{"id":1036,"date":"2026-09-16T17:18:21","date_gmt":"2026-09-16T08:18:21","guid":{"rendered":"http:\/\/kyunam.com\/?p=1036"},"modified":"2026-09-16T18:47:19","modified_gmt":"2026-09-16T09:47:19","slug":"cpu-moe-offloadinggpu-%eb%a9%94%eb%aa%a8%eb%a6%ac%ec%9d%98-%ed%95%9c%ea%b3%84%eb%a5%bc-%eb%84%98%ec%9c%bc%eb%a9%b4-%eb%ac%b4%ec%97%87%ec%9d%b4-%eb%8b%ac%eb%9d%bc%ec%a7%88%ea%b9%8c","status":"publish","type":"post","link":"http:\/\/kyunam.com\/?p=1036","title":{"rendered":"CPU MoE Offloading:GPU \uba54\ubaa8\ub9ac\uc758 \ud55c\uacc4\ub97c \ub118\uc73c\uba74 \ubb34\uc5c7\uc774 \ub2ec\ub77c\uc9c8\uae4c"},"content":{"rendered":"\n<style>\n\n.moe-post{--moe-ink:#223044;--moe-accent:#15617a;box-sizing:border-box;max-width:1100px;margin:0 auto;color:var(--moe-ink);font-family:Arial,\"Noto Sans CJK KR\",\"Malgun Gothic\",\"Apple SD Gothic Neo\",sans-serif;font-size:17px;line-height:1.88;word-break:keep-all;overflow-wrap:break-word;text-align:left;}\n.moe-post *{box-sizing:border-box;}\n.moe-post .moe-header{padding:0 0 22px;border-bottom:1px solid #dce4ea;margin-bottom:26px;}\n.moe-post .moe-eyebrow{font-size:12px;font-weight:700;letter-spacing:.13em;color:#15617a;margin:0 0 13px;}\n.moe-post .moe-title{margin:0 0 16px;font-size:clamp(29px,4.5vw,43px);font-weight:800;line-height:1.35;color:#192c40;letter-spacing:-.035em;}\n.moe-post .moe-deck{font-size:18px;color:#526477;margin:0;line-height:1.7;}\n.moe-post .moe-prose{max-width:870px;margin:0 auto;}\n.moe-post .moe-prose p{margin:0 0 19px;}\n.moe-post .moe-prose h2{margin:48px 0 21px;font-size:27px;line-height:1.45;font-weight:750;letter-spacing:-.02em;color:#18384d;}\n.moe-post .moe-prose h3{margin:30px 0 13px;font-size:20px;line-height:1.5;font-weight:700;color:#244354;}\n.moe-post .moe-prose strong{color:#152f43;font-weight:700;}\n.moe-post .moe-figure{margin:28px 0 35px;}\n.moe-post .moe-caption{font-size:13px;line-height:1.7;color:#657589;margin:12px 2px 0;word-break:keep-all;}\n.moe-post .moe-enlarge{margin:11px 0 0;font-size:13px;line-height:1.6;}\n.moe-post .moe-enlarge summary{cursor:pointer;color:#15617a;display:list-item;list-style:disclosure-closed;}\n.moe-post .moe-enlarge[open] summary{list-style:disclosure-open;}\n.moe-post .moe-zoom-viewport{overflow:auto;max-width:100%;border:1px solid #dbe4ec;margin-top:12px;background:white;}\n.moe-post .moe-zoom-detail{width:1500px;max-width:none;}\n.moe-post .moe-zoom-intro{width:1200px;max-width:none;}\n.moe-post .moe-callout{padding:20px 24px;margin:23px 0;background:#f3f8fb;border-left:4px solid #28728a;border-radius:0 6px 6px 0;line-height:1.75;}\n.moe-post .moe-note{font-size:14px;line-height:1.8;color:#586a7c;}\n.moe-post .moe-formula{padding:16px 20px;margin:20px 0;border:1px solid #d8e3eb;background:#fafcfd;border-radius:6px;font-size:15px;line-height:1.8;}\n.moe-post code{font-family:ui-monospace,SFMono-Regular,Consolas,\"Liberation Mono\",monospace;font-size:.88em;white-space:normal;overflow-wrap:anywhere;background:#eef3f6;padding:.08em .28em;border-radius:3px;}\n.moe-post .moe-code{padding:17px 20px;background:#f3f6f8;border:1px solid #dfe7ec;border-radius:6px;overflow-x:auto;font:14px\/1.9 ui-monospace,Consolas,monospace;white-space:pre;word-break:normal;}\n.moe-post .moe-table-wrap{overflow-x:auto;max-width:100%;margin:21px 0 14px;}\n.moe-post table.moe-table{width:100%;min-width:530px;border-collapse:collapse;border-spacing:0;border:1px solid #dce5ec;background:#fff;font-size:14px;line-height:1.65;margin:0;table-layout:fixed;}\n.moe-post .moe-table th,.moe-post .moe-table td{padding:12px 13px;border:1px solid #dce5ec;vertical-align:top;text-align:left;word-break:keep-all;font-weight:400;}\n.moe-post .moe-table th{font-weight:700;background:#eef5f8;color:#244659;}\n.moe-post .moe-table tbody tr:nth-child(even){background:#fafcfd;}\n.moe-post .moe-bar-chart{padding:22px 24px;background:#fbfcfe;border:1px solid #dbe5ee;border-radius:8px;margin:22px 0;}\n.moe-post .moe-bar-row{display:grid;grid-template-columns:135px 1fr 96px;gap:12px;align-items:center;font-size:14px;margin:12px 0;line-height:1.5;}\n.moe-post .moe-bar-track{height:20px;background:#e9eff4;border-radius:3px;overflow:hidden;}\n.moe-post .moe-bar-track span{height:100%;display:block;background:#27758d;border-radius:3px;min-width:2px;}\n.moe-post .moe-bar-value{font-weight:700;text-align:right;font-variant-numeric:tabular-nums;}\n.moe-post .moe-bar-scale{margin-top:13px;font-size:12px;color:#6c7d8e;}\n.moe-post .moe-motion-switch{vertical-align:middle;width:15px;height:15px;margin:0 6px 0 0;accent-color:#1d6b83;}\n.moe-post .moe-motion-label{vertical-align:middle;font-size:12px;color:#617283;cursor:pointer;}\n.moe-post .moe-source{font-size:12px;line-height:1.8;color:#7b8894;border-top:1px solid #e1e7ec;padding-top:13px;margin:23px 0;overflow-wrap:anywhere;}\n.moe-post .moe-ending{border-top:2px solid #d7e7ee;margin-top:32px;padding-top:23px;}\n@media(max-width:640px){.moe-post{font-size:16px;line-height:1.83;}.moe-post .moe-prose h2{font-size:23px;margin-top:37px;}.moe-post .moe-prose h3{font-size:18px;}.moe-post .moe-deck{font-size:16px;}.moe-post .moe-bar-row{grid-template-columns:99px 1fr 76px;gap:8px;font-size:12px;}.moe-post .moe-bar-chart{padding:15px 12px;}.moe-post .moe-callout{padding:16px 18px;}}\n\n.moe-diagram{position:relative;box-sizing:border-box;isolation:isolate;width:100%;background:#fff;text-align:left;direction:ltr;container-type:inline-size;}\n.moe-diagram *{box-sizing:border-box;}\n.moe-overview{aspect-ratio:5 \/ 2;border:1px solid #d7e7fa;border-radius:10px;overflow:hidden;background:#fbfdff;font-family:Arial,\"Noto Sans CJK KR\",\"Malgun Gothic\",sans-serif;color:#173b69;}\n.moe-ov-el{position:absolute;margin:0;padding:0;letter-spacing:0;word-break:keep-all;}\n.moe-ov-el br{display:block;}\n.moe-ov-arrow{position:absolute;transform-origin:0 50%;pointer-events:none;z-index:2;}\n.moe-ov-shaft{display:block;position:absolute;left:0;top:0;width:calc(100% - .5cqi);height:100%;background:var(--moe-ac);overflow:hidden;border-radius:1cqi;}\n.moe-ov-shaft:after{content:\"\";position:absolute;inset:27% 0;background:repeating-linear-gradient(90deg,transparent 0 .6cqi,var(--moe-al) .6cqi 1.2cqi,transparent 1.2cqi 2.4cqi);background-size:2.4cqi 100%;animation:moe-flow-x 1.2s linear infinite;}\n.moe-ov-arrowhead{position:absolute;right:-.02cqi;top:50%;width:1.05cqi;height:.92cqi;transform:translateY(-50%);background:var(--moe-ac);clip-path:polygon(0 0,100% 50%,0 100%);filter:drop-shadow(0 0 .03cqi rgba(0,0,0,.10));}\n\n.moe-chat-dots{position:absolute;inset:0;display:flex;justify-content:center;align-items:center;gap:.44cqi;}\n.moe-chat-dots i{display:block;width:.42cqi;height:.42cqi;background:#194c8c;border-radius:50%;}\n.moe-chat-tail{position:absolute;left:.35cqi;bottom:-.48cqi;width:.8cqi;height:.8cqi;border-left:.27cqi solid #194c8c;transform:skewY(-32deg);background:#f6f9fd;}\n.moe-ov-token{animation:moe-token-light 3s ease-in-out infinite;animation-delay:var(--moe-delay);}\n.moe-ov-router{animation:moe-route-glow 6s ease-in-out infinite;}\n.moe-ov-hot,.moe-ov-gpu{animation:moe-gpu-glow 6s ease-in-out infinite;}\n.moe-ov-cold,.moe-ov-cpu{animation:moe-cpu-glow 6s ease-in-out infinite;}\n.moe-ov-kv{animation:moe-kv-glow 3s ease-in-out infinite;}\n.moe-chip-face{animation:moe-chip-light 3s ease-in-out infinite;}\n.moe-chip-1{animation-delay:.2s}.moe-chip-2{animation-delay:.4s}.moe-chip-3{animation-delay:.6s}.moe-chip-4{animation-delay:.8s}\n@keyframes moe-flow-x{to{background-position:2.4cqi 0;}}\n@keyframes moe-flow-y{to{background-position:0 2.4cqi;}}\n@keyframes moe-token-light{0%,100%{filter:brightness(1)}35%{filter:brightness(1.2)}}\n@keyframes moe-route-glow{0%,16%,50%,100%{box-shadow:none}28%,40%{box-shadow:inset 0 0 0 .14cqi #3c83dd,0 0 .45cqi #c4ddfc}}\n@keyframes moe-gpu-glow{0%,25%,70%,100%{box-shadow:none}38%,55%{box-shadow:inset 0 0 0 .14cqi #46b875}}\n@keyframes moe-cpu-glow{0%,32%,82%,100%{box-shadow:none}46%,67%{box-shadow:inset 0 0 0 .14cqi #54a0eb}}\n@keyframes moe-kv-glow{0%,100%{filter:brightness(1)}50%{filter:brightness(.96)}}\n@keyframes moe-chip-light{0%,100%{filter:brightness(1)}42%{filter:brightness(1.14)}}\n.moe-detailed{overflow:hidden;}\n.moe-inner-flow{display:block;position:absolute;left:0;top:25%;width:100%;height:50%;background:repeating-linear-gradient(90deg,transparent 0 .65cqi,#f5fdff .65cqi 1.25cqi,transparent 1.25cqi 2.4cqi);background-size:2.4cqi 100%;animation:moe-flow-x 1.2s linear infinite;pointer-events:none;}\n.moe-line-up .moe-inner-flow,.moe-line-down .moe-inner-flow{left:25%;top:0;width:50%;height:100%;background:repeating-linear-gradient(180deg,transparent 0 .65cqi,#f5fdff .65cqi 1.25cqi,transparent 1.25cqi 2.4cqi);background-size:100% 2.4cqi;animation-name:moe-flow-y;}\n.moe-line-up .moe-inner-flow,.moe-line-left .moe-inner-flow{animation-direction:reverse;}\n.moe-stage{animation-duration:6s;animation-iteration-count:infinite;animation-timing-function:ease-in-out;}\n.moe-phase-attn{animation-name:moe-attn-phase}.moe-phase-route{animation-name:moe-router-phase}.moe-phase-cpu,.moe-phase-worker{animation-name:moe-cpu-phase}.moe-phase-combine{animation-name:moe-combine-phase}\n.moe-expert-active{animation:moe-expert-phase 6s ease-in-out infinite;}\n.moe-hbm-frame{animation:moe-hbm-phase 6s ease-in-out infinite;}\n.moe-cache-access{animation:moe-kv-glow 3s ease-in-out infinite;}\n@keyframes moe-attn-phase{0%,30%,100%{box-shadow:none}9%,21%{box-shadow:inset 0 0 0 .13cqi #a96cde}}\n@keyframes moe-router-phase{0%,15%,44%,100%{box-shadow:none}23%,35%{box-shadow:inset 0 0 0 .13cqi #a96cde}}\n@keyframes moe-cpu-phase{0%,35%,79%,100%{box-shadow:none}45%,67%{box-shadow:inset 0 0 0 .13cqi #629246}}\n@keyframes moe-combine-phase{0%,68%,100%{box-shadow:none}77%,90%{box-shadow:inset 0 0 0 .13cqi #629246}}\n@keyframes moe-expert-phase{0%,32%,80%,100%{filter:brightness(1)}46%,64%{filter:brightness(.91)}}\n@keyframes moe-hbm-phase{0%,100%{box-shadow:none}10%,22%,77%,88%{box-shadow:inset 0 0 0 .075cqi #c99d2b}}\n.moe-motion-switch:checked ~ .moe-article-content .moe-diagram *, .moe-motion-switch:checked ~ .moe-article-content .moe-diagram *:after{animation-play-state:paused!important;}\n@media(prefers-reduced-motion:reduce){.moe-diagram *,.moe-diagram *:after{animation-play-state:paused!important;}}\n@media print{.moe-diagram *,.moe-diagram *:after{animation:none!important;}.moe-motion-switch,.moe-motion-label,.moe-enlarge{display:none!important;}}\n\n\/* Hot\/cold residency and execution use the same animation primitives as the overview. *\/\n.moe-hc-label{position:absolute;display:flex;align-items:center;justify-content:flex-start;line-height:1.24;text-align:left;white-space:nowrap;word-break:keep-all;margin:0;padding:0;letter-spacing:0;z-index:4;}\n.moe-hc-label>span{display:block;max-width:100%;}\n.moe-phase-hot{animation-name:moe-gpu-glow;}\n.moe-hc-diagram [data-component=\"cold-expert-weights\"]{animation:moe-expert-phase 6s ease-in-out infinite;}\n.moe-hc-diagram .moe-cold-result{background:#629482;}\n\n\/* Directed routes: one terminal head for each complete connector, none at bends. *\/\n.moe-hc-diagram .moe-route{display:contents;}\n.moe-hc-diagram .moe-route-segment{\n  display:block;position:absolute;box-sizing:border-box;margin:0;padding:0;border:0;\n  pointer-events:none;z-index:2;transform-origin:0 50%;overflow:hidden;\n  background:var(--moe-route-color);border-radius:0;\n}\n.moe-hc-diagram .moe-route-segment::before,\n.moe-hc-diagram .moe-route-segment::after{content:none!important;display:none!important;}\n.moe-hc-diagram .moe-route-packets{\n  display:block;position:absolute;left:0;right:0;top:28%;height:44%;margin:0;padding:0;\n  border:0;pointer-events:none;z-index:0;\n  background:repeating-linear-gradient(90deg,transparent 0 .65cqi,#effdff .65cqi 1.25cqi,transparent 1.25cqi 3cqi);\n  background-size:3cqi 100%;animation:moe-routed-flow 1.2s linear infinite;\n}\n@keyframes moe-routed-flow{\n  from{background-position:var(--moe-route-phase,0cqi) 0;}\n  to{background-position:calc(var(--moe-route-phase,0cqi) + 3cqi) 0;}\n}\n.moe-hc-diagram .moe-route-head{\n  display:block;position:absolute;margin:0!important;padding:0!important;border:0!important;\n  box-sizing:border-box;z-index:8;pointer-events:none;background:var(--moe-route-color);\n  opacity:1!important;animation:none!important;border-radius:0!important;\n  filter:drop-shadow(0 0 .045cqi white);\n}\n.moe-hc-diagram .moe-route-head::before,\n.moe-hc-diagram .moe-route-head::after{content:none!important;display:none!important;}\n.moe-hc-diagram .moe-route-head-right{\n  clip-path:polygon(0 0,100% 50%,0 100%);transform:translate(-100%,-50%);\n}\n.moe-hc-diagram .moe-route-head-left{\n  clip-path:polygon(100% 0,0 50%,100% 100%);transform:translate(0,-50%);\n}\n.moe-hc-diagram .moe-route-head-up{\n  clip-path:polygon(0 100%,50% 0,100% 100%);transform:translate(-50%,0);\n}\n.moe-hc-diagram .moe-route-head-down{\n  clip-path:polygon(0 0,50% 100%,100% 0);transform:translate(-50%,-100%);\n}\n\n<\/style>\n<article class=\"moe-post\"><header class=\"moe-header\"><p class=\"moe-eyebrow\">LLM INFERENCE \u00b7 CPU\u2013GPU HYBRID<\/p><h1 class=\"moe-title\">CPU MoE Offloading:<br>GPU \uba54\ubaa8\ub9ac\uc758 \ud55c\uacc4\ub97c \ub118\uc73c\uba74 \ubb34\uc5c7\uc774 \ub2ec\ub77c\uc9c8\uae4c<\/h1><p class=\"moe-deck\">Hot expert\ub294 GPU\uc5d0, Cold expert\ub294 Host Memory\uc5d0: \uc2e4\ud589 \uad6c\uc870\uc640 \uae30\ubcf8 \uc131\ub2a5, \ucd5c\uc801\ud654\uc758 \ucd9c\ubc1c\uc810<\/p><\/header><input class=\"moe-motion-switch\" id=\"moe-pause-all\" type=\"checkbox\"><label class=\"moe-motion-label\" for=\"moe-pause-all\">\uadf8\ub9bc \uc6c0\uc9c1\uc784 \uba48\ucda4<\/label><div class=\"moe-article-content\"><div class=\"moe-prose\">\n<p>\ud070 \uc5b8\uc5b4 \ubaa8\ub378\uc744 \uc2e4\ud589\ud560 \ub54c\ub294 \uc5f0\uc0b0 \uc18d\ub3c4\uc5d0 \uc55e\uc11c <strong>\uac00\uc911\uce58\uc640 \uc2e4\ud589 \uc0c1\ud0dc\ub97c GPU \uba54\ubaa8\ub9ac\uc5d0 \ub2f4\uc744 \uc218 \uc788\ub294\uc9c0<\/strong>\ubd80\ud130 \ud655\uc778\ud574\uc57c \ud55c\ub2e4. MoE \ubaa8\ub378\uc740 \ud1a0\ud070\ub9c8\ub2e4 \uc77c\ubd80 \uc804\ubb38\uac00\ub9cc \uacc4\uc0b0\ud558\uc9c0\ub9cc, \uc120\ud0dd \uac00\ub2a5\ud55c \uc804\ubb38\uac00\uc758 \uac00\uc911\uce58\ub97c \uc2e4\ud589 \uac00\ub2a5\ud55c \uc704\uce58\uc5d0 \ubcf4\uad00\ud574\uc57c \ud55c\ub2e4. \ubaa8\ub4e0 \uac00\uc911\uce58\ub97c \uace0\uc18d GPU \uba54\ubaa8\ub9ac\uc5d0 \ub450\uae30 \uc5b4\ub835\ub2e4\uba74, \uac19\uc740 \uc11c\ubc84\uc758 CPU \uba54\ubaa8\ub9ac\ub97c \ud568\uaed8 \uc0ac\uc6a9\ud558\ub294 \ubc29\ubc95\uc774 \uc788\ub2e4.<\/p>\n<p><strong>Hot expert\ub294 GPU \uba54\ubaa8\ub9ac\uc5d0 \ub0a8\uaca8 GPU\uac00 \uacc4\uc0b0\ud558\uace0, Cold expert\ub294 Host Memory\uc5d0 \ub450\uc5b4 CPU\uac00 \uacc4\uc0b0\ud55c\ub2e4.<\/strong> \uc774 \uae00\uc5d0\uc11c \ub2e4\ub8e8\ub294 CPU MoE Offloading\uc740 \uc774\ub7f0 \ubd80\ubd84 \uc0c1\uc8fc \ubc29\uc2dd\uc774\ub2e4. GPU\ub294 Attention\uacfc Router, \uc790\uc8fc \uc120\ud0dd\ub418\ub294 \uc804\ubb38\uac00\ub97c \ucc98\ub9ac\ud558\uace0, CPU\ub294 \uc120\ud0dd\ub41c Cold expert\ub97c \uc2e4\ud589\ud55c\ub2e4. \uad00\uc2ec\uc0ac\ub294 \ub2e8\uc21c\ud788 \ubaa8\ub378\uc744 \uc801\uc7ac\ud558\ub294 \uac83\uc774 \uc544\ub2c8\ub77c, \ub450 \uc7a5\uce58\uac00 \uacc4\uc0b0\uc744 \ub098\ub204\uba74\uc11c \uc0dd\uae30\ub294 \uc804\ub2ec \ube44\uc6a9\uacfc \uacb0\uacfc \ub300\uae30\ub97c \uc5bc\ub9c8\ub098 \uc904\uc77c \uc218 \uc788\ub290\ub0d0\ub2e4.<\/p>\n<\/div><figure class=\"moe-figure\"><div aria-label=\"CPU MoE \uc804\ubb38\uac00 \ubd80\ubd84 \uc0c1\uc8fc \uac1c\ub150\ub3c4. \uc67c\ucabd \ud074\ub77c\uc774\uc5b8\ud2b8, \uac00\uc6b4\ub370 SGLang \ub7f0\ud0c0\uc784, \uc624\ub978\ucabd \ub3d9\uc77c \uc11c\ubc84\uc758 GPU\uc640 CPU. CPU\ub294 \ucd1d 112\ubb3c\ub9ac \ucf54\uc5b4, 224\ub17c\ub9ac \uc2a4\ub808\ub4dc.\" class=\"moe-diagram moe-overview\" role=\"img\"><div class=\"moe-ov-el\" style=\"left:1.60000%;top:3.75000%;width:0.50000%;height:6.25000%;background:#428ef4;border-radius:.3cqi;\"><\/div><div class=\"moe-ov-el\" data-text=\"CPU Offloading\uc744 \uc801\uc6a9\ud55c MoE \ucd94\ub860 \uad6c\uc870\" style=\"left:3.40000%;top:3.00000%;width:93.50000%;height:7.75000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.900cqi;font-weight:700;line-height:1.32;text-align:left;color:#173b69;\">CPU Offloading\uc744 \uc801\uc6a9\ud55c MoE \ucd94\ub860 \uad6c\uc870<\/div><div class=\"moe-ov-el\" data-text=\"GPU\uc640 CPU\ub294 \uac19\uc740 \ubb3c\ub9ac \uc11c\ubc84\uc5d0\uc11c \uc804\ubb38\uac00 \uacc4\uc0b0\uc744 \ub098\ub208\ub2e4.