u03bd\u03bf \u03ad\u03bd\u03b1 \u03c4\u03bc\u03ae\u03bc\u03b1 \u03c4\u03bf\u03c5 unit economics. \u03a4\u03bf <a href=\"https:\/\/twodots.gr\/ai-roi-axia-ergasias\/\">AI ROI \u03c3\u03b5 \u03b5\u03c0\u03af\u03c0\u03b5\u03b4\u03bf \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1\u03c2<\/a> \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bd\u03b1 \u03c3\u03c5\u03bd\u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03af\u03b6\u03b5\u03b9 \u03b1\u03c0\u03bf\u03c4\u03c5\u03c7\u03af\u03b5\u03c2, retries, human review \u03ba\u03b1\u03b9 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc outcome.<\/p>\n<p>\u0397 \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ae \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf \u03c3\u03c7\u03b5\u03c4\u03b9\u03ba\u03ae \u03b3\u03b9\u03b1 \u03bf\u03bc\u03ac\u03b4\u03b5\u03c2 \u03c0\u03bf\u03c5 \u03b5\u03bb\u03ad\u03b3\u03c7\u03bf\u03c5\u03bd \u03c4\u03bf inference stack \u03ae \u03c3\u03c5\u03bd\u03b5\u03c1\u03b3\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c0\u03ac\u03c1\u03bf\u03c7\u03bf \u03c0\u03bf\u03c5 \u03b5\u03ba\u03b8\u03ad\u03c4\u03b5\u03b9 \u03b1\u03c1\u03ba\u03b5\u03c4\u03ae telemetry. \u039c\u03b9\u03b1 \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ae \u03c0\u03bf\u03c5 \u03ba\u03b1\u03bb\u03b5\u03af \u03b1\u03c0\u03bb\u03ce\u03c2 \u03ad\u03bd\u03b1 \u03ba\u03bb\u03b5\u03b9\u03c3\u03c4\u03cc API \u03b4\u03b5\u03bd \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b8\u03b5\u03c9\u03c1\u03ae\u03c3\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03ad\u03c7\u03b5\u03b9 \u03c4\u03bf \u03af\u03b4\u03b9\u03bf memory profile. \u0393\u03b9\u03b1 \u03c4\u03bf \u03c0\u03ce\u03c2 runtime, scheduler \u03ba\u03b1\u03b9 \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd \u03c3\u03c5\u03bd\u03b4\u03ad\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c3\u03b5 \u03bc\u03af\u03b1 \u03b5\u03bd\u03b9\u03b1\u03af\u03b1 \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae, \u03b2\u03bf\u03b7\u03b8\u03ac \u03b7 \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03b3\u03b9\u03b1 \u03c4\u03bf <a href=\"https:\/\/twodots.gr\/full-stack-ai-integrated-digital-products\/\">full-stack AI<\/a>.<\/p>\n<h2 id=\"elegxos-paragogis\">\u03a0\u03ce\u03c2 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af\u03c4\u03b1\u03b9 \u03bc\u03b9\u03b1 \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ae KV compression \u03c3\u03c4\u03b7\u03bd \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae<\/h2>\n<p>\u03a4\u03bf \u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c2 pilot \u03b4\u03b5\u03bd \u03be\u03b5\u03ba\u03b9\u03bd\u03ac \u03b1\u03c0\u03cc \u03ad\u03bd\u03b1\u03bd headline ratio. \u039e\u03b5\u03ba\u03b9\u03bd\u03ac \u03b1\u03c0\u03cc \u03b1\u03bd\u03c4\u03b9\u03c0\u03c1\u03bf\u03c3\u03c9\u03c0\u03b5\u03c5\u03c4\u03b9\u03ba\u03ac traces \u03ba\u03b1\u03b9 \u03ad\u03bd\u03b1 dense control \u03bc\u03b5 \u03ba\u03bb\u03b5\u03b9\u03b4\u03c9\u03bc\u03ad\u03bd\u03b5\u03c2 \u03b5\u03ba\u03b4\u03cc\u03c3\u03b5\u03b9\u03c2 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf\u03c5, tokenizer, runtime, kernels \u03ba\u03b1\u03b9 scheduler. \u039c\u03b5\u03c4\u03ac \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03bc\u03cc\u03bd\u03bf \u03b7 \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae KV \u03ba\u03b1\u03b9 \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 \u03c4\u03b9 \u03c3\u03c5\u03bc\u03b2\u03b1\u03af\u03bd\u03b5\u03b9 \u03c3\u03b5 \u03bc\u03bd\u03ae\u03bc\u03b7, \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03ba\u03b1\u03b9 latency.<\/p>\n<div class=\"td-step-list\">\n<p class=\"td-step-list-title\">\u0391\u03c0\u03cc \u03c4\u03bf paper \u03c3\u03b5 \u03b5\u03bb\u03b5\u03b3\u03c7\u03cc\u03bc\u03b5\u03bd\u03bf production pilot<\/p>\n<ol>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 1<\/span><strong>\u039f\u03c1\u03af\u03c3\u03c4\u03b5 \u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc long-context workload<\/strong>\n<p>\u03a3\u03c5\u03bb\u03bb\u03ad\u03be\u03c4\u03b5 \u03ba\u03b1\u03c4\u03b1\u03bd\u03bf\u03bc\u03ad\u03c2 prompt length, output length, concurrency, prefix reuse \u03ba\u03b1\u03b9 \u03c4\u03b1 \u03c3\u03b7\u03bc\u03b5\u03af\u03b1 \u03c4\u03bf\u03c5 \u03c0\u03b1\u03bb\u03b9\u03bf\u03cd context \u03c0\u03bf\u03c5 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03bf\u03c5\u03bd \u03b1\u03bd\u03b1\u03ba\u03c4\u03ae\u03c3\u03b9\u03bc\u03b1.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 2<\/span><strong>\u039a\u03bb\u03b5\u03b9\u03b4\u03ce\u03c3\u03c4\u03b5 \u03ad\u03bd\u03b1 dense control<\/strong>\n<p>\u039a\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c8\u03c4\u03b5 model \u03ba\u03b1\u03b9 tokenizer revision, KV dtype, engine commit, CUDA stack, kernels, scheduler policy \u03ba\u03b1\u03b9 \u03b1\u03ba\u03c1\u03b9\u03b2\u03ad\u03c2 hardware.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 3<\/span><strong>\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 \u03c6\u03c5\u03c3\u03b9\u03ba\u03ae \u03ba\u03b1\u03b9 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ae \u03bc\u03bd\u03ae\u03bc\u03b7 \u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ac<\/strong>\n<p>\u03a0\u03b1\u03c1\u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03ae\u03c3\u03c4\u03b5 active KV bytes, metadata, scratch, allocator reserve, weights \u03ba\u03b1\u03b9 peak HBM \u03ce\u03c3\u03c4\u03b5 \u03bf codec ratio \u03bd\u03b1 \u03bc\u03b7 \u03c3\u03c5\u03b3\u03c7\u03ad\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c4\u03b7 \u03bc\u03bd\u03ae\u03bc\u03b7 \u03cc\u03bb\u03b7\u03c2 \u03c4\u03b7\u03c2 \u03b4\u03b9\u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1\u03c2.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 4<\/span><strong>\u0394\u03bf\u03ba\u03b9\u03bc\u03ac\u03c3\u03c4\u03b5 \u03c4\u03b9\u03c2 \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03b5\u03c2 \u03b1\u03bd\u03b1\u03ba\u03bb\u03ae\u03c3\u03b5\u03b9\u03c2<\/strong>\n<p>\u03a7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03c4\u03b5 \u03ad\u03b3\u03b3\u03c1\u03b1\u03c6\u03b1 \u03ba\u03b1\u03b9 \u03c3\u03c5\u03bd\u03bf\u03bc\u03b9\u03bb\u03af\u03b5\u03c2 \u03cc\u03c0\u03bf\u03c5 \u03bc\u03b9\u03b1 \u03c0\u03b1\u03bb\u03b9\u03ac \u03bf\u03b4\u03b7\u03b3\u03af\u03b1 \u03ae \u03c0\u03b7\u03b3\u03ae \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03bf\u03c1\u03b8\u03ae \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03c4\u03b5 \u03b6\u03b5\u03c5\u03b3\u03b1\u03c1\u03c9\u03bc\u03ad\u03bd\u03b1 outputs \u03bc\u03b5 \u03c4\u03bf control.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 5<\/span><strong>\u03a4\u03c1\u03ad\u03be\u03c4\u03b5 offered-load \u03ba\u03b1\u03bc\u03c0\u03cd\u03bb\u03b5\u03c2<\/strong>\n<p>\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 admission rate, queue time, TTFT, inter-token latency percentiles, throughput \u03ba\u03b1\u03b9 fallbacks \u03c3\u03b5 \u03c7\u03b1\u03bc\u03b7\u03bb\u03cc \u03ba\u03b1\u03b9 \u03c5\u03c8\u03b7\u03bb\u03cc \u03c6\u03bf\u03c1\u03c4\u03af\u03bf.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 6<\/span><strong>\u0395\u03bb\u03ad\u03b3\u03be\u03c4\u03b5 transitions \u03ba\u03b1\u03b9 prefix ownership<\/strong>\n<p>\u0394\u03bf\u03ba\u03b9\u03bc\u03ac\u03c3\u03c4\u03b5 retiering, copy-on-write, cancellation, cache eviction \u03ba\u03b1\u03b9 concurrent readers \u03b3\u03b9\u03b1 leaks, stale references \u03ba\u03b1\u03b9 \u03c3\u03b9\u03c9\u03c0\u03b7\u03bb\u03ad\u03c2 \u03b5\u03c0\u03b9\u03c3\u03c4\u03c1\u03bf\u03c6\u03ad\u03c2 \u03c3\u03b5 dense path.