\" style=\"left:3.40000%;top:11.50000%;width:93.00000%;height:6.00000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.300cqi;font-weight:400;line-height:1.32;text-align:left;color:#60738e;\">GPU\uc640 CPU\ub294 \uac19\uc740 \ubb3c\ub9ac \uc11c\ubc84\uc5d0\uc11c \uc804\ubb38\uac00 \uacc4\uc0b0\uc744 \ub098\ub208\ub2e4.<\/div><div class=\"moe-ov-el\" style=\"left:1.80000%;top:26.00000%;width:9.40000%;height:57.25000%;background:#f6f9fd;border:0.10cqi solid #c4d7ef;border-radius:0.600cqi;\"><\/div><div class=\"moe-ov-el\" style=\"left:1.80000%;top:26.00000%;width:9.40000%;height:10.25000%;background:#ecf3fb;border-radius:.6cqi .6cqi 0 0;\"><\/div><div class=\"moe-ov-el\" data-text=\"Client\" style=\"left:2.50000%;top:28.50000%;width:8.00000%;height:5.50000%;display:flex;align-items:center;justify-content:center;font-size:1.400cqi;font-weight:700;line-height:1.32;text-align:center;color:#173b69;\">Client<\/div><div aria-hidden=\"true\" class=\"moe-ov-el moe-prompt-icon\" style=\"left:4.50000%;top:41.50000%;width:3.70000%;height:7.50000%;\"><span style=\"display:block;position:absolute;left:0;top:0;width:100%;height:86%;border:.26cqi solid #194c8c;border-radius:40%;background:#f6f9fd;\"><span class=\"moe-chat-dots\" style=\"z-index:2;\"><i><\/i><i><\/i><i><\/i><\/span><\/span><span style=\"display:block;position:absolute;left:19%;top:64%;width:.80cqi;height:.80cqi;transform:rotate(45deg);border-right:.26cqi solid #194c8c;border-bottom:.26cqi solid #194c8c;background:#f6f9fd;z-index:1;\"><\/span><\/div><div class=\"moe-ov-el\" data-text=\"\uc694\uccad (\ud504\ub86c\ud504\ud2b8)\" style=\"left:2.20000%;top:51.75000%;width:8.60000%;height:10.25000%;display:flex;align-items:center;justify-content:center;font-size:1.300cqi;font-weight:700;line-height:1.32;text-align:center;color:#173b69;\">\uc694\uccad<br>(\ud504\ub86c\ud504\ud2b8)<\/div><div class=\"moe-ov-el\" data-text=\"\uc751\ub2f5 (\ud1a0\ud070 \uc2a4\ud2b8\ub9bc)\" style=\"left:2.20000%;top:64.75000%;width:8.60000%;height:10.50000%;display:flex;align-items:center;justify-content:center;font-size:1.300cqi;font-weight:700;line-height:1.32;text-align:center;color:#173b69;\">\uc751\ub2f5<br>(\ud1a0\ud070 \uc2a4\ud2b8\ub9bc)<\/div><div class=\"moe-ov-el\" style=\"left:16.80000%;top:22.00000%;width:53.10000%;height:64.00000%;background:#f4f8ff;border:0.10cqi solid #85b6fc;border-radius:0.700cqi;\"><\/div><div class=\"moe-ov-el\" data-text=\"\u25a5 SGLang Server (TP4)\" style=\"left:18.00000%;top:24.00000%;width:47.00000%;height:7.25000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.500cqi;font-weight:700;line-height:1.32;text-align:left;color:#173b69;\">\u25a5 SGLang Server (TP4)<\/div><div class=\"moe-ov-el\" style=\"left:17.70000%;top:35.75000%;width:8.70000%;height:32.25000%;background:#f7faff;border:0.10cqi solid #b5cee9;border-radius:0.500cqi;\"><\/div><div class=\"moe-ov-el\" data-text=\"Tokenizer &amp;amp; Scheduler\" style=\"left:18.10000%;top:45.50000%;width:7.90000%;height:12.25000%;display:flex;align-items:center;justify-content:center;font-size:1.250cqi;font-weight:700;line-height:1.32;text-align:center;color:#173b69;\">Tokenizer<br>&amp; Scheduler<\/div><div class=\"moe-ov-el moe-ov-dense\" style=\"left:27.40000%;top:36.00000%;width:12.40000%;height:32.00000%;background:#f7faff;border:0.10cqi solid #b5cee9;border-radius:0.500cqi;\"><\/div><div class=\"moe-ov-el\" data-text=\"Transformer (Attention + Dense)\" style=\"left:27.90000%;top:37.75000%;width:11.40000%;height:13.50000%;display:flex;align-items:center;justify-content:center;font-size:1.220cqi;font-weight:500;line-height:1.32;text-align:center;color:#173b69;\">Transformer<br>(Attention + Dense)<\/div><div class=\"moe-ov-el moe-ov-token\" style=\"left:28.20000%;top:53.00000%;width:1.60000%;height:6.50000%;background:#32bc6c;border-radius:.25cqi;--moe-delay:0.0s;\"><\/div><div class=\"moe-ov-el moe-ov-token\" style=\"left:30.00000%;top:53.00000%;width:1.60000%;height:6.50000%;background:#32a9eb;border-radius:.25cqi;--moe-delay:0.13s;\"><\/div><div class=\"moe-ov-el moe-ov-token\" style=\"left:31.80000%;top:53.00000%;width:1.60000%;height:6.50000%;background:#6688f2;border-radius:.25cqi;--moe-delay:0.26s;\"><\/div><div class=\"moe-ov-el moe-ov-token\" style=\"left:33.60000%;top:53.00000%;width:1.60000%;height:6.50000%;background:#6a6ceb;border-radius:.25cqi;--moe-delay:0.39s;\"><\/div><div class=\"moe-ov-el moe-ov-token\" style=\"left:35.40000%;top:53.00000%;width:1.60000%;height:6.50000%;background:#4b72c1;border-radius:.25cqi;--moe-delay:0.52s;\"><\/div><div class=\"moe-ov-el moe-ov-token\" style=\"left:37.20000%;top:53.00000%;width:1.60000%;height:6.50000%;background:#1d4179;border-radius:.25cqi;--moe-delay:0.65s;\"><\/div><div class=\"moe-ov-el moe-ov-moe\" style=\"left:41.60000%;top:30.75000%;width:27.00000%;height:40.25000%;background:#fbfdff;border:0.10cqi solid #468ffa;border-radius:0.700cqi;\"><\/div><div class=\"moe-ov-el\" data-text=\"MoE Layer\" style=\"left:45.50000%;top:32.50000%;width:19.80000%;height:6.00000%;display:flex;align-items:center;justify-content:center;font-size:1.350cqi;font-weight:700;line-height:1.32;text-align:center;color:#173b69;\">MoE Layer<\/div><div class=\"moe-ov-el moe-ov-router\" style=\"left:43.20000%;top:42.50000%;width:7.10000%;height:19.25000%;background:#f8fbff;border:0.10cqi solid #c5d9f3;border-radius:0.500cqi;\"><\/div><div class=\"moe-ov-el\" data-text=\"Router (Top-K)\" style=\"left:43.50000%;top:47.50000%;width:6.50000%;height:10.00000%;display:flex;align-items:center;justify-content:center;font-size:1.250cqi;font-weight:600;line-height:1.32;text-align:center;color:#173b69;\">Router<br>(Top-K)<\/div><div class=\"moe-ov-el moe-ov-hot\" style=\"left:56.00000%;top:39.50000%;width:11.50000%;height:14.00000%;background:#e4f7e9;border:0.10cqi solid #6cc78c;border-radius:0.500cqi;\"><\/div><div class=\"moe-ov-el\" data-text=\"Hot Experts (GPU, FP8)\" style=\"left:56.50000%;top:42.25000%;width:10.50000%;height:8.75000%;display:flex;align-items:center;justify-content:center;font-size:1.200cqi;font-weight:600;line-height:1.32;text-align:center;color:#143f36;\">Hot Experts<br>(GPU, FP8)<\/div><div class=\"moe-ov-el moe-ov-cold\" style=\"left:56.00000%;top:55.75000%;width:11.50000%;height:13.75000%;background:#eaf4ff;border:0.10cqi solid #8fc8ff;border-radius:0.500cqi;\"><\/div><div class=\"moe-ov-el\" data-text=\"Cold Experts (CPU, INT4)\" style=\"left:56.50000%;top:58.25000%;width:10.50000%;height:8.75000%;display:flex;align-items:center;justify-content:center;font-size:1.200cqi;font-weight:600;line-height:1.32;text-align:center;color:#173b69;\">Cold Experts<br>(CPU, INT4)<\/div><div class=\"moe-ov-el\" data-text=\"\ud65c\uc131\uac12\" style=\"left:50.70000%;top:42.75000%;width:4.90000%;height:4.25000%;display:flex;align-items:center;justify-content:center;font-size:1.050cqi;font-weight:500;line-height:1.32;text-align:center;color:#173b69;\">\ud65c\uc131\uac12<\/div><div class=\"moe-ov-el\" data-text=\"\uc120\ud0dd \uc815\ubcf4\" style=\"left:50.80000%;top:62.00000%;width:4.70000%;height:4.25000%;display:flex;align-items:center;justify-content:center;font-size:0.950cqi;font-weight:500;line-height:1.32;text-align:center;color:#173b69;\">\uc120\ud0dd \uc815\ubcf4<\/div><div class=\"moe-ov-el moe-ov-kv\" style=\"left:17.70000%;top:76.25000%;width:51.00000%;height:7.50000%;background:#e3f2ff;border:0.10cqi solid #78b7ff;border-radius:0.500cqi;\"><\/div><div class=\"moe-ov-el\" data-text=\"KV Cache \u00b7 GPU HBM\" style=\"left:18.50000%;top:77.25000%;width:49.40000%;height:5.75000%;display:flex;align-items:center;justify-content:center;font-size:1.350cqi;font-weight:600;line-height:1.32;text-align:center;color:#2458bb;\">KV Cache \u00b7 GPU HBM<\/div><div class=\"moe-ov-el moe-ov-gpu\" style=\"left:73.30000%;top:16.25000%;width:24.90000%;height:33.00000%;background:#eefbf3;border:0.10cqi solid #5fc984;border-radius:0.700cqi;\"><\/div><div class=\"moe-ov-el\" style=\"left:74.70000%;top:19.75000%;width:3.10000%;height:6.25000%;border-radius:.5cqi;background:#20ad59;box-shadow:inset .5cqi .3cqi 0 #49c673;\"><\/div><div class=\"moe-ov-el\" data-text=\"GPU (H100 \u00d7 4)\" style=\"left:79.80000%;top:18.50000%;width:17.70000%;height:7.25000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.500cqi;font-weight:700;line-height:1.32;text-align:left;color:#173b69;\">GPU (H100 \u00d7 4)<\/div><div class=\"moe-ov-el moe-chip-face moe-chip-0\" style=\"left:77.50000%;top:27.50000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#24b564;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el moe-chip-face moe-chip-1\" style=\"left:76.30000%;top:30.75000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#24b564;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el moe-chip-face moe-chip-2\" style=\"left:75.00000%;top:34.00000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#24b564;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el moe-chip-face moe-chip-3\" style=\"left:75.00000%;top:38.75000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#24b564;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el moe-chip-face moe-chip-4\" style=\"left:75.00000%;top:43.50000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#24b564;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el\" data-text=\"Hot Experts (\uc608: \uce35\ubcc4 96\uac1c)\" style=\"left:82.30000%;top:27.75000%;width:15.10000%;height:6.00000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.120cqi;font-weight:600;line-height:1.32;text-align:left;color:#173b69;\">Hot Experts (\uc608: \uce35\ubcc4 96\uac1c)<\/div><div class=\"moe-ov-el\" data-text=\"\u2022 Attention \/ Dense \u2022 \uc790\uc8fc \uc120\ud0dd\ub418\ub294 \uc804\ubb38\uac00 \u2022 FP8 \uc804\ubb38\uac00 \uac00\uc911\uce58\" style=\"left:82.30000%;top:34.25000%;width:15.30000%;height:12.00000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.140cqi;font-weight:400;line-height:1.32;text-align:left;color:#335167;\">\u2022 Attention \/ Dense<br>\u2022 \uc790\uc8fc \uc120\ud0dd\ub418\ub294 \uc804\ubb38\uac00<br>\u2022 FP8 \uc804\ubb38\uac00 \uac00\uc911\uce58<\/div><div class=\"moe-ov-el moe-ov-cpu\" style=\"left:73.30000%;top:59.75000%;width:24.90000%;height:34.25000%;background:#f3faff;border:0.10cqi solid #3498ff;border-radius:0.700cqi;\"><\/div><div class=\"moe-ov-el\" style=\"left:75.10000%;top:62.75000%;width:2.20000%;height:5.00000%;border:.24cqi solid #177ce8;box-shadow:.38cqi 0 0 -.17cqi #177ce8,-.38cqi 0 0 -.17cqi #177ce8;\"><span style=\"position:absolute;inset:22%;border:.18cqi solid #177ce8;\"><\/span><\/div><div class=\"moe-ov-el\" data-text=\"CPU (Xeon 8480+ \u00d7 2)\" style=\"left:78.60000%;top:62.00000%;width:18.60000%;height:5.75000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.330cqi;font-weight:700;line-height:1.32;text-align:left;color:#173b69;\">CPU (Xeon 8480+ \u00d7 2)<\/div><div class=\"moe-ov-el\" data-text=\"112\ucf54\uc5b4 \/ 224\ub17c\ub9ac \uc2a4\ub808\ub4dc\" style=\"left:78.60000%;top:67.75000%;width:18.60000%;height:4.25000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.050cqi;font-weight:400;line-height:1.32;text-align:left;color:#526d88;\">112\ucf54\uc5b4 \/ 224\ub17c\ub9ac \uc2a4\ub808\ub4dc<\/div><div class=\"moe-ov-el moe-chip-face moe-chip-0\" style=\"left:77.50000%;top:72.00000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#278be1;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el moe-chip-face moe-chip-1\" style=\"left:76.30000%;top:75.25000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#278be1;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el moe-chip-face moe-chip-2\" style=\"left:75.00000%;top:78.50000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#278be1;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el moe-chip-face moe-chip-3\" style=\"left:75.00000%;top:83.25000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#278be1;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el moe-chip-face moe-chip-4\" style=\"left:75.00000%;top:88.00000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#278be1;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el\" data-text=\"Cold Experts (\ub098\uba38\uc9c0)\" style=\"left:82.20000%;top:72.75000%;width:15.30000%;height:6.00000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.180cqi;font-weight:600;line-height:1.32;text-align:left;color:#173b69;\">Cold Experts (\ub098\uba38\uc9c0)<\/div><div class=\"moe-ov-el\" data-text=\"\u2022 \uc120\ud0dd\ub41c \uc804\ubb38\uac00\ub97c CPU\uc5d0\uc11c \uacc4\uc0b0 \u2022 INT4 \uc804\ubb38\uac00 \uac00\uc911\uce58 \u2022 CPU DRAM 2 TB\" style=\"left:82.30000%;top:80.00000%;width:15.40000%;height:11.75000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.080cqi;font-weight:400;line-height:1.32;text-align:left;color:#335167;\">\u2022 \uc120\ud0dd\ub41c \uc804\ubb38\uac00\ub97c CPU\uc5d0\uc11c \uacc4\uc0b0<br>\u2022 INT4 \uc804\ubb38\uac00 \uac00\uc911\uce58<br>\u2022 CPU DRAM 2 TB<\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:11.6000%;top:48.7500%;width:4.7000%;height:0.300cqi;transform:translateY(-50%) rotate(0.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:16.3000%;top:61.5000%;width:4.7000%;height:0.300cqi;transform:translateY(-50%) rotate(180.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:26.5000%;top:51.7500%;width:0.8000%;height:0.200cqi;transform:translateY(-50%) rotate(0.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:40.0000%;top:55.0000%;width:3.0000%;height:0.300cqi;transform:translateY(-50%) rotate(0.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:50.6000%;top:51.0000%;width:4.8836%;height:0.260cqi;transform:translateY(-50%) rotate(-10.6197deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:50.6000%;top:56.5000%;width:4.9729%;height:0.260cqi;transform:translateY(-50%) rotate(15.1541deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:70.1000%;top:38.2500%;width:2.8000%;height:0.280cqi;transform:translateY(-50%) rotate(0.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:81.0000%;top:51.2500%;width:2.8000%;height:0.280cqi;transform:translateY(-50%) rotate(90.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:82.9000%;top:58.2500%;width:2.8000%;height:0.280cqi;transform:translateY(-50%) rotate(-90.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:22.0000%;top:75.7500%;width:2.4000%;height:0.200cqi;transform:translateY(-50%) rotate(-90.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:33.7000%;top:75.7500%;width:2.4000%;height:0.200cqi;transform:translateY(-50%) rotate(-90.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:46.7000%;top:70.7500%;width:2.0000%;height:0.200cqi;transform:translateY(-50%) rotate(90.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div class=\"moe-ov-el\" data-text=\"PCIe \/ NUMA\" style=\"left:84.90000%;top:51.25000%;width:12.40000%;height:6.00000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.150cqi;font-weight:500;line-height:1.32;text-align:left;color:#173b69;\">PCIe \/ NUMA<\/div><div class=\"moe-ov-el\" data-text=\"\ubd80\ubd84 \uc0c1\uc8fc \uac1c\ub150\ub3c4 \u00b7 KV \uc815\ubc00\ub3c4\uc640 GPU \uc804\ubb38\uac00 \uc218\ub294 \uc124\uc815\uc5d0 \ub530\ub77c \ub2ec\ub77c\uc9c4\ub2e4.\" style=\"left:16.90000%;top:91.00000%;width:52.80000%;height:5.75000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.070cqi;font-weight:400;line-height:1.32;text-align:left;color:#61748a;\">\ubd80\ubd84 \uc0c1\uc8fc \uac1c\ub150\ub3c4 \u00b7 KV \uc815\ubc00\ub3c4\uc640 GPU \uc804\ubb38\uac00 \uc218\ub294 \uc124\uc815\uc5d0 \ub530\ub77c \ub2ec\ub77c\uc9c4\ub2e4.<\/div><\/div><figcaption class=\"moe-caption\">\uadf8\ub9bc 1. \uac19\uc740 \uc11c\ubc84\uc758 GPU\uc640 CPU\uac00 \uc804\ubb38\uac00 \uacc4\uc0b0\uc744 \ubd84\ub2f4\ud55c\ub2e4. Hot expert\ub294 GPU \uba54\ubaa8\ub9ac\uc5d0, Cold expert\ub294 Host Memory\uc5d0 \uc0c1\uc8fc\ud55c\ub2e4.