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 7<\/span><strong>\u039f\u03c1\u03af\u03c3\u03c4\u03b5 rollback \u03bc\u03b5 \u03b2\u03ac\u03c3\u03b7 SLO \u03ba\u03b1\u03b9 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1<\/strong>\n<p>\u039c\u03b7\u03bd \u03b5\u03bd\u03b5\u03c1\u03b3\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03b5\u03c4\u03b5 \u03ba\u03b1\u03b8\u03bf\u03bb\u03b9\u03ba\u03ac \u03c4\u03b7 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03b1\u03bd \u03bf\u03b9 \u03bf\u03c5\u03c1\u03ad\u03c2, \u03c4\u03b1 tail latencies \u03ae \u03c4\u03b1 task-specific evals \u03be\u03b5\u03c0\u03b5\u03c1\u03bd\u03bf\u03cd\u03bd \u03c4\u03b1 \u03c0\u03c1\u03bf\u03ba\u03b1\u03b8\u03bf\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03b1 \u03cc\u03c1\u03b9\u03b1.<\/p>\n<\/li>\n<\/ol>\n<\/div>\n<p>\u0391\u03c5\u03c4\u03cc\u03c2 \u03bf \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf\u03c2 \u03c3\u03c5\u03bd\u03b4\u03ad\u03b5\u03b9 \u03c4\u03b7\u03bd \u03c5\u03c0\u03bf\u03b4\u03bf\u03bc\u03ae \u03bc\u03b5 \u03c4\u03b7 \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03b9\u03c6\u03bf\u03c1\u03ac \u03c4\u03bf\u03c5 \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd\u03c4\u03bf\u03c2. \u0391\u03bd \u03bc\u03b9\u03b1 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03b8\u03ad\u03bb\u03b5\u03b9 \u03bd\u03b1 \u03c0\u03b5\u03c1\u03ac\u03c3\u03b5\u03b9 \u03b1\u03c0\u03cc demo \u03c3\u03b5 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03af\u03b1 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae\u03c2, \u03c4\u03b1 \u03af\u03b4\u03b9\u03b1 validation gates \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b5\u03bd\u03c3\u03c9\u03bc\u03b1\u03c4\u03c9\u03b8\u03bf\u03cd\u03bd \u03c3\u03c4\u03bf \u03b5\u03c5\u03c1\u03cd\u03c4\u03b5\u03c1\u03bf \u03c3\u03c7\u03ad\u03b4\u03b9\u03bf deployment, \u03cc\u03c0\u03c9\u03c2 \u03c0\u03b5\u03c1\u03b9\u03b3\u03c1\u03ac\u03c6\u03b5\u03c4\u03b1\u03b9 \u03c3\u03c4\u03b1 <a href=\"https:\/\/twodots.gr\/ai-agent-demo-paragogi-pente-teixi-epicheirisis\/\">\u03c0\u03ad\u03bd\u03c4\u03b5 \u03b5\u03bc\u03c0\u03cc\u03b4\u03b9\u03b1 \u03b1\u03c0\u03cc AI agent demo \u03c3\u03b5 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae<\/a>.<\/p>\n<h2 id=\"oria\">\u039f\u03b9 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03bf\u03af \u03c0\u03bf\u03c5 \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c7\u03b1\u03b8\u03bf\u03cd\u03bd<\/h2>\n<p>\u0397 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 \u03b1\u03c6\u03bf\u03c1\u03ac \u03c4\u03bf Qwen3.6-27B \u03c3\u03b5 \u03bc\u03af\u03b1 NVIDIA RTX PRO 6000 Blackwell 96 GB. \u0394\u03b5\u03bd \u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03b5\u03b9 tensor parallelism, multi-GPU serving, vision paths \u03ae \u03ac\u03bb\u03bb\u03b5\u03c2 \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ad\u03c2. \u0395\u03c0\u03b5\u03b9\u03b4\u03ae \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03b5\u03af\u03bd\u03b1\u03b9 \u03c5\u03b2\u03c1\u03b9\u03b4\u03b9\u03ba\u03cc, \u03bf\u03b9 \u03bb\u03cc\u03b3\u03bf\u03b9 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7\u03c2 \u03b1\u03c6\u03bf\u03c1\u03bf\u03cd\u03bd \u03c4\u03bf attention KV \u03c4\u03c9\u03bd 16 full-attention layers \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03b7 \u03c4\u03b7 \u03bc\u03bd\u03ae\u03bc\u03b7 \u03c4\u03bf\u03c5 process \u03ae \u03c4\u03bf Gated DeltaNet state.<\/p>\n<p>\u03a4\u03b1 profiles \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2, direct decode, deployment accounting \u03ba\u03b1\u03b9 prefix reuse \u03b5\u03af\u03bd\u03b1\u03b9 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b4\u03b9\u03b1\u03bc\u03bf\u03c1\u03c6\u03ce\u03c3\u03b5\u03b9\u03c2. \u0394\u03b5\u03bd \u03ad\u03c7\u03bf\u03c5\u03bd \u03bc\u03b5\u03c4\u03c1\u03b7\u03b8\u03b5\u03af conversion throughput, peak scratch, queue lag \u03ae \u03c4\u03bf \u03c0\u03bf\u03c3\u03bf\u03c3\u03c4\u03cc \u03bc\u03b5\u03c4\u03b1\u03c4\u03c1\u03bf\u03c0\u03ae\u03c2 \u03c0\u03bf\u03c5 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03ba\u03c1\u03c5\u03c6\u03c4\u03b5\u03af \u03c0\u03af\u03c3\u03c9 \u03b1\u03c0\u03cc \u03ac\u03bb\u03bb\u03bf serving work. \u03a4\u03bf \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03bf arithmetic profile \u03c0\u03b1\u03c1\u03bf\u03c5\u03c3\u03af\u03b1\u03c3\u03b5 \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 \u03c5\u03c0\u03bf\u03c7\u03ce\u03c1\u03b7\u03c3\u03b7 \u03c3\u03b5 \u03c4\u03c1\u03af\u03b1 \u03b1\u03c0\u03cc \u03c4\u03b1 \u03bf\u03ba\u03c4\u03ce 16K RULER tasks, \u03ac\u03c1\u03b1 \u03c4\u03bf speed result \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b5\u03bd\u03c9\u03b8\u03b5\u03af \u03bc\u03b5 \u03c4\u03bf\u03bd \u03b5\u03c5\u03c1\u03cd\u03c4\u03b5\u03c1\u03bf \u03b9\u03c3\u03c7\u03c5\u03c1\u03b9\u03c3\u03bc\u03cc \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2.<\/p>\n<p>\u03a4\u03ad\u03bb\u03bf\u03c2, \u03b7 \u03c5\u03bb\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 proprietary \u03ba\u03b1\u03b9 \u03b4\u03b5\u03bd \u03b4\u03b9\u03b1\u03c4\u03af\u03b8\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b1\u03b6\u03af \u03bc\u03b5 \u03c4\u03b7\u03bd \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1. \u039f\u03b9 \u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03b5\u03af\u03c2 \u03b1\u03bd\u03b1\u03c6\u03ad\u03c1\u03bf\u03c5\u03bd \u03cc\u03c4\u03b9 detailed manifests, telemetry \u03ba\u03b1\u03b9 raw logs \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9 \u03b5\u03c3\u03c9\u03c4\u03b5\u03c1\u03b9\u03ba\u03ac, \u03b1\u03bb\u03bb\u03ac \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03bf\u03cd\u03bd \u03b4\u03b7\u03bc\u03cc\u03c3\u03b9\u03bf artifact. \u0391\u03c5\u03c4\u03cc \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03af\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03b7 \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae \u03ba\u03b1\u03b9 \u03ba\u03ac\u03bd\u03b5\u03b9 \u03b1\u03ba\u03cc\u03bc\u03b7 \u03c0\u03b9\u03bf \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03ae \u03c4\u03b7\u03bd profile-specific \u03b1\u03bd\u03ac\u03b3\u03bd\u03c9\u03c3\u03b7.<\/p>\n<h2 id=\"ousia\">\u0397 \u03bf\u03c5\u03c3\u03af\u03b1: \u03b4\u03b9\u03b1\u03c7\u03b5\u03af\u03c1\u03b9\u03c3\u03b7 \u03c0\u03b9\u03c3\u03c4\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2, \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03bf \u03bb\u03b9\u03b3\u03cc\u03c4\u03b5\u03c1\u03b1 bits<\/h2>\n<p>\u03a4\u03bf \u03c0\u03b9\u03bf \u03b5\u03bd\u03b4\u03b9\u03b1\u03c6\u03ad\u03c1\u03bf\u03bd \u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af\u03bf \u03c4\u03bf\u03c5 Minima-KV \u03b5\u03af\u03bd\u03b1\u03b9 \u03b7 \u03bc\u03b5\u03c4\u03ac\u03b2\u03b1\u03c3\u03b7 \u03b1\u03c0\u03cc \u03bc\u03af\u03b1 \u03bf\u03bc\u03bf\u03b9\u03cc\u03bc\u03bf\u03c1\u03c6\u03b7 cache \u03c3\u03b5 \u03b9\u03b5\u03c1\u03b1\u03c1\u03c7\u03af\u03b1 \u03c0\u03b9\u03c3\u03c4\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2. \u039f\u03b9 \u03c0\u03c1\u03cc\u03c3\u03c6\u03b1\u03c4\u03b5\u03c2 \u03ba\u03b1\u03b9 \u03c0\u03c1\u03bf\u03c3\u03c4\u03b1\u03c4\u03b5\u03c5\u03bc\u03ad\u03bd\u03b5\u03c2 \u03c3\u03b5\u03bb\u03af\u03b4\u03b5\u03c2 \u03c0\u03b1\u03af\u03c1\u03bd\u03bf\u03c5\u03bd \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b1 bits, \u03bf\u03b9 \u03c0\u03b1\u03bb\u03b9\u03cc\u03c4\u03b5\u03c1\u03b5\u03c2 \u03bb\u03b9\u03b3\u03cc\u03c4\u03b5\u03c1\u03b1, \u03b1\u03bb\u03bb\u03ac \u03cc\u03bb\u03b5\u03c2 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03bf\u03c5\u03bd \u03b4\u03b9\u03b1\u03b8\u03ad\u03c3\u03b9\u03bc\u03b5\u03c2 \u03c3\u03c4\u03bf \u03b5\u03bd\u03b5\u03c1\u03b3\u03cc request. \u039f codec, \u03c4\u03b1 mixed-format kernels, \u03c4\u03bf global softmax merge, \u03c4\u03bf ownership \u03ba\u03b1\u03b9 \u03c4\u03b1 route counters \u03b5\u03af\u03bd\u03b1\u03b9 \u03bc\u03ad\u03c1\u03b7 \u03c4\u03bf\u03c5 \u03af\u03b4\u03b9\u03bf\u03c5 \u03c3\u03c5\u03c3\u03c4\u03ae\u03bc\u03b1\u03c4\u03bf\u03c2.