<\/figcaption><details class=\"moe-enlarge\"><summary>\uad6c\uc870\ub3c4 \ud655\ub300 \ubcf4\uae30<\/summary><div class=\"moe-zoom-viewport\" tabindex=\"0\"><div class=\"moe-zoom-intro\"><div aria-label=\"CPU MoE \uc804\ubb38\uac00 \ubd80\ubd84 \uc0c1\uc8fc \uac1c\ub150\ub3c4. \uc67c\ucabd \ud074\ub77c\uc774\uc5b8\ud2b8, \uac00\uc6b4\ub370 SGLang \ub7f0\ud0c0\uc784, \uc624\ub978\ucabd \ub3d9\uc77c \uc11c\ubc84\uc758 GPU\uc640 CPU. CPU\ub294 \ucd1d 112\ubb3c\ub9ac \ucf54\uc5b4, 224\ub17c\ub9ac \uc2a4\ub808\ub4dc.\" class=\"moe-diagram moe-overview\" role=\"img\"><div class=\"moe-ov-el\" style=\"left:1.60000%;top:3.75000%;width:0.50000%;height:6.25000%;background:#428ef4;border-radius:.3cqi;\"><\/div><div class=\"moe-ov-el\" data-text=\"CPU Offloading\uc744 \uc801\uc6a9\ud55c MoE \ucd94\ub860 \uad6c\uc870\" style=\"left:3.40000%;top:3.00000%;width:93.50000%;height:7.75000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.900cqi;font-weight:700;line-height:1.32;text-align:left;color:#173b69;\">CPU Offloading\uc744 \uc801\uc6a9\ud55c MoE \ucd94\ub860 \uad6c\uc870<\/div><div class=\"moe-ov-el\" data-text=\"GPU\uc640 CPU\ub294 \uac19\uc740 \ubb3c\ub9ac \uc11c\ubc84\uc5d0\uc11c \uc804\ubb38\uac00 \uacc4\uc0b0\uc744 \ub098\ub208\ub2e4.\" style=\"left:3.40000%;top:11.50000%;width:93.00000%;height:6.00000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.300cqi;font-weight:400;line-height:1.32;text-align:left;color:#60738e;\">GPU\uc640 CPU\ub294 \uac19\uc740 \ubb3c\ub9ac \uc11c\ubc84\uc5d0\uc11c \uc804\ubb38\uac00 \uacc4\uc0b0\uc744 \ub098\ub208\ub2e4.<\/div><div class=\"moe-ov-el\" style=\"left:1.80000%;top:26.00000%;width:9.40000%;height:57.25000%;background:#f6f9fd;border:0.10cqi solid #c4d7ef;border-radius:0.600cqi;\"><\/div><div class=\"moe-ov-el\" style=\"left:1.80000%;top:26.00000%;width:9.40000%;height:10.25000%;background:#ecf3fb;border-radius:.6cqi .6cqi 0 0;\"><\/div><div class=\"moe-ov-el\" data-text=\"Client\" style=\"left:2.50000%;top:28.50000%;width:8.00000%;height:5.50000%;display:flex;align-items:center;justify-content:center;font-size:1.400cqi;font-weight:700;line-height:1.32;text-align:center;color:#173b69;\">Client<\/div><div aria-hidden=\"true\" class=\"moe-ov-el moe-prompt-icon\" style=\"left:4.50000%;top:41.50000%;width:3.70000%;height:7.50000%;\"><span style=\"display:block;position:absolute;left:0;top:0;width:100%;height:86%;border:.26cqi solid #194c8c;border-radius:40%;background:#f6f9fd;\"><span class=\"moe-chat-dots\" style=\"z-index:2;\"><i><\/i><i><\/i><i><\/i><\/span><\/span><span style=\"display:block;position:absolute;left:19%;top:64%;width:.80cqi;height:.80cqi;transform:rotate(45deg);border-right:.26cqi solid #194c8c;border-bottom:.26cqi solid #194c8c;background:#f6f9fd;z-index:1;\"><\/span><\/div><div class=\"moe-ov-el\" data-text=\"\uc694\uccad (\ud504\ub86c\ud504\ud2b8)\" style=\"left:2.20000%;top:51.75000%;width:8.60000%;height:10.25000%;display:flex;align-items:center;justify-content:center;font-size:1.300cqi;font-weight:700;line-height:1.32;text-align:center;color:#173b69;\">\uc694\uccad<br>(\ud504\ub86c\ud504\ud2b8)<\/div><div class=\"moe-ov-el\" data-text=\"\uc751\ub2f5 (\ud1a0\ud070 \uc2a4\ud2b8\ub9bc)\" style=\"left:2.20000%;top:64.75000%;width:8.60000%;height:10.50000%;display:flex;align-items:center;justify-content:center;font-size:1.300cqi;font-weight:700;line-height:1.32;text-align:center;color:#173b69;\">\uc751\ub2f5<br>(\ud1a0\ud070 \uc2a4\ud2b8\ub9bc)<\/div><div class=\"moe-ov-el\" style=\"left:16.80000%;top:22.00000%;width:53.10000%;height:64.00000%;background:#f4f8ff;border:0.10cqi solid #85b6fc;border-radius:0.700cqi;\"><\/div><div class=\"moe-ov-el\" data-text=\"\u25a5 SGLang Server (TP4)\" style=\"left:18.00000%;top:24.00000%;width:47.00000%;height:7.25000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.500cqi;font-weight:700;line-height:1.32;text-align:left;color:#173b69;\">\u25a5 SGLang Server (TP4)<\/div><div class=\"moe-ov-el\" style=\"left:17.70000%;top:35.75000%;width:8.70000%;height:32.25000%;background:#f7faff;border:0.10cqi solid #b5cee9;border-radius:0.500cqi;\"><\/div><div class=\"moe-ov-el\" data-text=\"Tokenizer &amp;amp; Scheduler\" style=\"left:18.10000%;top:45.50000%;width:7.90000%;height:12.25000%;display:flex;align-items:center;justify-content:center;font-size:1.250cqi;font-weight:700;line-height:1.32;text-align:center;color:#173b69;\">Tokenizer<br>&amp; Scheduler<\/div><div class=\"moe-ov-el moe-ov-dense\" style=\"left:27.40000%;top:36.00000%;width:12.40000%;height:32.00000%;background:#f7faff;border:0.10cqi solid #b5cee9;border-radius:0.500cqi;\"><\/div><div class=\"moe-ov-el\" data-text=\"Transformer (Attention + Dense)\" style=\"left:27.90000%;top:37.75000%;width:11.40000%;height:13.50000%;display:flex;align-items:center;justify-content:center;font-size:1.220cqi;font-weight:500;line-height:1.32;text-align:center;color:#173b69;\">Transformer<br>(Attention + Dense)<\/div><div class=\"moe-ov-el moe-ov-token\" style=\"left:28.20000%;top:53.00000%;width:1.60000%;height:6.50000%;background:#32bc6c;border-radius:.25cqi;--moe-delay:0.0s;\"><\/div><div class=\"moe-ov-el moe-ov-token\" style=\"left:30.00000%;top:53.00000%;width:1.60000%;height:6.50000%;background:#32a9eb;border-radius:.25cqi;--moe-delay:0.13s;\"><\/div><div class=\"moe-ov-el moe-ov-token\" style=\"left:31.80000%;top:53.00000%;width:1.60000%;height:6.50000%;background:#6688f2;border-radius:.25cqi;--moe-delay:0.26s;\"><\/div><div class=\"moe-ov-el moe-ov-token\" style=\"left:33.60000%;top:53.00000%;width:1.60000%;height:6.50000%;background:#6a6ceb;border-radius:.25cqi;--moe-delay:0.39s;\"><\/div><div class=\"moe-ov-el moe-ov-token\" style=\"left:35.40000%;top:53.00000%;width:1.60000%;height:6.50000%;background:#4b72c1;border-radius:.25cqi;--moe-delay:0.52s;\"><\/div><div class=\"moe-ov-el moe-ov-token\" style=\"left:37.20000%;top:53.00000%;width:1.60000%;height:6.50000%;background:#1d4179;border-radius:.25cqi;--moe-delay:0.65s;\"><\/div><div class=\"moe-ov-el moe-ov-moe\" style=\"left:41.60000%;top:30.75000%;width:27.00000%;height:40.25000%;background:#fbfdff;border:0.10cqi solid #468ffa;border-radius:0.700cqi;\"><\/div><div class=\"moe-ov-el\" data-text=\"MoE Layer\" style=\"left:45.50000%;top:32.50000%;width:19.80000%;height:6.00000%;display:flex;align-items:center;justify-content:center;font-size:1.350cqi;font-weight:700;line-height:1.32;text-align:center;color:#173b69;\">MoE Layer<\/div><div class=\"moe-ov-el moe-ov-router\" style=\"left:43.20000%;top:42.50000%;width:7.10000%;height:19.25000%;background:#f8fbff;border:0.10cqi solid #c5d9f3;border-radius:0.500cqi;\"><\/div><div class=\"moe-ov-el\" data-text=\"Router (Top-K)\" style=\"left:43.50000%;top:47.50000%;width:6.50000%;height:10.00000%;display:flex;align-items:center;justify-content:center;font-size:1.250cqi;font-weight:600;line-height:1.32;text-align:center;color:#173b69;\">Router<br>(Top-K)<\/div><div class=\"moe-ov-el moe-ov-hot\" style=\"left:56.00000%;top:39.50000%;width:11.50000%;height:14.00000%;background:#e4f7e9;border:0.10cqi solid #6cc78c;border-radius:0.500cqi;\"><\/div><div class=\"moe-ov-el\" data-text=\"Hot Experts (GPU, FP8)\" style=\"left:56.50000%;top:42.25000%;width:10.50000%;height:8.75000%;display:flex;align-items:center;justify-content:center;font-size:1.200cqi;font-weight:600;line-height:1.32;text-align:center;color:#143f36;\">Hot Experts<br>(GPU, FP8)<\/div><div class=\"moe-ov-el moe-ov-cold\" style=\"left:56.00000%;top:55.75000%;width:11.50000%;height:13.75000%;background:#eaf4ff;border:0.10cqi solid #8fc8ff;border-radius:0.500cqi;\"><\/div><div class=\"moe-ov-el\" data-text=\"Cold Experts (CPU, INT4)\" style=\"left:56.50000%;top:58.25000%;width:10.50000%;height:8.75000%;display:flex;align-items:center;justify-content:center;font-size:1.200cqi;font-weight:600;line-height:1.32;text-align:center;color:#173b69;\">Cold Experts<br>(CPU, INT4)<\/div><div class=\"moe-ov-el\" data-text=\"\ud65c\uc131\uac12\" style=\"left:50.70000%;top:42.75000%;width:4.90000%;height:4.25000%;display:flex;align-items:center;justify-content:center;font-size:1.050cqi;font-weight:500;line-height:1.32;text-align:center;color:#173b69;\">\ud65c\uc131\uac12<\/div><div class=\"moe-ov-el\" data-text=\"\uc120\ud0dd \uc815\ubcf4\" style=\"left:50.80000%;top:62.00000%;width:4.70000%;height:4.25000%;display:flex;align-items:center;justify-content:center;font-size:0.950cqi;font-weight:500;line-height:1.32;text-align:center;color:#173b69;\">\uc120\ud0dd \uc815\ubcf4<\/div><div class=\"moe-ov-el moe-ov-kv\" style=\"left:17.70000%;top:76.25000%;width:51.00000%;height:7.50000%;background:#e3f2ff;border:0.10cqi solid #78b7ff;border-radius:0.500cqi;\"><\/div><div class=\"moe-ov-el\" data-text=\"KV Cache \u00b7 GPU HBM\" style=\"left:18.50000%;top:77.25000%;width:49.40000%;height:5.75000%;display:flex;align-items:center;justify-content:center;font-size:1.350cqi;font-weight:600;line-height:1.32;text-align:center;color:#2458bb;\">KV Cache \u00b7 GPU HBM<\/div><div class=\"moe-ov-el moe-ov-gpu\" style=\"left:73.30000%;top:16.25000%;width:24.90000%;height:33.00000%;background:#eefbf3;border:0.10cqi solid #5fc984;border-radius:0.700cqi;\"><\/div><div class=\"moe-ov-el\" style=\"left:74.70000%;top:19.75000%;width:3.10000%;height:6.25000%;border-radius:.5cqi;background:#20ad59;box-shadow:inset .5cqi .3cqi 0 #49c673;\"><\/div><div class=\"moe-ov-el\" data-text=\"GPU (H100 \u00d7 4)\" style=\"left:79.80000%;top:18.50000%;width:17.70000%;height:7.25000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.500cqi;font-weight:700;line-height:1.32;text-align:left;color:#173b69;\">GPU (H100 \u00d7 4)<\/div><div class=\"moe-ov-el moe-chip-face moe-chip-0\" style=\"left:77.50000%;top:27.50000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#24b564;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el moe-chip-face moe-chip-1\" style=\"left:76.30000%;top:30.75000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#24b564;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el moe-chip-face moe-chip-2\" style=\"left:75.00000%;top:34.00000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#24b564;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el moe-chip-face moe-chip-3\" style=\"left:75.00000%;top:38.75000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#24b564;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el moe-chip-face moe-chip-4\" style=\"left:75.00000%;top:43.50000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#24b564;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el\" data-text=\"Hot Experts (\uc608: \uce35\ubcc4 96\uac1c)\" style=\"left:82.30000%;top:27.75000%;width:15.10000%;height:6.00000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.120cqi;font-weight:600;line-height:1.32;text-align:left;color:#173b69;\">Hot Experts (\uc608: \uce35\ubcc4 96\uac1c)<\/div><div class=\"moe-ov-el\" data-text=\"\u2022 Attention \/ Dense \u2022 \uc790\uc8fc \uc120\ud0dd\ub418\ub294 \uc804\ubb38\uac00 \u2022 FP8 \uc804\ubb38\uac00 \uac00\uc911\uce58\" style=\"left:82.30000%;top:34.25000%;width:15.30000%;height:12.00000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.140cqi;font-weight:400;line-height:1.32;text-align:left;color:#335167;\">\u2022 Attention \/ Dense<br>\u2022 \uc790\uc8fc \uc120\ud0dd\ub418\ub294 \uc804\ubb38\uac00<br>\u2022 FP8 \uc804\ubb38\uac00 \uac00\uc911\uce58<\/div><div class=\"moe-ov-el moe-ov-cpu\" style=\"left:73.30000%;top:59.75000%;width:24.90000%;height:34.25000%;background:#f3faff;border:0.10cqi solid #3498ff;border-radius:0.700cqi;\"><\/div><div class=\"moe-ov-el\" style=\"left:75.10000%;top:62.75000%;width:2.20000%;height:5.00000%;border:.24cqi solid #177ce8;box-shadow:.38cqi 0 0 -.17cqi #177ce8,-.38cqi 0 0 -.17cqi #177ce8;\"><span style=\"position:absolute;inset:22%;border:.18cqi solid #177ce8;\"><\/span><\/div><div class=\"moe-ov-el\" data-text=\"CPU (Xeon 8480+ \u00d7 2)\" style=\"left:78.60000%;top:62.00000%;width:18.60000%;height:5.75000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.330cqi;font-weight:700;line-height:1.32;text-align:left;color:#173b69;\">CPU (Xeon 8480+ \u00d7 2)<\/div><div class=\"moe-ov-el\" data-text=\"112\ucf54\uc5b4 \/ 224\ub17c\ub9ac \uc2a4\ub808\ub4dc\" style=\"left:78.60000%;top:67.75000%;width:18.60000%;height:4.25000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.050cqi;font-weight:400;line-height:1.32;text-align:left;color:#526d88;\">112\ucf54\uc5b4 \/ 224\ub17c\ub9ac \uc2a4\ub808\ub4dc<\/div><div class=\"moe-ov-el moe-chip-face moe-chip-0\" style=\"left:77.50000%;top:72.00000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#278be1;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el moe-chip-face moe-chip-1\" style=\"left:76.30000%;top:75.25000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#278be1;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el moe-chip-face moe-chip-2\" style=\"left:75.00000%;top:78.50000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#278be1;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el moe-chip-face moe-chip-3\" style=\"left:75.00000%;top:83.25000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#278be1;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el moe-chip-face moe-chip-4\" style=\"left:75.00000%;top:88.00000%;width:3.00000%;height:4.50000%;border-radius:.3cqi;background:#278be1;border:.10cqi solid rgba(0,0,0,.09);box-shadow:inset .38cqi .25cqi 0 rgba(255,255,255,.19);\"><\/div><div class=\"moe-ov-el\" data-text=\"Cold Experts (\ub098\uba38\uc9c0)\" style=\"left:82.20000%;top:72.75000%;width:15.30000%;height:6.00000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.180cqi;font-weight:600;line-height:1.32;text-align:left;color:#173b69;\">Cold Experts (\ub098\uba38\uc9c0)<\/div><div class=\"moe-ov-el\" data-text=\"\u2022 \uc120\ud0dd\ub41c \uc804\ubb38\uac00\ub97c CPU\uc5d0\uc11c \uacc4\uc0b0 \u2022 INT4 \uc804\ubb38\uac00 \uac00\uc911\uce58 \u2022 CPU DRAM 2 TB\" style=\"left:82.30000%;top:80.00000%;width:15.40000%;height:11.75000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.080cqi;font-weight:400;line-height:1.32;text-align:left;color:#335167;\">\u2022 \uc120\ud0dd\ub41c \uc804\ubb38\uac00\ub97c CPU\uc5d0\uc11c \uacc4\uc0b0<br>\u2022 INT4 \uc804\ubb38\uac00 \uac00\uc911\uce58<br>\u2022 CPU DRAM 2 TB<\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:11.6000%;top:48.7500%;width:4.7000%;height:0.300cqi;transform:translateY(-50%) rotate(0.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:16.3000%;top:61.5000%;width:4.7000%;height:0.300cqi;transform:translateY(-50%) rotate(180.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:26.5000%;top:51.7500%;width:0.8000%;height:0.200cqi;transform:translateY(-50%) rotate(0.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:40.0000%;top:55.0000%;width:3.0000%;height:0.300cqi;transform:translateY(-50%) rotate(0.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:50.6000%;top:51.0000%;width:4.8836%;height:0.260cqi;transform:translateY(-50%) rotate(-10.6197deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:50.6000%;top:56.5000%;width:4.9729%;height:0.260cqi;transform:translateY(-50%) rotate(15.1541deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:70.1000%;top:38.2500%;width:2.8000%;height:0.280cqi;transform:translateY(-50%) rotate(0.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:81.0000%;top:51.2500%;width:2.8000%;height:0.280cqi;transform:translateY(-50%) rotate(90.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:82.9000%;top:58.2500%;width:2.8000%;height:0.280cqi;transform:translateY(-50%) rotate(-90.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:22.0000%;top:75.7500%;width:2.4000%;height:0.200cqi;transform:translateY(-50%) rotate(-90.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:33.7000%;top:75.7500%;width:2.4000%;height:0.200cqi;transform:translateY(-50%) rotate(-90.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-ov-arrow\" style=\"left:46.7000%;top:70.7500%;width:2.0000%;height:0.200cqi;transform:translateY(-50%) rotate(90.0000deg);--moe-ac:#194c8c;--moe-al:#bce9ff;\"><span class=\"moe-ov-shaft\"><\/span><span class=\"moe-ov-arrowhead\"><\/span><\/div><div class=\"moe-ov-el\" data-text=\"PCIe \/ NUMA\" style=\"left:84.90000%;top:51.25000%;width:12.40000%;height:6.00000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.150cqi;font-weight:500;line-height:1.32;text-align:left;color:#173b69;\">PCIe \/ NUMA<\/div><div class=\"moe-ov-el\" data-text=\"\ubd80\ubd84 \uc0c1\uc8fc \uac1c\ub150\ub3c4 \u00b7 KV \uc815\ubc00\ub3c4\uc640 GPU \uc804\ubb38\uac00 \uc218\ub294 \uc124\uc815\uc5d0 \ub530\ub77c \ub2ec\ub77c\uc9c4\ub2e4.