<\/p>\n<p>\u03a4\u03b1 \u03b4\u03b9\u03b1\u03b8\u03ad\u03c3\u03b9\u03bc\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03b4\u03b5\u03af\u03c7\u03bd\u03bf\u03c5\u03bd \u03bc\u03b9\u03b1 \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ac \u03c1\u03b5\u03b1\u03bb\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae \u03ba\u03b1\u03c4\u03b5\u03cd\u03b8\u03c5\u03bd\u03c3\u03b7: \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03ae \u03bc\u03b5\u03af\u03c9\u03c3\u03b7 attention KV, \u03bc\u03b9\u03ba\u03c1\u03ad\u03c2 \u03b1\u03c0\u03bf\u03ba\u03bb\u03af\u03c3\u03b5\u03b9\u03c2 \u03c3\u03c4\u03b1 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b1 quality tests \u03ba\u03b1\u03b9 direct decode \u03c7\u03c9\u03c1\u03af\u03c2 dense shadow. \u0394\u03b5\u03bd \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03bf\u03cd\u03bd \u03cc\u03bc\u03c9\u03c2 \u03b3\u03b5\u03bd\u03b9\u03ba\u03ae \u03b5\u03b3\u03b3\u03cd\u03b7\u03c3\u03b7 \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf, GPU \u03ae workload. \u0397 \u03b5\u03c0\u03cc\u03bc\u03b5\u03bd\u03b7 \u03c0\u03b5\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae \u03b1\u03c0\u03cc\u03b4\u03b5\u03b9\u03be\u03b7 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b4\u03b7\u03bc\u03cc\u03c3\u03b9\u03b1 \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae, \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b1 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03ba\u03b1\u03b9 \u03b5\u03c0\u03b1\u03bd\u03b1\u03bb\u03b1\u03bc\u03b2\u03b1\u03bd\u03cc\u03bc\u03b5\u03bd\u03b5\u03c2 offered-load \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03bc\u03b5 tail latency \u03ba\u03b1\u03b9 conversion overhead.<\/p>\n<p>\u0393\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7, \u03c4\u03bf \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03cc \u03c3\u03c5\u03bc\u03c0\u03ad\u03c1\u03b1\u03c3\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03c4\u03bf long-context AI \u03b4\u03b5\u03bd \u03ba\u03bb\u03b9\u03bc\u03b1\u03ba\u03ce\u03bd\u03b5\u03c4\u03b1\u03b9 \u03bc\u03cc\u03bd\u03bf \u03bc\u03b5 \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf context window. \u03a7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b5\u03bb\u03b5\u03b3\u03c7\u03cc\u03bc\u03b5\u03bd\u03b7 \u03b4\u03b9\u03b1\u03c7\u03b5\u03af\u03c1\u03b9\u03c3\u03b7 state, \u03c3\u03b1\u03c6\u03b5\u03af\u03c2 \u03b1\u03bd\u03c4\u03b1\u03bb\u03bb\u03b1\u03b3\u03ad\u03c2 \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03b5 \u03bc\u03bd\u03ae\u03bc\u03b7 \u03ba\u03b1\u03b9 \u03c0\u03b9\u03c3\u03c4\u03cc\u03c4\u03b7\u03c4\u03b1 \u03ba\u03b1\u03b9 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 \u03c0\u03ac\u03bd\u03c9 \u03c3\u03c4\u03b7 \u03b4\u03b9\u03ba\u03ae \u03c4\u03b7\u03c2 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1.