\" style=\"left:16.90000%;top:91.00000%;width:52.80000%;height:5.75000%;display:flex;align-items:center;justify-content:flex-start;font-size:1.070cqi;font-weight:400;line-height:1.32;text-align:left;color:#61748a;\">\ubd80\ubd84 \uc0c1\uc8fc \uac1c\ub150\ub3c4 \u00b7 KV \uc815\ubc00\ub3c4\uc640 GPU \uc804\ubb38\uac00 \uc218\ub294 \uc124\uc815\uc5d0 \ub530\ub77c \ub2ec\ub77c\uc9c4\ub2e4.<\/div><\/div><\/div><\/div><\/details><\/figure><div class=\"moe-prose\">\n<h2 id=\"moe-concept\">1. \uc77c\ubd80\ub9cc \uacc4\uc0b0\ud574\ub3c4, \uc804\ubb38\uac00 \uac00\uc911\uce58\ub294 \ud544\uc694\ud558\ub2e4<\/h2>\n<p>MoE(Mixture of Experts)\ub294 \uc2e0\uacbd\ub9dd\uc758 \uc77c\ubd80 \uc5f0\uc0b0\uc744 \uc5ec\ub7ec <strong>\uc804\ubb38\uac00 \uc2e0\uacbd\ub9dd<\/strong>\uc73c\ub85c \ub098\ub204\uace0, <strong>Router<\/strong>\uac00 \ud1a0\ud070\uc758 \uc911\uac04 \ud45c\ud604\uc744 \ubcf4\uace0 \uc0ac\uc6a9\ud560 \uc804\ubb38\uac00\ub97c \uc120\ud0dd\ud558\ub294 \uad6c\uc870\ub2e4. \uc120\ud0dd\ub41c \uc804\ubb38\uac00\ub4e4\uc758 \ucd9c\ub825\uc740 \ub77c\uc6b0\ud305 \uac00\uc911\uce58\uc5d0 \ub530\ub77c \ud569\uccd0\uc9c4\ub2e4. \uc804\uccb4 \ud30c\ub77c\ubbf8\ud130 \uc218\uac00 \uac00\uc911\uce58 \ubcf4\uad00 \uacf5\uac04\uacfc \uad00\ub828\ub41c\ub2e4\uba74, \ud65c\uc131 \ud30c\ub77c\ubbf8\ud130 \uc218\ub294 \ud1a0\ud070 \ud558\ub098\uc758 \uacc4\uc0b0\ub7c9\uacfc \uad00\ub828\ub41c\ub2e4.<\/p>\n<p>\uc2e4\ud5d8\uc5d0 \uc0ac\uc6a9\ud55c <strong>Qwen3-Coder-480B-A35B-Instruct-FP8<\/strong>\uc740 \uce35\ub9c8\ub2e4 160\uac1c \uc804\ubb38\uac00 \uc911 \ud1a0\ud070\ubcc4\ub85c 8\uac1c\ub97c \uc120\ud0dd\ud55c\ub2e4. \ud55c \ud1a0\ud070\uc5d0\uc11c \uc0ac\uc6a9\ud558\uc9c0 \uc54a\uc740 \uc804\ubb38\uac00\ub3c4 \ub2e4\ub978 \ud1a0\ud070\uc5d0\uc11c\ub294 \ud544\uc694\ud558\ub2e4. \uc77c\ubd80\ub9cc \uacc4\uc0b0\ud55c\ub2e4\ub294 \uc774\uc720\ub85c \ub098\uba38\uc9c0 \uac00\uc911\uce58\ub97c \ubc84\ub9b4 \uc218\ub294 \uc5c6\ub2e4. <a aria-label=\"\ubaa8\ub378 \uad6c\uc131 \uadfc\uac70\" href=\"#moe-source-model\">[1]<\/a><\/p>\n<p><strong>Hot\uacfc Cold\ub294 \ubaa8\ub378\uc5d0 \uace0\uc815\ub41c \ub450 \uc885\ub958\uc758 \uc804\ubb38\uac00\uac00 \uc544\ub2c8\ub77c, \uc120\ud0dd \ube48\ub3c4\uc640 \ubc30\uce58 \uc815\ucc45\uc5d0 \ub530\ub978 \uad6c\ubd84<\/strong>\uc774\ub2e4. \uc790\uc8fc \uc120\ud0dd\ub418\ub294 \uc804\ubb38\uac00\ub97c GPU\uc5d0 \ub0a8\uae30\uace0 \ub098\uba38\uc9c0\ub97c CPU\uac00 \ub9e1\ub294\ub2e4. Cold expert\ub3c4 \uc120\ud0dd\ub418\uba74 \uacc4\uc0b0\ud574\uc57c \ud558\uba70, \uc911\uc694\ud558\uc9c0 \uc54a\uc740 \uc5f0\uc0b0\uc774\ub77c\ub294 \ub73b\uc740 \uc544\ub2c8\ub2e4. \ub610\ud55c GPU\uc5d0 \uc804\ubb38\uac00 96\uac1c\ub97c \ub454\ub2e4\uace0 \ud574\uc11c \ud1a0\ud070\ubcc4 8\uac1c \uc120\ud0dd \uc911 \uc77c\uc815\ud55c \uc218\uac00 \ubc18\ub4dc\uc2dc GPU\ub098 CPU\uc5d0 \ubc30\uc815\ub418\ub294 \uac83\uc740 \uc544\ub2c8\ub2e4.<\/p>\n<p>GPU\uc758 \uace0\ub300\uc5ed\ud3ed \uba54\ubaa8\ub9ac\uc778 <strong>HBM<\/strong>\uc5d0\ub294 \uac00\uc911\uce58\ubfd0 \uc544\ub2c8\ub77c <strong>KV cache<\/strong>\ub3c4 \ub4e4\uc5b4\uac04\ub2e4. KV cache\ub294 \uc774\ubbf8 \ucc98\ub9ac\ud55c \ud1a0\ud070\uc758 Attention \uc0c1\ud0dc\ub97c \ubcf4\uad00\ud558\ub294 \uacf5\uac04\uc774\ub2e4. \ubb38\ub9e5\uc774 \uae38\uc5b4\uc9c0\uac70\ub098 \ub3d9\uc2dc \uc694\uccad\uc774 \ub298\uba74 \uce90\uc2dc\uc640 \uc2e4\ud589 \ubc84\ud37c\uc758 \uacf5\uac04 \uc694\uad6c\ub3c4 \ucee4\uc9c4\ub2e4.<\/p>\n<div class=\"moe-formula\"><strong>GPU \uba54\ubaa8\ub9ac \uc608\uc0b0<\/strong><br>Hot expert \uac00\uc911\uce58 + Attention\u00b7Router \ub4f1 \ube44\uc804\ubb38\uac00 \uac00\uc911\uce58 + KV cache + \ud65c\uc131\uac12\u00b7\uc2e4\ud589 \ubc84\ud37c<\/div>\n<p>\ub530\ub77c\uc11c Hot expert\ub97c \uac00\ub2a5\ud55c \ud55c \ub9ce\uc774 \uc62c\ub9ac\ub294 \uac83\ub9cc\uc73c\ub85c\ub294 \ucda9\ubd84\ud558\uc9c0 \uc54a\ub2e4. GPU\uc5d0\uc11c \uc9c1\uc811 \uacc4\uc0b0\ud560 \uc804\ubb38\uac00\uc640 \uc694\uccad\uc744 \uc720\uc9c0\ud560 \uce90\uc2dc \uc0ac\uc774\uc5d0 \uacf5\uac04\uc744 \ub098\ub204\uc5b4\uc57c \ud55c\ub2e4. \uc804\ubb38\uac00 \ubc30\uce58\uc640 \uc11c\ube44\uc2a4\uac00 \uc694\uad6c\ud558\ub294 \ubb38\ub9e5 \uae38\uc774\u00b7\ub3d9\uc2dc\uc131\uc744 \ud568\uaed8 \uacb0\uc815\ud574\uc57c \ud558\ub294 \uc774\uc720\ub2e4.<\/p>\n<h2 id=\"moe-offloading-types\">2. CPU\uc5d0 \ubcf4\uad00\ud558\ub294 \uac83\uacfc CPU\uac00 \uacc4\uc0b0\ud558\ub294 \uac83\uc740 \ub2e4\ub974\ub2e4<\/h2>\n<p>CPU offloading\uc774\ub77c\ub294 \uc774\ub984 \uc544\ub798\uc5d0\ub3c4 \uc11c\ub85c \ub2e4\ub978 \uc2e4\ud589 \ubc29\uc2dd\uc774 \uc788\ub2e4. \uac00\uc7a5 \uc911\uc694\ud55c \ucc28\uc774\ub294 <strong>\uc804\ubb38\uac00 \uac00\uc911\uce58\ub97c GPU\ub85c \uac00\uc838\uc640 \uacc4\uc0b0\ud558\ub294\uc9c0, CPU\uac00 \uc790\uc2e0\uc758 \uba54\ubaa8\ub9ac\ub97c \uc77d\uc5b4 \uc9c1\uc811 \uacc4\uc0b0\ud558\ub294\uc9c0<\/strong>\ub2e4.<\/p>\n<div class=\"moe-table-wrap\"><table class=\"moe-table\"><thead><tr><th style=\"width:28%\">\ubc29\uc2dd<\/th><th style=\"width:36%\">\uc804\ubb38\uac00 \uacc4\uc0b0 \uc8fc\uccb4<\/th><th>CPU\u2013GPU \uac04 \uc8fc\uc694 \uc804\ub2ec \ub370\uc774\ud130<\/th><\/tr><\/thead><tbody>\n<tr><td>GPU \uc804\uc6a9 \uc2e4\ud589<\/td><td>GPU\uac00 \ubaa8\ub4e0 \uc120\ud0dd \uc804\ubb38\uac00\ub97c \uacc4\uc0b0<\/td><td>CPU \uc804\ubb38\uac00 \uc5f0\uc0b0\uc744 \uc704\ud55c \uc655\ubcf5 \uc5c6\uc74c<\/td><\/tr>\n<tr><td>CPU \uc800\uc7a5\u00b7GPU \uacc4\uc0b0<\/td><td>\ud544\uc694\ud55c \uac00\uc911\uce58\ub97c GPU\ub85c \uac00\uc838\uc628 \ub4a4 \uacc4\uc0b0<\/td><td>\uc804\ubb38\uac00 \uac00\uc911\uce58<\/td><\/tr>\n<tr><td><strong>Hot\u00b7Cold \ubd80\ubd84 \uc0c1\uc8fc<\/strong><\/td><td><strong>GPU\ub294 Hot, CPU\ub294 Cold\ub97c \uacc4\uc0b0<\/strong><\/td><td><strong>Cold \uacbd\ub85c\uc758 \ud65c\uc131\uac12\u00b7\uc120\ud0dd \uc815\ubcf4\u00b7\uc5f0\uc0b0 \uacb0\uacfc<\/strong><\/td><\/tr>\n<\/tbody><\/table><\/div>\n<p>\ubd80\ubd84 \uc0c1\uc8fc \ubc29\uc2dd\uc5d0\uc11c\ub294 GPU\uc758 Hot expert\uc640 CPU\uc758 Cold expert\uac00 \uc11c\ub85c \ub2e4\ub978 \uc785\ub825 \uc9d1\ud569\uc744 \ucc98\ub9ac\ud55c\ub2e4. CPU\uc5d0 \ud544\uc694\ud55c \ud65c\uc131\uac12, \uc804\ubb38\uac00 ID, \ub77c\uc6b0\ud305 \uac00\uc911\uce58\ub97c \uc804\ub2ec\ud558\uace0 \uc5f0\uc0b0 \uacb0\uacfc\ub97c \ubc1b\ub294\ub2e4. <strong>Cold expert\uc758 \uc804\uccb4 \uac00\uc911\uce58\ub97c \ud1a0\ud070\ub9c8\ub2e4 GPU\ub85c \uc62e\uae30\uc9c0 \uc54a\ub294\ub2e4.<\/strong> \ub300\uc2e0 CPU\uc758 \uc5f0\uc0b0 \uc18d\ub3c4\uc640 \uba54\ubaa8\ub9ac \ub300\uc5ed\ud3ed, \uacb0\uacfc\ub97c \ub3cc\ub824\uc8fc\ub294 \uc9c0\uc5f0\uc774 \uc911\uc694\ud574\uc9c4\ub2e4.<\/p>\n<h2 id=\"moe-execution\">3. \uac19\uc740 \uc11c\ubc84\uc5d0\uc11c \ub098\ub258\uace0 \ub2e4\uc2dc \ud569\uccd0\uc9c0\ub294 \uc2e4\ud589<\/h2>\n<p>\uc544\ub798 \uad6c\uc870\ub294 CPU 2\uc18c\ucf13\uacfc GPU 4\uc7a5\uc774 \ud558\ub098\uc758 \ubaa8\ub378\uc744 \uc2e4\ud589\ud558\ub294 \uad6c\uc131\uc774\ub2e4. Hot expert\uc640 Attention\u00b7Router\u00b7KV cache\ub294 GPU\uc5d0, Cold expert \uac00\uc911\uce58\ub294 Host Memory\uc5d0 \ub454\ub2e4. \uadf8\ub9bc\uc758 <strong>TP4\ub294 \ud150\uc11c \ubcd1\ub82c\ub85c \ubaa8\ub378\uc758 \ud589\ub82c \uc5f0\uc0b0\uc744 GPU 4\uc7a5\uc5d0 \ub098\ub208\ub2e4\ub294 \ub73b<\/strong>\uc774\ub2e4. \uc804\ubb38\uac00\ub97c CPU\uc640 GPU \uc911 \uc5b4\ub514\uc5d0 \ub458\uc9c0 \uacb0\uc815\ud558\ub294 \ubc30\uce58 \uc815\ucc45\uacfc\ub294 \ubcc4\uac1c\uc758 \ucd95\uc774\ub2e4.<\/p>\n<\/div><figure class=\"moe-figure\"><div aria-label=\"\ub2e8\uc77c \uc11c\ubc84 CPU MoE \ubd80\ubd84 \uc0c1\uc8fc \uad6c\uc870. Hot expert\ub294 GPU HBM\uc5d0 \uc0c1\uc8fc\ud558\uace0 GPU\uc5d0\uc11c \uc2e4\ud589\ud55c\ub2e4. Cold expert\ub294 Host Memory\uc5d0 \uc0c1\uc8fc\ud558\uace0 CPU\uc5d0\uc11c \uc2e4\ud589\ud55c\ub2e4. Router\uac00 \uc120\ud0dd\ud55c \uc804\ubb38\uac00\uc758 \ubc30\uce58\uc5d0 \ub530\ub77c \uacbd\ub85c\uac00 \ub098\ub258\uba70 \ub450 \uacb0\uacfc\ub294 GPU\uc5d0\uc11c \uacb0\ud569\ub41c\ub2e4.\" class=\"moe-diagram moe-detailed moe-hc-diagram\" id=\"moe-full-canvas\" role=\"group\" style=\"aspect-ratio:10 \/ 7;overflow:hidden;font-family:Arial,'Noto Sans CJK KR','Malgun Gothic',sans-serif;color:#20252B;line-height:1.24;\">\n<div class=\"moe-hc-label\" data-label=\"CPU MoE Offloading | Hot\uc740 GPU, Cold\ub294 Host Memory\" style=\"left:2.500000%;top:1.607143%;width:94.375000%;height:4.910714%;font-size:2.25000cqi;font-weight:700;color:#20252B;justify-content:flex-start;text-align:left;\"><span>CPU MoE Offloading | Hot\uc740 GPU, Cold\ub294 Host Memory<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"\ub2e8\uc77c \uc11c\ubc84\uc758 \uac00\uc911\uce58 \ubc30\uce58\uc640 \uc2e4\ud589 \uacbd\ub85c \u00b7 Qwen3-Coder-480B \u00b7 TP4 \ubd80\ubd84 \uc0c1\uc8fc \uc608\uc2dc\" style=\"left:2.625000%;top:7.053571%;width:93.750000%;height:3.125000%;font-size:1.43750cqi;font-weight:400;color:#697581;justify-content:flex-start;text-align:left;\"><span>\ub2e8\uc77c \uc11c\ubc84\uc758 \uac00\uc911\uce58 \ubc30\uce58\uc640 \uc2e4\ud589 \uacbd\ub85c \u00b7 Qwen3-Coder-480B \u00b7 TP4 \ubd80\ubd84 \uc0c1\uc8fc \uc608\uc2dc<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:2.000000%;top:11.785714%;width:96.000000%;height:83.928571%;box-sizing:border-box;background:#FFFFFF;z-index:1;border:0.08750cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:2.000000%;top:11.785714%;width:96.000000%;height:4.464286%;box-sizing:border-box;background:#FAFAFA;z-index:1;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:2.000000%;top:11.785714%;width:0.437500%;height:4.464286%;box-sizing:border-box;background:#538134;z-index:1;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\ub2e8\uc77c \ubb3c\ub9ac \uc11c\ubc84 \u00b7 CPU\uc640 GPU\uac00 \ud568\uaed8 \uc2e4\ud589\" style=\"left:3.250000%;top:12.321429%;width:43.750000%;height:3.482143%;font-size:1.75000cqi;font-weight:700;color:#20252B;justify-content:flex-start;text-align:left;\"><span>\ub2e8\uc77c \ubb3c\ub9ac \uc11c\ubc84 \u00b7 CPU\uc640 GPU\uac00 \ud568\uaed8 \uc2e4\ud589<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"Xeon 8480+ \u00d72 \u00b7 Host DRAM 2 TB \u00b7 H100 80 GB \u00d74\" style=\"left:54.250000%;top:12.678571%;width:41.750000%;height:2.946429%;font-size:1.43750cqi;font-weight:400;color:#68737E;justify-content:flex-end;text-align:right;\"><span>Xeon 8480+ \u00d72 \u00b7 Host DRAM 2 TB \u00b7 H100 80 GB \u00d74<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"runtime\" style=\"position:absolute;left:4.500000%;top:18.214286%;width:91.000000%;height:5.000000%;box-sizing:border-box;background:#FFFFFF;z-index:1;border:0.08750cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:4.500000%;top:18.214286%;width:91.000000%;height:3.928571%;box-sizing:border-box;background:#FAFAFA;z-index:1;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:4.500000%;top:18.214286%;width:0.375000%;height:3.928571%;box-sizing:border-box;background:#FF7E8A;z-index:1;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"SGLang \uc11c\ube59 \ub7f0\ud0c0\uc784 | \uc694\uccad \uc2a4\ucf00\uc904\ub9c1 \u00b7 TP4 GPU \uc6cc\ucee4 \u00b7 KT CPUInfer \uc5f0\ub3d9\" style=\"left:5.500000%;top:18.482143%;width:89.125000%;height:3.303571%;font-size:1.62500cqi;font-weight:700;color:#20252B;justify-content:flex-start;text-align:left;\"><span>SGLang \uc11c\ube59 \ub7f0\ud0c0\uc784 | \uc694\uccad \uc2a4\ucf00\uc904\ub9c1 \u00b7 TP4 GPU \uc6cc\ucee4 \u00b7 KT CPUInfer \uc5f0\ub3d9<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"model-files\" style=\"position:absolute;left:4.500000%;top:26.785714%;width:17.500000%;height:23.214286%;box-sizing:border-box;background:#FFFFFF;z-index:1;border:0.08750cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:4.500000%;top:26.785714%;width:17.500000%;height:3.928571%;box-sizing:border-box;background:#FAFAFA;z-index:1;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:4.500000%;top:26.785714%;width:0.375000%;height:3.928571%;box-sizing:border-box;background:#737373;z-index:1;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\ubaa8\ub378 \ud30c\uc77c \u00b7 \uae30\ub3d9 \uc2dc \uc801\uc7ac\" style=\"left:5.500000%;top:27.053571%;width:15.625000%;height:3.303571%;font-size:1.43750cqi;font-weight:700;color:#20252B;justify-content:flex-start;text-align:left;\"><span>\ubaa8\ub378 \ud30c\uc77c \u00b7 \uae30\ub3d9 \uc2dc \uc801\uc7ac<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"fp8-checkpoint\" style=\"position:absolute;left:5.750000%;top:32.589286%;width:15.000000%;height:5.625000%;box-sizing:border-box;background:#F1E9F8;z-index:3;border:0.09375cqi solid #C78AFF;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\uc6d0\ubcf8 FP8 checkpoint\nmodel_path\" style=\"left:6.250000%;top:32.946429%;width:14.000000%;height:4.910714%;font-size:1.37500cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>\uc6d0\ubcf8 FP8 checkpoint<br>model_path<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"amxint4-checkpoint\" style=\"position:absolute;left:5.750000%;top:42.321429%;width:15.000000%;height:5.625000%;box-sizing:border-box;background:#F1E9F8;z-index:3;border:0.09375cqi solid #C78AFF;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"CPU\uc6a9 INT4 \ubcc0\ud658\ubcf8\nkt_weight_path\" style=\"left:6.250000%;top:42.678571%;width:14.000000%;height:4.910714%;font-size:1.37500cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>CPU\uc6a9 INT4 \ubcc0\ud658\ubcf8<br>kt_weight_path<\/span><\/div>\n<!-- FP8 -> INT4 offline conversion -->\n<div class=\"moe-hc-label\" data-label=\"kt quant \u00b7 INT4\" style=\"left:13.875000%;top:38.660714%;width:7.250000%;height:3.214286%;font-size:1.06250cqi;font-weight:400;color:#737373;justify-content:flex-start;text-align:left;\"><span>kt quant \u00b7 INT4<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"host-memory\" style=\"position:absolute;left:24.500000%;top:26.785714%;width:31.250000%;height:23.214286%;box-sizing:border-box;background:#FFFFFF;z-index:1;border:0.08750cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:24.500000%;top:26.785714%;width:31.250000%;height:3.928571%;box-sizing:border-box;background:#E2F0D9;z-index:1;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:24.500000%;top:26.785714%;width:0.375000%;height:3.928571%;box-sizing:border-box;background:#538134;z-index:1;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Host Memory \u00b7 DDR5 2 TB\" style=\"left:25.500000%;top:27.053571%;width:29.375000%;height:3.303571%;font-size:1.62500cqi;font-weight:700;color:#20252B;justify-content:flex-start;text-align:left;\"><span>Host Memory \u00b7 DDR5 2 TB<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"cold-expert-weights\" style=\"position:absolute;left:25.750000%;top:32.589286%;width:28.750000%;height:13.571429%;box-sizing:border-box;background:#F5FAFF;z-index:3;border:0.09375cqi solid #8DBBF3;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\" style=\"left:26.250000%;top:32.946429%;width:27.750000%;height:12.857143%;font-size:1.43750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span><\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"Cold expert \uac00\uc911\uce58 \u00b7 INT4\" style=\"left:26.375000%;top:33.303571%;width:27.500000%;height:3.482143%;font-size:1.75000cqi;font-weight:700;color:#1C5997;justify-content:center;text-align:center;\"><span>Cold expert \uac00\uc911\uce58 \u00b7 INT4<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:26.687500%;top:38.303571%;width:5.875000%;height:3.392857%;box-sizing:border-box;background:#EAF3FF;z-index:3;border:0.09375cqi solid #79ADE9;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Cold E\u2090\" style=\"left:27.187500%;top:38.392857%;width:4.875000%;height:3.214286%;font-size:1.25000cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>Cold E\u2090<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:33.437500%;top:38.303571%;width:5.875000%;height:3.392857%;box-sizing:border-box;background:#EAF3FF;z-index:3;border:0.09375cqi solid #79ADE9;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Cold E\u1d66\" style=\"left:33.937500%;top:38.392857%;width:4.875000%;height:3.214286%;font-size:1.25000cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>Cold E\u1d66<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:40.187500%;top:38.303571%;width:5.875000%;height:3.392857%;box-sizing:border-box;background:#EAF3FF;z-index:3;border:0.09375cqi solid #79ADE9;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\u2026\" style=\"left:40.687500%;top:38.392857%;width:4.875000%;height:3.214286%;font-size:1.25000cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>\u2026<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:46.937500%;top:38.303571%;width:5.875000%;height:3.392857%;box-sizing:border-box;background:#EAF3FF;z-index:3;border:0.09375cqi solid #79ADE9;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Cold E\u2099\" style=\"left:47.437500%;top:38.392857%;width:4.875000%;height:3.214286%;font-size:1.25000cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>Cold E\u2099<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"\ubc30\uce58\ud45c\uc5d0\uc11c CPU\ub85c \uc9c0\uc815\ud55c \uc804\ubb38\uac00 \uc9d1\ud569\" style=\"left:26.437500%;top:42.678571%;width:27.375000%;height:2.589286%;font-size:1.31250cqi;font-weight:400;color:#56728C;justify-content:center;text-align:center;\"><span>\ubc30\uce58\ud45c\uc5d0\uc11c CPU\ub85c \uc9c0\uc815\ud55c \uc804\ubb38\uac00 \uc9d1\ud569<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"\uac00\uc911\uce58\ub294 Host Memory\uc5d0 \uc0c1\uc8fc\" style=\"left:25.687500%;top:47.410714%;width:29.062500%;height:2.500000%;font-size:1.37500cqi;font-weight:700;color:#347543;justify-content:center;text-align:center;\"><span>\uac00\uc911\uce58\ub294 Host Memory\uc5d0 \uc0c1\uc8fc<\/span><\/div>\n<!