<\/p>\n<section class=\"td-service-cta\">\n<div class=\"td-service-cta-content\">\n<p class=\"td-service-cta-kicker\">AI \u03c5\u03c0\u03bf\u03b4\u03bf\u03bc\u03ae \u03bc\u03b5 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b9\u03bc\u03b1 \u03cc\u03c1\u03b9\u03b1<\/p>\n<p class=\"td-service-cta-title\">\u03a3\u03c7\u03b5\u03b4\u03b9\u03ac\u03c3\u03c4\u03b5 long-context workflows \u03c0\u03bf\u03c5 \u03b1\u03bd\u03c4\u03ad\u03c7\u03bf\u03c5\u03bd \u03c3\u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03c6\u03bf\u03c1\u03c4\u03af\u03bf<\/p>\n<p>\u0397 TWO DOTS \u03c3\u03c5\u03bd\u03b4\u03ad\u03b5\u03b9 AI \u03b1\u03c5\u03c4\u03bf\u03bc\u03b1\u03c4\u03b9\u03c3\u03bc\u03bf\u03cd\u03c2 \u03bc\u03b5 workload-specific evals, telemetry, validation gates \u03ba\u03b1\u03b9 human review, \u03ce\u03c3\u03c4\u03b5 \u03bf\u03b9 \u03b1\u03c0\u03bf\u03c6\u03ac\u03c3\u03b5\u03b9\u03c2 \u03b3\u03b9\u03b1 context, \u03bc\u03bd\u03ae\u03bc\u03b7 \u03ba\u03b1\u03b9 \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03bd\u03b1 \u03b2\u03b1\u03c3\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c3\u03c4\u03b7 \u03b4\u03b9\u03ba\u03ae \u03c3\u03b1\u03c2 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03af\u03b1.<\/p>\n<div class=\"td-service-cta-actions\"><a class=\"td-service-cta-button\" href=\"https:\/\/twodots.gr\/aftomatismoi-epicheiriseon-ai\/\">\u0394\u03b5\u03af\u03c4\u03b5 \u0391\u03c5\u03c4\u03bf\u03bc\u03b1\u03c4\u03b9\u03c3\u03bc\u03bf\u03cd\u03c2 \u0395\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03ae\u03c3\u03b5\u03c9\u03bd &amp; AI<\/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\">\u0427\u0435\u0441\u0442\u043e \u0437\u0430\u0434\u0430\u0432\u0430\u043d\u0438 \u0432\u044a\u043f\u0440\u043e\u0441\u0438<\/p>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03b9 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b7 KV cache;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0395\u03af\u03bd\u03b1\u03b9 \u03b7 \u03b1\u03c0\u03bf\u03b8\u03b7\u03ba\u03b5\u03c5\u03bc\u03ad\u03bd\u03b7 \u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 keys \u03ba\u03b1\u03b9 values \u03b1\u03c0\u03cc \u03c0\u03c1\u03bf\u03b7\u03b3\u03bf\u03cd\u03bc\u03b5\u03bd\u03b1 tokens, \u03ce\u03c3\u03c4\u03b5 \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03bd\u03b1 \u03bc\u03b7\u03bd \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03af\u03b6\u03b5\u03b9 \u03be\u03b1\u03bd\u03ac \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03bf \u03c4\u03bf \u03b9\u03c3\u03c4\u03bf\u03c1\u03b9\u03ba\u03cc \u03c3\u03b5 \u03ba\u03ac\u03b8\u03b5 \u03b2\u03ae\u03bc\u03b1 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae\u03c2.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03b9 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03cc \u03ba\u03ac\u03bd\u03b5\u03b9 \u03c4\u03bf Minima-KV;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a7\u03c9\u03c1\u03af\u03b6\u03b5\u03b9 \u03c4\u03b9\u03c2 \u03c3\u03b5\u03bb\u03af\u03b4\u03b5\u03c2 \u03c3\u03b5 Recent, Anchor \u03ba\u03b1\u03b9 Stale. \u039f\u03b9 Recent \u03ba\u03b1\u03b9 Anchor \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03bf\u03c5\u03bd \u03c3\u03b5 FP8, \u03b5\u03bd\u03ce \u03bf\u03b9 Stale \u03c3\u03c5\u03bc\u03c0\u03b9\u03ad\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c3\u03b5 TQ3 \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03b4\u03b9\u03b1\u03b3\u03c1\u03ac\u03c6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf \u03b5\u03bd\u03b5\u03c1\u03b3\u03cc request.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03cc\u03c3\u03b7 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03b1\u03bd\u03b1\u03c6\u03ad\u03c1\u03b5\u03b9 \u03b7 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a4\u03bf deployment aggregate \u03c4\u03c9\u03bd 18,3 KiB \u03b1\u03bd\u03ac live token \u03b1\u03bd\u03c4\u03b9\u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af \u03c3\u03b5 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 3,50 \u03c6\u03bf\u03c1\u03ad\u03c2 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf attention KV \u03b1\u03c0\u03cc BF16 \u03ba\u03b1\u03b9 1,75 \u03c6\u03bf\u03c1\u03ad\u03c2 \u03b1\u03c0\u03cc FP8. \u039e\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03cc direct canary \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b5 3,625 \u03c6\u03bf\u03c1\u03ad\u03c2 active-KV compression.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0394\u03b9\u03b1\u03c4\u03b7\u03c1\u03b5\u03af\u03c4\u03b1\u03b9 \u03c0\u03bb\u03ae\u03c1\u03c9\u03c2 \u03b7 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u038c\u03c7\u03b9 \u03c9\u03c2 \u03b3\u03b5\u03bd\u03b9\u03ba\u03cc\u03c2 \u03b9\u03c3\u03c7\u03c5\u03c1\u03b9\u03c3\u03bc\u03cc\u03c2. \u03a5\u03c0\u03ae\u03c1\u03be\u03b5 \u03b9\u03c3\u03bf\u03c0\u03b1\u03bb\u03af\u03b1 \u03c3\u03c4\u03bf 16K RULER profile, \u03bc\u03b9\u03ba\u03c1\u03ae \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03c3\u03b7 \u03c3\u03c4\u03b1 8K \u03ba\u03b1\u03b9 \u03c5\u03c0\u03bf\u03c7\u03c9\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03c3\u03c4\u03b1 4K \u03ba\u03b1\u03b9 \u03c3\u03c4\u03bf LongBench v2. \u0394\u03b5\u03bd \u03c0\u03b1\u03c1\u03ad\u03c7\u03bf\u03bd\u03c4\u03b1\u03b9 confidence intervals \u03b3\u03b9\u03b1 \u03cc\u03bb\u03b5\u03c2 \u03c4\u03b9\u03c2 \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03c3\u03b5\u03b9\u03c2.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0393\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03c4\u03bf inference \u03b3\u03c1\u03b7\u03b3\u03bf\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0394\u03b5\u03bd \u03ad\u03c7\u03b5\u03b9 \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03c7\u03b8\u03b5\u03af \u03ba\u03b1\u03b8\u03bf\u03bb\u03b9\u03ba\u03ae \u03b5\u03c0\u03b9\u03c4\u03ac\u03c7\u03c5\u03bd\u03c3\u03b7. \u03a3\u03c4\u03bf \u03bc\u03bf\u03bd\u03b1\u03b4\u03b9\u03ba\u03cc direct pair \u03c4\u03bf throughput ratio \u03ae\u03c4\u03b1\u03bd 0,9821 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 \u03c4\u03bf\u03c5 control. \u0397 \u03c0\u03b9\u03b8\u03b1\u03bd\u03ae \u03b1\u03be\u03af\u03b1 \u03b2\u03c1\u03af\u03c3\u03ba\u03b5\u03c4\u03b1\u03b9 \u03ba\u03c5\u03c1\u03af\u03c9\u03c2 \u03c3\u03c4\u03b7 \u03bc\u03b5\u03b9\u03c9\u03bc\u03ad\u03bd\u03b7 \u03bc\u03bd\u03ae\u03bc\u03b7 \u03ba\u03b1\u03b9 \u03c3\u03c4\u03bf admission capacity \u03c5\u03c0\u03cc \u03c6\u03bf\u03c1\u03c4\u03af\u03bf.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03bf Minima-KV \u03c0\u03b5\u03c4\u03ac \u03c0\u03b1\u03bb\u03b9\u03ac tokens;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u038c\u03c7\u03b9 \u03b3\u03b9\u03b1 \u03b5\u03bd\u03b5\u03c1\u03b3\u03ac requests. \u039a\u03c1\u03b1\u03c4\u03ac \u03c3\u03c5\u03bc\u03c0\u03b9\u03b5\u03c3\u03bc\u03ad\u03bd\u03b7 \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 \u03c4\u03c9\u03bd \u03c0\u03b1\u03bb\u03b9\u03ce\u03bd \u03bc\u03b7 \u03c0\u03c1\u03bf\u03c3\u03c4\u03b1\u03c4\u03b5\u03c5\u03bc\u03ad\u03bd\u03c9\u03bd \u03c3\u03b5\u03bb\u03af\u03b4\u03c9\u03bd \u03ce\u03c3\u03c4\u03b5 \u03bd\u03b1 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03bf\u03c5\u03bd addressable \u03b1\u03bd \u03be\u03b1\u03bd\u03b1\u03c7\u03c1\u03b5\u03b9\u03b1\u03c3\u03c4\u03bf\u03cd\u03bd.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0395\u03af\u03bd\u03b1\u03b9 \u03ad\u03c4\u03bf\u03b9\u03bc\u03bf \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03b9\u03ba\u03cc \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u038c\u03c7\u03b9 \u03c7\u03c9\u03c1\u03af\u03c2 \u03b5\u03c0\u03b9\u03c0\u03bb\u03ad\u03bf\u03bd validation. \u03a4\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03b1\u03c6\u03bf\u03c1\u03bf\u03cd\u03bd \u03ad\u03bd\u03b1 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03ba\u03b1\u03b9 \u03bc\u03af\u03b1 GPU, \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ac evidence profiles \u03ba\u03b1\u03b9 proprietary \u03c5\u03bb\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03c0\u03bf\u03c5 \u03b4\u03b5\u03bd \u03ad\u03c7\u03b5\u03b9 \u03b4\u03b7\u03bc\u03bf\u03c3\u03b9\u03b5\u03c5\u03c4\u03b5\u03af.