-- INT4 load into host memory -->\n<div aria-hidden=\"true\" class=\"\" data-component=\"cpu-compute\" style=\"position:absolute;left:61.250000%;top:26.785714%;width:34.250000%;height:23.214286%;box-sizing:border-box;background:#FFFFFF;z-index:1;border:0.08750cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:61.250000%;top:26.785714%;width:34.250000%;height:3.928571%;box-sizing:border-box;background:#E2F0D9;z-index:1;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:61.250000%;top:26.785714%;width:0.375000%;height:3.928571%;box-sizing:border-box;background:#538134;z-index:1;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"CPU \u00b7 Xeon 8480+ \u00d72\" style=\"left:62.250000%;top:27.053571%;width:32.375000%;height:3.303571%;font-size:1.62500cqi;font-weight:700;color:#20252B;justify-content:flex-start;text-align:left;\"><span>CPU \u00b7 Xeon 8480+ \u00d72<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-stage moe-phase-cpu\" data-component=\"kt-cpuinfer\" style=\"position:absolute;left:62.500000%;top:32.589286%;width:31.750000%;height:9.107143%;box-sizing:border-box;background:#F8FCF8;z-index:3;border:0.09375cqi solid #629A6C;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\" style=\"left:63.000000%;top:32.946429%;width:30.750000%;height:8.392857%;font-size:1.43750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span><\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"\u2462-C  Cold expert \uacc4\uc0b0 \u00b7 KT CPUInfer\" style=\"left:63.125000%;top:33.214286%;width:30.500000%;height:3.214286%;font-size:1.56250cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>\u2462-C  Cold expert \uacc4\uc0b0 \u00b7 KT CPUInfer<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"CPU\uac00 Host Memory\uc758 \uac00\uc911\uce58\ub97c \uc77d\uc5b4 \uc5f0\uc0b0\" style=\"left:63.125000%;top:36.875000%;width:30.500000%;height:2.678571%;font-size:1.31250cqi;font-weight:400;color:#20252B;justify-content:center;text-align:center;\"><span>CPU\uac00 Host Memory\uc758 \uac00\uc911\uce58\ub97c \uc77d\uc5b4 \uc5f0\uc0b0<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-stage moe-phase-worker\" data-component=\"numa-0\" style=\"position:absolute;left:62.500000%;top:43.928571%;width:15.000000%;height:3.660714%;box-sizing:border-box;background:#FFFFFF;z-index:3;border:0.09375cqi solid #BEBEBE;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"NUMA 0 \u00b7 48 workers\" style=\"left:63.000000%;top:44.017857%;width:14.000000%;height:3.482143%;font-size:1.31250cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>NUMA 0 \u00b7 48 workers<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-stage moe-phase-worker\" data-component=\"numa-1\" style=\"position:absolute;left:79.250000%;top:43.928571%;width:15.000000%;height:3.660714%;box-sizing:border-box;background:#FFFFFF;z-index:3;border:0.09375cqi solid #BEBEBE;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"NUMA 1 \u00b7 48 workers\" style=\"left:79.750000%;top:44.017857%;width:14.000000%;height:3.482143%;font-size:1.31250cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>NUMA 1 \u00b7 48 workers<\/span><\/div>\n<!-- Cold weights: Host Memory -> CPU compute (local read) -->\n<div class=\"moe-hc-label\" data-label=\"\uac00\uc911\uce58\n\uc77d\uae30\" style=\"left:56.062500%;top:30.625000%;width:4.812500%;height:4.107143%;font-size:1.18750cqi;font-weight:400;color:#737373;justify-content:center;text-align:center;\"><span>\uac00\uc911\uce58<br>\uc77d\uae30<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"gpu-bank\" style=\"position:absolute;left:4.500000%;top:57.500000%;width:91.000000%;height:34.285714%;box-sizing:border-box;background:#FFFFFF;z-index:1;border:0.08750cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:4.500000%;top:57.500000%;width:91.000000%;height:3.928571%;box-sizing:border-box;background:#FFF2CC;z-index:1;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:4.500000%;top:57.500000%;width:0.375000%;height:3.928571%;box-sizing:border-box;background:#BF9000;z-index:1;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"GPU \u00b7 H100 80 GB \u00d74 \u00b7 TP4\" style=\"left:5.500000%;top:57.767857%;width:89.125000%;height:3.303571%;font-size:1.68750cqi;font-weight:700;color:#20252B;justify-content:flex-start;text-align:left;\"><span>GPU \u00b7 H100 80 GB \u00d74 \u00b7 TP4<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"Hot expert + Attention \/ Router + KV cache\" style=\"left:52.812500%;top:57.857143%;width:41.437500%;height:3.125000%;font-size:1.43750cqi;font-weight:700;color:#986B00;justify-content:flex-end;text-align:right;\"><span>Hot expert + Attention \/ Router + KV cache<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"input-hidden-state\" style=\"position:absolute;left:6.062500%;top:63.392857%;width:8.687500%;height:7.142857%;box-sizing:border-box;background:#FFFFFF;z-index:3;border:0.09375cqi solid #BEBEBE;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\uc785\ub825\nhidden state\" style=\"left:6.562500%;top:63.750000%;width:7.687500%;height:6.428571%;font-size:1.25000cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>\uc785\ub825<br>hidden state<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-stage moe-phase-attn\" data-component=\"attention\" style=\"position:absolute;left:16.812500%;top:63.392857%;width:12.750000%;height:7.142857%;box-sizing:border-box;background:#F1E9F8;z-index:3;border:0.09375cqi solid #C78AFF;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\u2460 Attention\nMoE \uc785\ub825 \uc900\ube44\" style=\"left:17.312500%;top:63.750000%;width:11.750000%;height:6.428571%;font-size:1.56250cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>\u2460 Attention<br>MoE \uc785\ub825 \uc900\ube44<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-stage moe-phase-route\" data-component=\"router-dispatch\" style=\"position:absolute;left:33.125000%;top:63.392857%;width:15.125000%;height:7.142857%;box-sizing:border-box;background:#F1E9F8;z-index:3;border:0.09375cqi solid #C78AFF;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\u2461 Router \/ dispatch\ntop-8 \u2192 Hot \/ Cold \ubd84\uae30\" style=\"left:33.625000%;top:63.750000%;width:14.125000%;height:6.428571%;font-size:1.25000cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>\u2461 Router \/ dispatch<br>top-8 \u2192 Hot \/ Cold \ubd84\uae30<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-stage moe-phase-hot\" data-component=\"gpu-hot-compute\" style=\"position:absolute;left:52.562500%;top:63.392857%;width:14.687500%;height:7.142857%;box-sizing:border-box;background:#E9F7E8;z-index:3;border:0.09375cqi solid #4B9B58;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\u2462-G  Hot expert\nGPU \uc5f0\uc0b0 \u00b7 FP8\" style=\"left:53.062500%;top:63.750000%;width:13.687500%;height:6.428571%;font-size:1.50000cqi;font-weight:700;color:#236638;justify-content:center;text-align:center;\"><span>\u2462-G  Hot expert<br>GPU \uc5f0\uc0b0 \u00b7 FP8<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-stage moe-phase-combine\" data-component=\"result-combine\" style=\"position:absolute;left:75.625000%;top:63.392857%;width:18.000000%;height:7.142857%;box-sizing:border-box;background:#E2F0D9;z-index:3;border:0.09375cqi solid #538134;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\u2463 Hot + Cold \uacb0\uacfc \uacb0\ud569\n\uc794\ucc28 \uc5f0\uacb0 \u2192 \ub2e4\uc74c \uce35\" style=\"left:76.125000%;top:63.750000%;width:17.000000%;height:6.428571%;font-size:1.50000cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>\u2463 Hot + Cold \uacb0\uacfc \uacb0\ud569<br>\uc794\ucc28 \uc5f0\uacb0 \u2192 \ub2e4\uc74c \uce35<\/span><\/div>\n<!-- Input -> GPU Attention -->\n<!-- MoE input activation -> Router \/ dispatch -->\n<div class=\"moe-hc-label\" data-label=\"x\" style=\"left:29.812500%;top:63.839286%;width:3.062500%;height:2.232143%;font-size:1.25000cqi;font-weight:400;color:#68737E;justify-content:center;text-align:center;\"><span>x<\/span><\/div>\n<!-- Selected Hot expert input -> GPU computation -->\n<div class=\"moe-hc-label\" data-label=\"Hot\" style=\"left:48.562500%;top:63.571429%;width:3.625000%;height:2.232143%;font-size:1.18750cqi;font-weight:700;color:#347E42;justify-content:center;text-align:center;\"><span>Hot<\/span><\/div>\n<!-- GPU Hot result -> combine -->\n<div class=\"moe-hc-label\" data-label=\"Hot \uacb0\uacfc\" style=\"left:67.812500%;top:63.750000%;width:7.312500%;height:2.410714%;font-size:1.25000cqi;font-weight:700;color:#347E42;justify-content:center;text-align:center;\"><span>Hot \uacb0\uacfc<\/span><\/div>\n<!-- Cold activation + selected IDs \/ routing weights -> CPU -->\n<div class=\"moe-hc-label\" data-label=\"Cold \uacbd\ub85c: \ud65c\uc131\uac12 + \uc120\ud0dd ID \u00b7 \ub77c\uc6b0\ud305 \uac00\uc911\uce58\" style=\"left:41.562500%;top:50.446429%;width:35.000000%;height:2.857143%;font-size:1.50000cqi;font-weight:700;color:#25869B;justify-content:flex-start;text-align:left;\"><span>Cold \uacbd\ub85c: \ud65c\uc131\uac12 + \uc120\ud0dd ID \u00b7 \ub77c\uc6b0\ud305 \uac00\uc911\uce58<\/span><\/div>\n<!-- CPU Cold result -> GPU combine -->\n<div class=\"moe-hc-label\" data-label=\"Cold \uc5f0\uc0b0 \uacb0\uacfc \ubc18\ud658\" style=\"left:81.312500%;top:51.696429%;width:14.062500%;height:2.857143%;font-size:1.43750cqi;font-weight:700;color:#407D61;justify-content:flex-end;text-align:right;\"><span>Cold \uc5f0\uc0b0 \uacb0\uacfc \ubc18\ud658<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"\ud65c\uc131\uac12\uacfc \uacb0\uacfc\ub9cc \ub3d9\uc77c \uc11c\ubc84 \ub0b4\ubd80\uc5d0\uc11c \uc655\ubcf5\" style=\"left:45.000000%;top:54.285714%;width:30.000000%;height:2.410714%;font-size:1.12500cqi;font-weight:400;color:#73808A;justify-content:center;text-align:center;\"><span>\ud65c\uc131\uac12\uacfc \uacb0\uacfc\ub9cc \ub3d9\uc77c \uc11c\ubc84 \ub0b4\ubd80\uc5d0\uc11c \uc655\ubcf5<\/span><\/div>\n<!-- Residual input bypass -> residual connection -->\n<div class=\"moe-hc-label\" data-label=\"\uc794\ucc28 \uc785\ub825 \u2014 \uc804\ubb38\uac00 \uacc4\uc0b0 \uacbd\ub85c\uc640 \ubcc4\ub3c4\ub85c \uc804\ub2ec\" style=\"left:34.750000%;top:70.892857%;width:35.625000%;height:2.321429%;font-size:1.12500cqi;font-weight:400;color:#79838D;justify-content:center;text-align:center;\"><span>\uc794\ucc28 \uc785\ub825 \u2014 \uc804\ubb38\uac00 \uacc4\uc0b0 \uacbd\ub85c\uc640 \ubcc4\ub3c4\ub85c \uc804\ub2ec<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"\uac01 GPU HBM\uc5d0 \uc0c1\uc8fc\ud558\ub294 \ub0b4\uc6a9 \u00b7 TP4 \ubd84\ud560\" style=\"left:5.875000%;top:74.821429%;width:56.250000%;height:2.767857%;font-size:1.50000cqi;font-weight:700;color:#AD810B;justify-content:flex-start;text-align:left;\"><span>\uac01 GPU HBM\uc5d0 \uc0c1\uc8fc\ud558\ub294 \ub0b4\uc6a9 \u00b7 TP4 \ubd84\ud560<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"GPU \uac04 \uc5f0\uacb0: NVLink\" style=\"left:69.375000%;top:74.910714%;width:24.500000%;height:2.500000%;font-size:1.31250cqi;font-weight:400;color:#778390;justify-content:flex-end;text-align:right;\"><span>GPU \uac04 \uc5f0\uacb0: NVLink<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-hbm-frame\" data-component=\"gpu-0\" style=\"position:absolute;left:6.125000%;top:78.125000%;width:20.437500%;height:11.696429%;box-sizing:border-box;background:#FFFFFF;z-index:3;border:0.06250cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:6.187500%;top:78.214286%;width:20.312500%;height:2.500000%;box-sizing:border-box;background:#FFF2CC;z-index:3;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"GPU 0 \u00b7 rank 0\" style=\"left:6.625000%;top:78.303571%;width:19.437500%;height:2.410714%;font-size:1.31250cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>GPU 0 \u00b7 rank 0<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"gpu-0-hot-weights\" style=\"position:absolute;left:6.625000%;top:81.428571%;width:19.437500%;height:2.232143%;box-sizing:border-box;background:#E9F7E8;z-index:3;border:0.09375cqi solid #80BC83;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Hot expert TP \uc870\uac01 \u00b7 FP8\" style=\"left:7.125000%;top:81.517857%;width:18.437500%;height:2.053571%;font-size:1.25000cqi;font-weight:700;color:#246337;justify-content:center;text-align:center;\"><span>Hot expert TP \uc870\uac01 \u00b7 FP8<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"gpu-0-nonexpert-weights\" style=\"position:absolute;left:6.625000%;top:84.285714%;width:19.437500%;height:2.142857%;box-sizing:border-box;background:#F1E9F8;z-index:3;border:0.09375cqi solid #C78AFF;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Attention \ub4f1 \ube44\uc804\ubb38\uac00 TP \uc870\uac01\" style=\"left:7.125000%;top:84.375000%;width:18.437500%;height:1.964286%;font-size:1.18750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>Attention \ub4f1 \ube44\uc804\ubb38\uac00 TP \uc870\uac01<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-cache-access\" data-component=\"gpu-0-kv-cache\" style=\"position:absolute;left:6.625000%;top:87.053571%;width:19.437500%;height:2.142857%;box-sizing:border-box;background:#E7F5FE;z-index:3;border:0.09375cqi solid #6CBCDA;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"KV cache \u00b7 GPU HBM\" style=\"left:7.125000%;top:87.142857%;width:18.437500%;height:1.964286%;font-size:1.18750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>KV cache \u00b7 GPU HBM<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-hbm-frame\" data-component=\"gpu-1\" style=\"position:absolute;left:28.125000%;top:78.125000%;width:20.437500%;height:11.696429%;box-sizing:border-box;background:#FFFFFF;z-index:3;border:0.06250cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:28.187500%;top:78.214286%;width:20.312500%;height:2.500000%;box-sizing:border-box;background:#FFF2CC;z-index:3;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"GPU 1 \u00b7 rank 1\" style=\"left:28.625000%;top:78.303571%;width:19.437500%;height:2.410714%;font-size:1.31250cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>GPU 1 \u00b7 rank 1<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"gpu-1-hot-weights\" style=\"position:absolute;left:28.625000%;top:81.428571%;width:19.437500%;height:2.232143%;box-sizing:border-box;background:#E9F7E8;z-index:3;border:0.09375cqi solid #80BC83;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Hot expert TP \uc870\uac01 \u00b7 FP8\" style=\"left:29.125000%;top:81.517857%;width:18.437500%;height:2.053571%;font-size:1.25000cqi;font-weight:700;color:#246337;justify-content:center;text-align:center;\"><span>Hot expert TP \uc870\uac01 \u00b7 