<\/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 \u03b2\u03b1\u03c3\u03b9\u03ba\u03cc \u03bc\u03ac\u03b8\u03b7\u03bc\u03b1 \u03b3\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u039d\u03b1 \u03bc\u03b5\u03c4\u03c1\u03ac \u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ac KV bytes, \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc HBM, \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1, admission rate, queue time \u03ba\u03b1\u03b9 latency \u03c5\u03c0\u03cc \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03c6\u03bf\u03c1\u03c4\u03af\u03bf. \u039f \u03bb\u03cc\u03b3\u03bf\u03c2 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7\u03c2 \u03b4\u03b5\u03bd \u03bc\u03b5\u03c4\u03b1\u03c4\u03c1\u03ad\u03c0\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c5\u03c4\u03cc\u03bc\u03b1\u03c4\u03b1 \u03c3\u03b5 \u03af\u03b4\u03b9\u03bf \u03c0\u03bf\u03c3\u03bf\u03c3\u03c4\u03cc \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03c9\u03bd \u03c7\u03c1\u03b7\u03c3\u03c4\u03ce\u03bd.<\/p>\n<\/div>\n<\/details>\n<\/div>\n<\/section>\n<div class=\"td-source-list\">\n<p id=\"piges\" class=\"td-source-list-title\">\u0418\u0437\u0442\u043e\u0447\u043d\u0438\u0446\u0438<\/p>\n<ul>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2608.23834\" target=\"_blank\" rel=\"noopener\">Kozyrev &amp; Maiboroda \u2014 Minima-KV: Retention-Preserving KV Cache Compression with Mixed-Format Paged Attention<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2309.06180\" target=\"_blank\" rel=\"noopener\">Kwon et al. \u2014 Efficient Memory Management for Large Language Model Serving with PagedAttention<\/a><\/li>\n<li><a href=\"https:\/\/docs.vllm.ai\/en\/latest\/design\/paged_attention\/\" target=\"_blank\" rel=\"noopener\">vLLM \u2014 Paged Attention documentation<\/a><\/li>\n<\/ul>\n<\/div>","protected":false},"excerpt":{"rendered":"<p>\u03a4\u03bf Minima-KV \u03c3\u03c5\u03bc\u03c0\u03b9\u03ad\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd KV cache \u03c7\u03c9\u03c1\u03af\u03c2 \u03b4\u03b9\u03b1\u03b3\u03c1\u03b1\u03c6\u03ae \u03c0\u03b1\u03bb\u03b9\u03bf\u03cd context. \u03a4\u03b9 \u03ad\u03b4\u03b5\u03b9\u03be\u03b1\u03bd \u03bf\u03b9 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03bc\u03bd\u03ae\u03bc\u03b7\u03c2, \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2 \u03ba\u03b1\u03b9 throughput \u03c3\u03c4\u03bf long-context AI.<\/p>","protected":false},"author":1,"featured_media":105889,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_gspb_post_css":"","content-type":"","footnotes":""},"categories":[199],"tags":[20912,20911,20246,20910,20909],"class_list":["post-98250","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-techniti-noimosyni","tag-ai-performance","tag-gpu-memory","tag-kv-cache","tag-llm-infrastructure","tag-long-context-ai"],"blocksy_meta":{"styles_descriptor":{"styles":{"desktop":"","tablet":"","mobile":""},"google_fonts":[],"version":8}},"_links":{"self":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/posts\/98250","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/comments?post=98250"}],"version-history":[{"count":0,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/posts\/98250\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/media\/105889"}],"wp:attachment":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/media?parent=98250"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/categories?post=98250"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/tags?post=98250"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}