FP8<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"gpu-1-nonexpert-weights\" style=\"position:absolute;left:28.625000%;top:84.285714%;width:19.437500%;height:2.142857%;box-sizing:border-box;background:#F1E9F8;z-index:3;border:0.09375cqi solid #C78AFF;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Attention \ub4f1 \ube44\uc804\ubb38\uac00 TP \uc870\uac01\" style=\"left:29.125000%;top:84.375000%;width:18.437500%;height:1.964286%;font-size:1.18750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>Attention \ub4f1 \ube44\uc804\ubb38\uac00 TP \uc870\uac01<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-cache-access\" data-component=\"gpu-1-kv-cache\" style=\"position:absolute;left:28.625000%;top:87.053571%;width:19.437500%;height:2.142857%;box-sizing:border-box;background:#E7F5FE;z-index:3;border:0.09375cqi solid #6CBCDA;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"KV cache \u00b7 GPU HBM\" style=\"left:29.125000%;top:87.142857%;width:18.437500%;height:1.964286%;font-size:1.18750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>KV cache \u00b7 GPU HBM<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-hbm-frame\" data-component=\"gpu-2\" style=\"position:absolute;left:50.125000%;top:78.125000%;width:20.437500%;height:11.696429%;box-sizing:border-box;background:#FFFFFF;z-index:3;border:0.06250cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:50.187500%;top:78.214286%;width:20.312500%;height:2.500000%;box-sizing:border-box;background:#FFF2CC;z-index:3;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"GPU 2 \u00b7 rank 2\" style=\"left:50.625000%;top:78.303571%;width:19.437500%;height:2.410714%;font-size:1.31250cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>GPU 2 \u00b7 rank 2<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"gpu-2-hot-weights\" style=\"position:absolute;left:50.625000%;top:81.428571%;width:19.437500%;height:2.232143%;box-sizing:border-box;background:#E9F7E8;z-index:3;border:0.09375cqi solid #80BC83;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Hot expert TP \uc870\uac01 \u00b7 FP8\" style=\"left:51.125000%;top:81.517857%;width:18.437500%;height:2.053571%;font-size:1.25000cqi;font-weight:700;color:#246337;justify-content:center;text-align:center;\"><span>Hot expert TP \uc870\uac01 \u00b7 FP8<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"gpu-2-nonexpert-weights\" style=\"position:absolute;left:50.625000%;top:84.285714%;width:19.437500%;height:2.142857%;box-sizing:border-box;background:#F1E9F8;z-index:3;border:0.09375cqi solid #C78AFF;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Attention \ub4f1 \ube44\uc804\ubb38\uac00 TP \uc870\uac01\" style=\"left:51.125000%;top:84.375000%;width:18.437500%;height:1.964286%;font-size:1.18750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>Attention \ub4f1 \ube44\uc804\ubb38\uac00 TP \uc870\uac01<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-cache-access\" data-component=\"gpu-2-kv-cache\" style=\"position:absolute;left:50.625000%;top:87.053571%;width:19.437500%;height:2.142857%;box-sizing:border-box;background:#E7F5FE;z-index:3;border:0.09375cqi solid #6CBCDA;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"KV cache \u00b7 GPU HBM\" style=\"left:51.125000%;top:87.142857%;width:18.437500%;height:1.964286%;font-size:1.18750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>KV cache \u00b7 GPU HBM<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-hbm-frame\" data-component=\"gpu-3\" style=\"position:absolute;left:72.125000%;top:78.125000%;width:20.437500%;height:11.696429%;box-sizing:border-box;background:#FFFFFF;z-index:3;border:0.06250cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:72.187500%;top:78.214286%;width:20.312500%;height:2.500000%;box-sizing:border-box;background:#FFF2CC;z-index:3;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"GPU 3 \u00b7 rank 3\" style=\"left:72.625000%;top:78.303571%;width:19.437500%;height:2.410714%;font-size:1.31250cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>GPU 3 \u00b7 rank 3<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"gpu-3-hot-weights\" style=\"position:absolute;left:72.625000%;top:81.428571%;width:19.437500%;height:2.232143%;box-sizing:border-box;background:#E9F7E8;z-index:3;border:0.09375cqi solid #80BC83;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Hot expert TP \uc870\uac01 \u00b7 FP8\" style=\"left:73.125000%;top:81.517857%;width:18.437500%;height:2.053571%;font-size:1.25000cqi;font-weight:700;color:#246337;justify-content:center;text-align:center;\"><span>Hot expert TP \uc870\uac01 \u00b7 FP8<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"gpu-3-nonexpert-weights\" style=\"position:absolute;left:72.625000%;top:84.285714%;width:19.437500%;height:2.142857%;box-sizing:border-box;background:#F1E9F8;z-index:3;border:0.09375cqi solid #C78AFF;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Attention \ub4f1 \ube44\uc804\ubb38\uac00 TP \uc870\uac01\" style=\"left:73.125000%;top:84.375000%;width:18.437500%;height:1.964286%;font-size:1.18750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>Attention \ub4f1 \ube44\uc804\ubb38\uac00 TP \uc870\uac01<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-cache-access\" data-component=\"gpu-3-kv-cache\" style=\"position:absolute;left:72.625000%;top:87.053571%;width:19.437500%;height:2.142857%;box-sizing:border-box;background:#E7F5FE;z-index:3;border:0.09375cqi solid #6CBCDA;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"KV cache \u00b7 GPU HBM\" style=\"left:73.125000%;top:87.142857%;width:18.437500%;height:1.964286%;font-size:1.18750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>KV cache \u00b7 GPU HBM<\/span><\/div>\n<!-- FP8 Hot\/non-expert checkpoint load -> GPU HBM -->\n<div class=\"moe-hc-label\" data-label=\"\uc810\uc120: \uac00\uc911\uce58 \uc801\uc7ac\u00b7\ucc38\uc870 \/ \uc794\ucc28 \uc785\ub825\" style=\"left:5.250000%;top:51.607143%;width:32.625000%;height:2.321429%;font-size:1.25000cqi;font-weight:400;color:#76818B;justify-content:flex-start;text-align:left;\"><span>\uc810\uc120: \uac00\uc911\uce58 \uc801\uc7ac\u00b7\ucc38\uc870 \/ \uc794\ucc28 \uc785\ub825<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"\uc0c9 \uc2e4\uc120: \ud1a0\ud070 \ub370\uc774\ud130\uc640 \uc804\ubb38\uac00 \uc5f0\uc0b0 \uacb0\uacfc\" style=\"left:5.250000%;top:54.196429%;width:32.625000%;height:2.321429%;font-size:1.25000cqi;font-weight:400;color:#76818B;justify-content:flex-start;text-align:left;\"><span>\uc0c9 \uc2e4\uc120: \ud1a0\ud070 \ub370\uc774\ud130\uc640 \uc804\ubb38\uac00 \uc5f0\uc0b0 \uacb0\uacfc<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"Hot expert \uac00\uc911\uce58\ub294 GPU HBM\uc5d0, Cold expert \uac00\uc911\uce58\ub294 Host Memory\uc5d0 \ub454\ub2e4.\" style=\"left:4.000000%;top:92.321429%;width:92.000000%;height:3.035714%;font-size:1.56250cqi;font-weight:700;color:#235C3A;justify-content:center;text-align:center;\"><span>Hot expert \uac00\uc911\uce58\ub294 GPU HBM\uc5d0, Cold expert \uac00\uc911\uce58\ub294 Host Memory\uc5d0 \ub454\ub2e4.<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"\ub17c\ub9ac\uc801 \uc2e4\ud589 \uad6c\uc870 \u00b7 \uac15\uc870 \uc21c\uc11c\uc640 \uc560\ub2c8\uba54\uc774\uc158 \uc18d\ub3c4\ub294 \uc2e4\uc81c \uc5f0\uc0b0 \uc2dc\uac04\uc774\ub098 \uc2e4\ud589 \uc911\ucca9\uc758 \uce21\uc815\uac12\uc774 \uc544\ub2c8\ub2e4.\" style=\"left:2.625000%;top:96.517857%;width:94.375000%;height:2.410714%;font-size:1.09375cqi;font-weight:400;color:#737373;justify-content:flex-start;text-align:left;\"><span>\ub17c\ub9ac\uc801 \uc2e4\ud589 \uad6c\uc870 \u00b7 \uac15\uc870 \uc21c\uc11c\uc640 \uc560\ub2c8\uba54\uc774\uc158 \uc18d\ub3c4\ub294 \uc2e4\uc81c \uc5f0\uc0b0 \uc2dc\uac04\uc774\ub098 \uc2e4\ud589 \uc911\ucca9\uc758 \uce21\uc815\uac12\uc774 \uc544\ub2c8\ub2e4.<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-route\" data-route=\"checkpoint-convert\" data-source=\"fp8-checkpoint\" data-target=\"amxint4-checkpoint\" style=\"--moe-route-color:#A363CC;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:13.125000%;top:38.214286%;width:2.120000%;height:0.125000cqi;transform:translateY(-50%) rotate(90.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-head moe-route-head-down\" data-head-at=\"target\" data-tip-x=\"210.0\" data-tip-y=\"472.0\" style=\"left:13.125000%;top:42.142857%;width:0.717500cqi;height:0.875000cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"load-cold-weights\" data-source=\"amxint4-checkpoint\" data-target=\"cold-expert-weights\" style=\"--moe-route-color:#A363CC;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:20.750000%;top:45.178571%;width:2.562500%;height:0.125000cqi;transform:translateY(-50%) rotate(0.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-segment\" data-segment=\"1\" style=\"left:23.312500%;top:45.178571%;width:3.812500%;height:0.125000cqi;transform:translateY(-50%) rotate(-90.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-segment\" data-segment=\"2\" style=\"left:23.312500%;top:39.732143%;width:1.682500%;height:0.125000cqi;transform:translateY(-50%) rotate(0.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-head moe-route-head-right\" data-head-at=\"target\" data-tip-x=\"410.0\" data-tip-y=\"445.0\" style=\"left:25.625000%;top:39.732143%;width:0.875000cqi;height:0.717500cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"numa-0-link\" data-source=\"kt-cpuinfer\" data-target=\"numa-0\" style=\"--moe-route-color:#8D96A0;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:69.875000%;top:41.696429%;width:1.562500%;height:0.081250cqi;transform:translateY(-50%) rotate(90.000deg);\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"numa-1-link\" data-source=\"kt-cpuinfer\" data-target=\"numa-1\" style=\"--moe-route-color:#8D96A0;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:86.750000%;top:41.696429%;width:1.562500%;height:0.081250cqi;transform:translateY(-50%) rotate(90.000deg);\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"read-cold-weights\" data-source=\"cold-expert-weights\" data-target=\"kt-cpuinfer\" style=\"--moe-route-color:#A363CC;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:54.500000%;top:35.267857%;width:7.245000%;height:0.125000cqi;transform:translateY(-50%) rotate(0.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-head moe-route-head-right\" data-head-at=\"target\" data-tip-x=\"998.0\" data-tip-y=\"395.0\" style=\"left:62.375000%;top:35.267857%;width:0.875000cqi;height:0.717500cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"input-attention\" data-source=\"input-hidden-state\" data-target=\"attention\" style=\"--moe-route-color:#697B90;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:14.750000%;top:66.964286%;width:1.307500%;height:0.187500cqi;transform:translateY(-50%) rotate(0.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-0.000000cqi;\"><\/span><\/span><span class=\"moe-route-head moe-route-head-right\" data-head-at=\"target\" data-tip-x=\"267.0\" data-tip-y=\"750.0\" style=\"left:16.687500%;top:66.964286%;width:0.875000cqi;height:0.717500cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"attention-router\" data-source=\"attention\" data-target=\"router-dispatch\" style=\"--moe-route-color:#697B90;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:29.562500%;top:66.964286%;width:2.807500%;height:0.187500cqi;transform:translateY(-50%) rotate(0.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-0.000000cqi;\"><\/span><\/span><span class=\"moe-route-head moe-route-head-right\" data-head-at=\"target\" data-tip-x=\"528.0\" data-tip-y=\"750.0\" style=\"left:33.000000%;top:66.964286%;width:0.875000cqi;height:0.717500cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"hot-dispatch\" data-source=\"router-dispatch\" data-target=\"gpu-hot-compute\" style=\"--moe-route-color:#24873F;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:48.250000%;top:66.964286%;width:3.197500%;height:0.312500cqi;transform:translateY(-50%) rotate(0.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-0.000000cqi;\"><\/span><\/span><span class=\"moe-route-head moe-route-head-right\" data-head-at=\"target\" data-tip-x=\"839.0\" data-tip-y=\"750.0\" style=\"left:52.437500%;top:66.964286%;width:1.375000cqi;height:1.127500cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"hot-result\" data-source=\"gpu-hot-compute\" data-target=\"result-combine\" style=\"--moe-route-color:#24873F;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:67.250000%;top:66.964286%;width:7.260000%;height:0.312500cqi;transform:translateY(-50%) rotate(0.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-0.000000cqi;\"><\/span><\/span><span class=\"moe-route-head moe-route-head-right\" data-head-at=\"target\" data-tip-x=\"1208.0\" data-tip-y=\"750.0\" style=\"left:75.500000%;top:66.964286%;width:1.375000cqi;height:1.127500cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"cold-dispatch\" data-source=\"router-dispatch\" data-target=\"kt-cpuinfer\" style=\"--moe-route-color:#1688A4;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:40.687500%;top:63.392857%;width:6.750000%;height:0.312500cqi;transform:translateY(-50%) rotate(-90.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-0.000000cqi;\"><\/span><\/span><span class=\"moe-route-segment\" data-segment=\"1\" style=\"left:40.687500%;top:53.750000%;width:37.687500%;height:0.312500cqi;transform:translateY(-50%) rotate(0.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-6.750000cqi;\"><\/span><\/span><span class=\"moe-route-segment\" data-segment=\"2\" style=\"left:78.375000%;top:53.750000%;width:7.322500%;height:0.312500cqi;transform:translateY(-50%) rotate(-90.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-44.437500cqi;\"><\/span><\/span><span class=\"moe-route-head moe-route-head-up\" data-head-at=\"target\" data-tip-x=\"1254.0\" data-tip-y=\"469.0\" style=\"left:78.375000%;top:41.875000%;width:1.127500cqi;height:1.375000cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"cold-result\" data-source=\"kt-cpuinfer\" data-target=\"result-combine\" style=\"--moe-route-color:#357F64;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:94.250000%;top:37.232143%;width:2.375000%;height:0.312500cqi;transform:translateY(-50%) rotate(0.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-0.000000cqi;\"><\/span><\/span><span class=\"moe-route-segment\" data-segment=\"1\" style=\"left:96.625000%;top:37.232143%;width:20.812500%;height:0.312500cqi;transform:translateY(-50%) rotate(90.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-2.375000cqi;\"><\/span><\/span><span class=\"moe-route-segment\" data-segment=\"2\" style=\"left:96.625000%;top:66.964286%;width:1.885000%;height:0.312500cqi;transform:translateY(-50%) rotate(180.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-23.187500cqi;\"><\/span><\/span><span class=\"moe-route-head moe-route-head-left\" data-head-at=\"target\" data-tip-x=\"1500.0\" data-tip-y=\"750.0\" style=\"left:93.750000%;top:66.964286%;width:1.375000cqi;height:1.127500cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"residual-input\" data-source=\"attention\" data-target=\"result-combine\" style=\"--moe-route-color:#8C99A7;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:31.312500%;top:66.964286%;width:4.437500%;height:0.125000cqi;transform:translateY(-50%) rotate(90.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-segment\" data-segment=\"1\" style=\"left:31.312500%;top:73.303571%;width:58.500000%;height:0.125000cqi;transform:translateY(-50%) rotate(0.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-segment\" data-segment=\"2\" style=\"left:89.812500%;top:73.303571%;width:1.182500%;height:0.125000cqi;transform:translateY(-50%) rotate(-90.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-head moe-route-head-up\" data-head-at=\"target\" data-tip-x=\"1437.0\" data-tip-y=\"792.0\" style=\"left:89.812500%;top:70.714286%;width:0.717500cqi;height:0.875000cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"load-gpu-weights\" data-source=\"fp8-checkpoint\" data-target=\"gpu-0\" style=\"--moe-route-color:#A363CC;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:5.750000%;top:35.446429%;width:2.437500%;height:0.125000cqi;transform:translateY(-50%) rotate(180.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-segment\" data-segment=\"1\" style=\"left:3.312500%;top:35.446429%;width:33.000000%;height:0.125000cqi;transform:translateY(-50%) rotate(90.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-segment\" data-segment=\"2\" style=\"left:3.312500%;top:82.589286%;width:2.057500%;height:0.125000cqi;transform:translateY(-50%) rotate(0.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-head moe-route-head-right\" data-head-at=\"target\" data-tip-x=\"96.0\" data-tip-y=\"925.0\" style=\"left:6.000000%;top:82.589286%;width:0.875000cqi;height:0.717500cqi;\"><\/span><\/div><\/div><figcaption class=\"moe-caption\">\uadf8\ub9bc 2. Qwen3-Coder-480B\uc758 GPU 4\uc7a5 \ubd80\ubd84 \uc0c1\uc8fc \uad6c\uc870. Hot expert\ub294 GPU\uc5d0\uc11c, Cold expert\ub294 CPU\uc5d0\uc11c \uacc4\uc0b0\ud558\uba70 \ub450 \uacbd\ub85c\uc758 \uacb0\uacfc\ub97c GPU\uc5d0\uc11c \uacb0\ud569\ud55c\ub2e4.<\/figcaption><details class=\"moe-enlarge\"><summary>\uad6c\uc870\ub3c4 \ud655\ub300 \ubcf4\uae30<\/summary><div class=\"moe-zoom-viewport\" tabindex=\"0\"><div class=\"moe-zoom-detail\"><div aria-label=\"\ub2e8\uc77c \uc11c\ubc84 CPU MoE \ubd80\ubd84 \uc0c1\uc8fc \uad6c\uc870. Hot expert\ub294 GPU HBM\uc5d0 \uc0c1\uc8fc\ud558\uace0 GPU\uc5d0\uc11c \uc2e4\ud589\ud55c\ub2e4. Cold expert\ub294 Host Memory\uc5d0 \uc0c1\uc8fc\ud558\uace0 CPU\uc5d0\uc11c \uc2e4\ud589\ud55c\ub2e4. Router\uac00 \uc120\ud0dd\ud55c \uc804\ubb38\uac00\uc758 \ubc30\uce58\uc5d0 \ub530\ub77c \uacbd\ub85c\uac00 \ub098\ub258\uba70 \ub450 \uacb0\uacfc\ub294 GPU\uc5d0\uc11c \uacb0\ud569\ub41c\ub2e4.\" class=\"moe-diagram moe-detailed moe-hc-diagram\" role=\"group\" style=\"aspect-ratio:10 \/ 7;overflow:hidden;font-family:Arial,'Noto Sans CJK KR','Malgun Gothic',sans-serif;color:#20252B;line-height:1.24;\">\n<div class=\"moe-hc-label\" data-label=\"CPU MoE Offloading | Hot\uc740 GPU, Cold\ub294 Host Memory\" style=\"left:2.500000%;top:1.607143%;width:94.375000%;height:4.910714%;font-size:2.25000cqi;font-weight:700;color:#20252B;justify-content:flex-start;text-align:left;\"><span>CPU MoE Offloading | Hot\uc740 GPU, Cold\ub294 Host Memory<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"\ub2e8\uc77c \uc11c\ubc84\uc758 \uac00\uc911\uce58 \ubc30\uce58\uc640 \uc2e4\ud589 \uacbd\ub85c \u00b7 Qwen3-Coder-480B \u00b7 TP4 \ubd80\ubd84 \uc0c1\uc8fc \uc608\uc2dc\" style=\"left:2.625000%;top:7.053571%;width:93.750000%;height:3.125000%;font-size:1.43750cqi;font-weight:400;color:#697581;justify-content:flex-start;text-align:left;\"><span>\ub2e8\uc77c \uc11c\ubc84\uc758 \uac00\uc911\uce58 \ubc30\uce58\uc640 \uc2e4\ud589 \uacbd\ub85c \u00b7 Qwen3-Coder-480B \u00b7 TP4 \ubd80\ubd84 \uc0c1\uc8fc \uc608\uc2dc<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:2.000000%;top:11.785714%;width:96.000000%;height:83.928571%;box-sizing:border-box;background:#FFFFFF;z-index:1;border:0.08750cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:2.000000%;top:11.785714%;width:96.000000%;height:4.464286%;box-sizing:border-box;background:#FAFAFA;z-index:1;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:2.000000%;top:11.785714%;width:0.437500%;height:4.464286%;box-sizing:border-box;background:#538134;z-index:1;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\ub2e8\uc77c \ubb3c\ub9ac \uc11c\ubc84 \u00b7 CPU\uc640 GPU\uac00 \ud568\uaed8 \uc2e4\ud589\" style=\"left:3.250000%;top:12.321429%;width:43.750000%;height:3.482143%;font-size:1.75000cqi;font-weight:700;color:#20252B;justify-content:flex-start;text-align:left;\"><span>\ub2e8\uc77c \ubb3c\ub9ac \uc11c\ubc84 \u00b7 CPU\uc640 GPU\uac00 \ud568\uaed8 \uc2e4\ud589<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"Xeon 8480+ \u00d72 \u00b7 Host DRAM 2 TB \u00b7 H100 80 GB \u00d74\" style=\"left:54.250000%;top:12.678571%;width:41.750000%;height:2.946429%;font-size:1.43750cqi;font-weight:400;color:#68737E;justify-content:flex-end;text-align:right;\"><span>Xeon 8480+ \u00d72 \u00b7 Host DRAM 2 TB \u00b7 H100 80 GB \u00d74<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"runtime\" style=\"position:absolute;left:4.500000%;top:18.214286%;width:91.000000%;height:5.000000%;box-sizing:border-box;background:#FFFFFF;z-index:1;border:0.08750cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:4.500000%;top:18.214286%;width:91.000000%;height:3.928571%;box-sizing:border-box;background:#FAFAFA;z-index:1;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:4.500000%;top:18.214286%;width:0.375000%;height:3.928571%;box-sizing:border-box;background:#FF7E8A;z-index:1;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"SGLang \uc11c\ube59 \ub7f0\ud0c0\uc784 | \uc694\uccad \uc2a4\ucf00\uc904\ub9c1 \u00b7 TP4 GPU \uc6cc\ucee4 \u00b7 KT CPUInfer \uc5f0\ub3d9\" style=\"left:5.500000%;top:18.482143%;width:89.125000%;height:3.303571%;font-size:1.62500cqi;font-weight:700;color:#20252B;justify-content:flex-start;text-align:left;\"><span>SGLang \uc11c\ube59 \ub7f0\ud0c0\uc784 | \uc694\uccad \uc2a4\ucf00\uc904\ub9c1 \u00b7 TP4 GPU \uc6cc\ucee4 \u00b7 KT CPUInfer \uc5f0\ub3d9<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"model-files\" style=\"position:absolute;left:4.500000%;top:26.785714%;width:17.500000%;height:23.214286%;box-sizing:border-box;background:#FFFFFF;z-index:1;border:0.08750cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:4.500000%;top:26.785714%;width:17.500000%;height:3.928571%;box-sizing:border-box;background:#FAFAFA;z-index:1;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:4.500000%;top:26.785714%;width:0.375000%;height:3.928571%;box-sizing:border-box;background:#737373;z-index:1;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\ubaa8\ub378 \ud30c\uc77c \u00b7 \uae30\ub3d9 \uc2dc \uc801\uc7ac\" style=\"left:5.500000%;top:27.053571%;width:15.625000%;height:3.303571%;font-size:1.43750cqi;font-weight:700;color:#20252B;justify-content:flex-start;text-align:left;\"><span>\ubaa8\ub378 \ud30c\uc77c \u00b7 \uae30\ub3d9 \uc2dc \uc801\uc7ac<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"fp8-checkpoint\" style=\"position:absolute;left:5.750000%;top:32.589286%;width:15.000000%;height:5.625000%;box-sizing:border-box;background:#F1E9F8;z-index:3;border:0.09375cqi solid #C78AFF;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\uc6d0\ubcf8 FP8 checkpoint\nmodel_path\" style=\"left:6.250000%;top:32.946429%;width:14.000000%;height:4.910714%;font-size:1.37500cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>\uc6d0\ubcf8 FP8 checkpoint<br>model_path<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"amxint4-checkpoint\" style=\"position:absolute;left:5.750000%;top:42.321429%;width:15.000000%;height:5.625000%;box-sizing:border-box;background:#F1E9F8;z-index:3;border:0.09375cqi solid #C78AFF;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"CPU\uc6a9 INT4 \ubcc0\ud658\ubcf8\nkt_weight_path\" style=\"left:6.250000%;top:42.678571%;width:14.000000%;height:4.910714%;font-size:1.37500cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>CPU\uc6a9 INT4 \ubcc0\ud658\ubcf8<br>kt_weight_path<\/span><\/div>\n<!-- FP8 -> INT4 offline conversion -->\n<div class=\"moe-hc-label\" data-label=\"kt quant \u00b7 INT4\" style=\"left:13.875000%;top:38.660714%;width:7.250000%;height:3.214286%;font-size:1.06250cqi;font-weight:400;color:#737373;justify-content:flex-start;text-align:left;\"><span>kt quant \u00b7 INT4<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"host-memory\" style=\"position:absolute;left:24.500000%;top:26.785714%;width:31.250000%;height:23.214286%;box-sizing:border-box;background:#FFFFFF;z-index:1;border:0.08750cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:24.500000%;top:26.785714%;width:31.250000%;height:3.928571%;box-sizing:border-box;background:#E2F0D9;z-index:1;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:24.500000%;top:26.785714%;width:0.375000%;height:3.928571%;box-sizing:border-box;background:#538134;z-index:1;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Host Memory \u00b7 DDR5 2 TB\" style=\"left:25.500000%;top:27.053571%;width:29.375000%;height:3.303571%;font-size:1.62500cqi;font-weight:700;color:#20252B;justify-content:flex-start;text-align:left;\"><span>Host Memory \u00b7 DDR5 2 TB<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"cold-expert-weights\" style=\"position:absolute;left:25.750000%;top:32.589286%;width:28.750000%;height:13.571429%;box-sizing:border-box;background:#F5FAFF;z-index:3;border:0.09375cqi solid #8DBBF3;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\" style=\"left:26.250000%;top:32.946429%;width:27.750000%;height:12.857143%;font-size:1.43750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span><\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"Cold expert \uac00\uc911\uce58 \u00b7 INT4\" style=\"left:26.375000%;top:33.303571%;width:27.500000%;height:3.482143%;font-size:1.75000cqi;font-weight:700;color:#1C5997;justify-content:center;text-align:center;\"><span>Cold expert \uac00\uc911\uce58 \u00b7 INT4<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:26.687500%;top:38.303571%;width:5.875000%;height:3.392857%;box-sizing:border-box;background:#EAF3FF;z-index:3;border:0.09375cqi solid #79ADE9;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Cold E\u2090\" style=\"left:27.187500%;top:38.392857%;width:4.875000%;height:3.214286%;font-size:1.25000cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>Cold E\u2090<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:33.437500%;top:38.303571%;width:5.875000%;height:3.392857%;box-sizing:border-box;background:#EAF3FF;z-index:3;border:0.09375cqi solid #79ADE9;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Cold E\u1d66\" style=\"left:33.937500%;top:38.392857%;width:4.875000%;height:3.214286%;font-size:1.25000cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>Cold E\u1d66<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:40.187500%;top:38.303571%;width:5.875000%;height:3.392857%;box-sizing:border-box;background:#EAF3FF;z-index:3;border:0.09375cqi solid #79ADE9;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\u2026\" style=\"left:40.687500%;top:38.392857%;width:4.875000%;height:3.214286%;font-size:1.25000cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>\u2026<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:46.937500%;top:38.303571%;width:5.875000%;height:3.392857%;box-sizing:border-box;background:#EAF3FF;z-index:3;border:0.09375cqi solid #79ADE9;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Cold E\u2099\" style=\"left:47.437500%;top:38.392857%;width:4.875000%;height:3.214286%;font-size:1.25000cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>Cold E\u2099<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"\ubc30\uce58\ud45c\uc5d0\uc11c CPU\ub85c \uc9c0\uc815\ud55c \uc804\ubb38\uac00 \uc9d1\ud569\" style=\"left:26.437500%;top:42.678571%;width:27.375000%;height:2.589286%;font-size:1.31250cqi;font-weight:400;color:#56728C;justify-content:center;text-align:center;\"><span>\ubc30\uce58\ud45c\uc5d0\uc11c CPU\ub85c \uc9c0\uc815\ud55c \uc804\ubb38\uac00 \uc9d1\ud569<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"\uac00\uc911\uce58\ub294 Host Memory\uc5d0 \uc0c1\uc8fc\" style=\"left:25.687500%;top:47.410714%;width:29.062500%;height:2.500000%;font-size:1.37500cqi;font-weight:700;color:#347543;justify-content:center;text-align:center;\"><span>\uac00\uc911\uce58\ub294 Host Memory\uc5d0 \uc0c1\uc8fc<\/span><\/div>\n<!-- INT4 load into host memory -->\n<div aria-hidden=\"true\" class=\"\" data-component=\"cpu-compute\" style=\"position:absolute;left:61.250000%;top:26.785714%;width:34.250000%;height:23.214286%;box-sizing:border-box;background:#FFFFFF;z-index:1;border:0.08750cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:61.250000%;top:26.785714%;width:34.250000%;height:3.928571%;box-sizing:border-box;background:#E2F0D9;z-index:1;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:61.250000%;top:26.785714%;width:0.375000%;height:3.928571%;box-sizing:border-box;background:#538134;z-index:1;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"CPU \u00b7 Xeon 8480+ \u00d72\" style=\"left:62.250000%;top:27.053571%;width:32.375000%;height:3.303571%;font-size:1.62500cqi;font-weight:700;color:#20252B;justify-content:flex-start;text-align:left;\"><span>CPU \u00b7 Xeon 8480+ \u00d72<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-stage moe-phase-cpu\" data-component=\"kt-cpuinfer\" style=\"position:absolute;left:62.500000%;top:32.589286%;width:31.750000%;height:9.107143%;box-sizing:border-box;background:#F8FCF8;z-index:3;border:0.09375cqi solid #629A6C;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\" style=\"left:63.000000%;top:32.946429%;width:30.750000%;height:8.392857%;font-size:1.43750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span><\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"\u2462-C  Cold expert \uacc4\uc0b0 \u00b7 KT CPUInfer\" style=\"left:63.125000%;top:33.214286%;width:30.500000%;height:3.214286%;font-size:1.56250cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>\u2462-C  Cold expert \uacc4\uc0b0 \u00b7 KT CPUInfer<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"CPU\uac00 Host Memory\uc758 \uac00\uc911\uce58\ub97c \uc77d\uc5b4 \uc5f0\uc0b0\" style=\"left:63.125000%;top:36.875000%;width:30.500000%;height:2.678571%;font-size:1.31250cqi;font-weight:400;color:#20252B;justify-content:center;text-align:center;\"><span>CPU\uac00 Host Memory\uc758 \uac00\uc911\uce58\ub97c \uc77d\uc5b4 \uc5f0\uc0b0<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-stage moe-phase-worker\" data-component=\"numa-0\" style=\"position:absolute;left:62.500000%;top:43.928571%;width:15.000000%;height:3.660714%;box-sizing:border-box;background:#FFFFFF;z-index:3;border:0.09375cqi solid #BEBEBE;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"NUMA 0 \u00b7 48 workers\" style=\"left:63.000000%;top:44.017857%;width:14.000000%;height:3.482143%;font-size:1.31250cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>NUMA 0 \u00b7 48 workers<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-stage moe-phase-worker\" data-component=\"numa-1\" style=\"position:absolute;left:79.250000%;top:43.928571%;width:15.000000%;height:3.660714%;box-sizing:border-box;background:#FFFFFF;z-index:3;border:0.09375cqi solid #BEBEBE;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"NUMA 1 \u00b7 48 workers\" style=\"left:79.750000%;top:44.017857%;width:14.000000%;height:3.482143%;font-size:1.31250cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>NUMA 1 \u00b7 48 workers<\/span><\/div>\n<!-- Cold weights: Host Memory -> CPU compute (local read) -->\n<div class=\"moe-hc-label\" data-label=\"\uac00\uc911\uce58\n\uc77d\uae30\" style=\"left:56.062500%;top:30.625000%;width:4.812500%;height:4.107143%;font-size:1.18750cqi;font-weight:400;color:#737373;justify-content:center;text-align:center;\"><span>\uac00\uc911\uce58<br>\uc77d\uae30<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"gpu-bank\" style=\"position:absolute;left:4.500000%;top:57.500000%;width:91.000000%;height:34.285714%;box-sizing:border-box;background:#FFFFFF;z-index:1;border:0.08750cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:4.500000%;top:57.500000%;width:91.000000%;height:3.928571%;box-sizing:border-box;background:#FFF2CC;z-index:1;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:4.500000%;top:57.500000%;width:0.375000%;height:3.928571%;box-sizing:border-box;background:#BF9000;z-index:1;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"GPU \u00b7 H100 80 GB \u00d74 \u00b7 TP4\" style=\"left:5.500000%;top:57.767857%;width:89.125000%;height:3.303571%;font-size:1.68750cqi;font-weight:700;color:#20252B;justify-content:flex-start;text-align:left;\"><span>GPU \u00b7 H100 80 GB \u00d74 \u00b7 TP4<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"Hot expert + Attention \/ Router + KV cache\" style=\"left:52.812500%;top:57.857143%;width:41.437500%;height:3.125000%;font-size:1.43750cqi;font-weight:700;color:#986B00;justify-content:flex-end;text-align:right;\"><span>Hot expert + Attention \/ Router + KV cache<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"input-hidden-state\" style=\"position:absolute;left:6.062500%;top:63.392857%;width:8.687500%;height:7.142857%;box-sizing:border-box;background:#FFFFFF;z-index:3;border:0.09375cqi solid #BEBEBE;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\uc785\ub825\nhidden state\" style=\"left:6.562500%;top:63.750000%;width:7.687500%;height:6.428571%;font-size:1.25000cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>\uc785\ub825<br>hidden state<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-stage moe-phase-attn\" data-component=\"attention\" style=\"position:absolute;left:16.812500%;top:63.392857%;width:12.750000%;height:7.142857%;box-sizing:border-box;background:#F1E9F8;z-index:3;border:0.09375cqi solid #C78AFF;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\u2460 Attention\nMoE \uc785\ub825 \uc900\ube44\" style=\"left:17.312500%;top:63.750000%;width:11.750000%;height:6.428571%;font-size:1.56250cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>\u2460 Attention<br>MoE \uc785\ub825 \uc900\ube44<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-stage moe-phase-route\" data-component=\"router-dispatch\" style=\"position:absolute;left:33.125000%;top:63.392857%;width:15.125000%;height:7.142857%;box-sizing:border-box;background:#F1E9F8;z-index:3;border:0.09375cqi solid #C78AFF;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\u2461 Router \/ dispatch\ntop-8 \u2192 Hot \/ Cold \ubd84\uae30\" style=\"left:33.625000%;top:63.750000%;width:14.125000%;height:6.428571%;font-size:1.25000cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>\u2461 Router \/ dispatch<br>top-8 \u2192 Hot \/ Cold \ubd84\uae30<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-stage moe-phase-hot\" data-component=\"gpu-hot-compute\" style=\"position:absolute;left:52.562500%;top:63.392857%;width:14.687500%;height:7.142857%;box-sizing:border-box;background:#E9F7E8;z-index:3;border:0.09375cqi solid #4B9B58;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\u2462-G  Hot expert\nGPU \uc5f0\uc0b0 \u00b7 FP8\" style=\"left:53.062500%;top:63.750000%;width:13.687500%;height:6.428571%;font-size:1.50000cqi;font-weight:700;color:#236638;justify-content:center;text-align:center;\"><span>\u2462-G  Hot expert<br>GPU \uc5f0\uc0b0 \u00b7 FP8<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-stage moe-phase-combine\" data-component=\"result-combine\" style=\"position:absolute;left:75.625000%;top:63.392857%;width:18.000000%;height:7.142857%;box-sizing:border-box;background:#E2F0D9;z-index:3;border:0.09375cqi solid #538134;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"\u2463 Hot + Cold \uacb0\uacfc \uacb0\ud569\n\uc794\ucc28 \uc5f0\uacb0 \u2192 \ub2e4\uc74c \uce35\" style=\"left:76.125000%;top:63.750000%;width:17.000000%;height:6.428571%;font-size:1.50000cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>\u2463 Hot + Cold \uacb0\uacfc \uacb0\ud569<br>\uc794\ucc28 \uc5f0\uacb0 \u2192 \ub2e4\uc74c \uce35<\/span><\/div>\n<!-- Input -> GPU Attention -->\n<!-- MoE input activation -> Router \/ dispatch -->\n<div class=\"moe-hc-label\" data-label=\"x\" style=\"left:29.812500%;top:63.839286%;width:3.062500%;height:2.232143%;font-size:1.25000cqi;font-weight:400;color:#68737E;justify-content:center;text-align:center;\"><span>x<\/span><\/div>\n<!-- Selected Hot expert input -> GPU computation -->\n<div class=\"moe-hc-label\" data-label=\"Hot\" style=\"left:48.562500%;top:63.571429%;width:3.625000%;height:2.232143%;font-size:1.18750cqi;font-weight:700;color:#347E42;justify-content:center;text-align:center;\"><span>Hot<\/span><\/div>\n<!-- GPU Hot result -> combine -->\n<div class=\"moe-hc-label\" data-label=\"Hot \uacb0\uacfc\" style=\"left:67.812500%;top:63.750000%;width:7.312500%;height:2.410714%;font-size:1.25000cqi;font-weight:700;color:#347E42;justify-content:center;text-align:center;\"><span>Hot \uacb0\uacfc<\/span><\/div>\n<!-- Cold activation + selected IDs \/ routing weights -> CPU -->\n<div class=\"moe-hc-label\" data-label=\"Cold \uacbd\ub85c: \ud65c\uc131\uac12 + \uc120\ud0dd ID \u00b7 \ub77c\uc6b0\ud305 \uac00\uc911\uce58\" style=\"left:41.562500%;top:50.446429%;width:35.000000%;height:2.857143%;font-size:1.50000cqi;font-weight:700;color:#25869B;justify-content:flex-start;text-align:left;\"><span>Cold \uacbd\ub85c: \ud65c\uc131\uac12 + \uc120\ud0dd ID \u00b7 \ub77c\uc6b0\ud305 \uac00\uc911\uce58<\/span><\/div>\n<!-- CPU Cold result -> GPU combine -->\n<div class=\"moe-hc-label\" data-label=\"Cold \uc5f0\uc0b0 \uacb0\uacfc \ubc18\ud658\" style=\"left:81.312500%;top:51.696429%;width:14.062500%;height:2.857143%;font-size:1.43750cqi;font-weight:700;color:#407D61;justify-content:flex-end;text-align:right;\"><span>Cold \uc5f0\uc0b0 \uacb0\uacfc \ubc18\ud658<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"\ud65c\uc131\uac12\uacfc \uacb0\uacfc\ub9cc \ub3d9\uc77c \uc11c\ubc84 \ub0b4\ubd80\uc5d0\uc11c \uc655\ubcf5\" style=\"left:45.000000%;top:54.285714%;width:30.000000%;height:2.410714%;font-size:1.12500cqi;font-weight:400;color:#73808A;justify-content:center;text-align:center;\"><span>\ud65c\uc131\uac12\uacfc \uacb0\uacfc\ub9cc \ub3d9\uc77c \uc11c\ubc84 \ub0b4\ubd80\uc5d0\uc11c \uc655\ubcf5<\/span><\/div>\n<!-- Residual input bypass -> residual connection -->\n<div class=\"moe-hc-label\" data-label=\"\uc794\ucc28 \uc785\ub825 \u2014 \uc804\ubb38\uac00 \uacc4\uc0b0 \uacbd\ub85c\uc640 \ubcc4\ub3c4\ub85c \uc804\ub2ec\" style=\"left:34.750000%;top:70.892857%;width:35.625000%;height:2.321429%;font-size:1.12500cqi;font-weight:400;color:#79838D;justify-content:center;text-align:center;\"><span>\uc794\ucc28 \uc785\ub825 \u2014 \uc804\ubb38\uac00 \uacc4\uc0b0 \uacbd\ub85c\uc640 \ubcc4\ub3c4\ub85c \uc804\ub2ec<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"\uac01 GPU HBM\uc5d0 \uc0c1\uc8fc\ud558\ub294 \ub0b4\uc6a9 \u00b7 TP4 \ubd84\ud560\" style=\"left:5.875000%;top:74.821429%;width:56.250000%;height:2.767857%;font-size:1.50000cqi;font-weight:700;color:#AD810B;justify-content:flex-start;text-align:left;\"><span>\uac01 GPU HBM\uc5d0 \uc0c1\uc8fc\ud558\ub294 \ub0b4\uc6a9 \u00b7 TP4 \ubd84\ud560<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"GPU \uac04 \uc5f0\uacb0: NVLink\" style=\"left:69.375000%;top:74.910714%;width:24.500000%;height:2.500000%;font-size:1.31250cqi;font-weight:400;color:#778390;justify-content:flex-end;text-align:right;\"><span>GPU \uac04 \uc5f0\uacb0: NVLink<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-hbm-frame\" data-component=\"gpu-0\" style=\"position:absolute;left:6.125000%;top:78.125000%;width:20.437500%;height:11.696429%;box-sizing:border-box;background:#FFFFFF;z-index:3;border:0.06250cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:6.187500%;top:78.214286%;width:20.312500%;height:2.500000%;box-sizing:border-box;background:#FFF2CC;z-index:3;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"GPU 0 \u00b7 rank 0\" style=\"left:6.625000%;top:78.303571%;width:19.437500%;height:2.410714%;font-size:1.31250cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>GPU 0 \u00b7 rank 0<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"gpu-0-hot-weights\" style=\"position:absolute;left:6.625000%;top:81.428571%;width:19.437500%;height:2.232143%;box-sizing:border-box;background:#E9F7E8;z-index:3;border:0.09375cqi solid #80BC83;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Hot expert TP \uc870\uac01 \u00b7 FP8\" style=\"left:7.125000%;top:81.517857%;width:18.437500%;height:2.053571%;font-size:1.25000cqi;font-weight:700;color:#246337;justify-content:center;text-align:center;\"><span>Hot expert TP \uc870\uac01 \u00b7 FP8<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"gpu-0-nonexpert-weights\" style=\"position:absolute;left:6.625000%;top:84.285714%;width:19.437500%;height:2.142857%;box-sizing:border-box;background:#F1E9F8;z-index:3;border:0.09375cqi solid #C78AFF;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Attention \ub4f1 \ube44\uc804\ubb38\uac00 TP \uc870\uac01\" style=\"left:7.125000%;top:84.375000%;width:18.437500%;height:1.964286%;font-size:1.18750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>Attention \ub4f1 \ube44\uc804\ubb38\uac00 TP \uc870\uac01<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-cache-access\" data-component=\"gpu-0-kv-cache\" style=\"position:absolute;left:6.625000%;top:87.053571%;width:19.437500%;height:2.142857%;box-sizing:border-box;background:#E7F5FE;z-index:3;border:0.09375cqi solid #6CBCDA;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"KV cache \u00b7 GPU HBM\" style=\"left:7.125000%;top:87.142857%;width:18.437500%;height:1.964286%;font-size:1.18750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>KV cache \u00b7 GPU HBM<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-hbm-frame\" data-component=\"gpu-1\" style=\"position:absolute;left:28.125000%;top:78.125000%;width:20.437500%;height:11.696429%;box-sizing:border-box;background:#FFFFFF;z-index:3;border:0.06250cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:28.187500%;top:78.214286%;width:20.312500%;height:2.500000%;box-sizing:border-box;background:#FFF2CC;z-index:3;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"GPU 1 \u00b7 rank 1\" style=\"left:28.625000%;top:78.303571%;width:19.437500%;height:2.410714%;font-size:1.31250cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>GPU 1 \u00b7 rank 1<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"gpu-1-hot-weights\" style=\"position:absolute;left:28.625000%;top:81.428571%;width:19.437500%;height:2.232143%;box-sizing:border-box;background:#E9F7E8;z-index:3;border:0.09375cqi solid #80BC83;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Hot expert TP \uc870\uac01 \u00b7 FP8\" style=\"left:29.125000%;top:81.517857%;width:18.437500%;height:2.053571%;font-size:1.25000cqi;font-weight:700;color:#246337;justify-content:center;text-align:center;\"><span>Hot expert TP \uc870\uac01 \u00b7 FP8<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"gpu-1-nonexpert-weights\" style=\"position:absolute;left:28.625000%;top:84.285714%;width:19.437500%;height:2.142857%;box-sizing:border-box;background:#F1E9F8;z-index:3;border:0.09375cqi solid #C78AFF;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Attention \ub4f1 \ube44\uc804\ubb38\uac00 TP \uc870\uac01\" style=\"left:29.125000%;top:84.375000%;width:18.437500%;height:1.964286%;font-size:1.18750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>Attention \ub4f1 \ube44\uc804\ubb38\uac00 TP \uc870\uac01<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-cache-access\" data-component=\"gpu-1-kv-cache\" style=\"position:absolute;left:28.625000%;top:87.053571%;width:19.437500%;height:2.142857%;box-sizing:border-box;background:#E7F5FE;z-index:3;border:0.09375cqi solid #6CBCDA;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"KV cache \u00b7 GPU HBM\" style=\"left:29.125000%;top:87.142857%;width:18.437500%;height:1.964286%;font-size:1.18750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>KV cache \u00b7 GPU HBM<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-hbm-frame\" data-component=\"gpu-2\" style=\"position:absolute;left:50.125000%;top:78.125000%;width:20.437500%;height:11.696429%;box-sizing:border-box;background:#FFFFFF;z-index:3;border:0.06250cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:50.187500%;top:78.214286%;width:20.312500%;height:2.500000%;box-sizing:border-box;background:#FFF2CC;z-index:3;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"GPU 2 \u00b7 rank 2\" style=\"left:50.625000%;top:78.303571%;width:19.437500%;height:2.410714%;font-size:1.31250cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>GPU 2 \u00b7 rank 2<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"gpu-2-hot-weights\" style=\"position:absolute;left:50.625000%;top:81.428571%;width:19.437500%;height:2.232143%;box-sizing:border-box;background:#E9F7E8;z-index:3;border:0.09375cqi solid #80BC83;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Hot expert TP \uc870\uac01 \u00b7 FP8\" style=\"left:51.125000%;top:81.517857%;width:18.437500%;height:2.053571%;font-size:1.25000cqi;font-weight:700;color:#246337;justify-content:center;text-align:center;\"><span>Hot expert TP \uc870\uac01 \u00b7 FP8<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"gpu-2-nonexpert-weights\" style=\"position:absolute;left:50.625000%;top:84.285714%;width:19.437500%;height:2.142857%;box-sizing:border-box;background:#F1E9F8;z-index:3;border:0.09375cqi solid #C78AFF;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Attention \ub4f1 \ube44\uc804\ubb38\uac00 TP \uc870\uac01\" style=\"left:51.125000%;top:84.375000%;width:18.437500%;height:1.964286%;font-size:1.18750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>Attention \ub4f1 \ube44\uc804\ubb38\uac00 TP \uc870\uac01<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-cache-access\" data-component=\"gpu-2-kv-cache\" style=\"position:absolute;left:50.625000%;top:87.053571%;width:19.437500%;height:2.142857%;box-sizing:border-box;background:#E7F5FE;z-index:3;border:0.09375cqi solid #6CBCDA;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"KV cache \u00b7 GPU HBM\" style=\"left:51.125000%;top:87.142857%;width:18.437500%;height:1.964286%;font-size:1.18750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>KV cache \u00b7 GPU HBM<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-hbm-frame\" data-component=\"gpu-3\" style=\"position:absolute;left:72.125000%;top:78.125000%;width:20.437500%;height:11.696429%;box-sizing:border-box;background:#FFFFFF;z-index:3;border:0.06250cqi solid #BEBEBE;\"><\/div>\n<div aria-hidden=\"true\" class=\"\" style=\"position:absolute;left:72.187500%;top:78.214286%;width:20.312500%;height:2.500000%;box-sizing:border-box;background:#FFF2CC;z-index:3;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"GPU 3 \u00b7 rank 3\" style=\"left:72.625000%;top:78.303571%;width:19.437500%;height:2.410714%;font-size:1.31250cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>GPU 3 \u00b7 rank 3<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"gpu-3-hot-weights\" style=\"position:absolute;left:72.625000%;top:81.428571%;width:19.437500%;height:2.232143%;box-sizing:border-box;background:#E9F7E8;z-index:3;border:0.09375cqi solid #80BC83;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Hot expert TP \uc870\uac01 \u00b7 FP8\" style=\"left:73.125000%;top:81.517857%;width:18.437500%;height:2.053571%;font-size:1.25000cqi;font-weight:700;color:#246337;justify-content:center;text-align:center;\"><span>Hot expert TP \uc870\uac01 \u00b7 FP8<\/span><\/div>\n<div aria-hidden=\"true\" class=\"\" data-component=\"gpu-3-nonexpert-weights\" style=\"position:absolute;left:72.625000%;top:84.285714%;width:19.437500%;height:2.142857%;box-sizing:border-box;background:#F1E9F8;z-index:3;border:0.09375cqi solid #C78AFF;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"Attention \ub4f1 \ube44\uc804\ubb38\uac00 TP \uc870\uac01\" style=\"left:73.125000%;top:84.375000%;width:18.437500%;height:1.964286%;font-size:1.18750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>Attention \ub4f1 \ube44\uc804\ubb38\uac00 TP \uc870\uac01<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-cache-access\" data-component=\"gpu-3-kv-cache\" style=\"position:absolute;left:72.625000%;top:87.053571%;width:19.437500%;height:2.142857%;box-sizing:border-box;background:#E7F5FE;z-index:3;border:0.09375cqi solid #6CBCDA;\"><\/div>\n<div class=\"moe-hc-label\" data-label=\"KV cache \u00b7 GPU HBM\" style=\"left:73.125000%;top:87.142857%;width:18.437500%;height:1.964286%;font-size:1.18750cqi;font-weight:700;color:#20252B;justify-content:center;text-align:center;\"><span>KV cache \u00b7 GPU HBM<\/span><\/div>\n<!-- FP8 Hot\/non-expert checkpoint load -> GPU HBM -->\n<div class=\"moe-hc-label\" data-label=\"\uc810\uc120: \uac00\uc911\uce58 \uc801\uc7ac\u00b7\ucc38\uc870 \/ \uc794\ucc28 \uc785\ub825\" style=\"left:5.250000%;top:51.607143%;width:32.625000%;height:2.321429%;font-size:1.25000cqi;font-weight:400;color:#76818B;justify-content:flex-start;text-align:left;\"><span>\uc810\uc120: \uac00\uc911\uce58 \uc801\uc7ac\u00b7\ucc38\uc870 \/ \uc794\ucc28 \uc785\ub825<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"\uc0c9 \uc2e4\uc120: \ud1a0\ud070 \ub370\uc774\ud130\uc640 \uc804\ubb38\uac00 \uc5f0\uc0b0 \uacb0\uacfc\" style=\"left:5.250000%;top:54.196429%;width:32.625000%;height:2.321429%;font-size:1.25000cqi;font-weight:400;color:#76818B;justify-content:flex-start;text-align:left;\"><span>\uc0c9 \uc2e4\uc120: \ud1a0\ud070 \ub370\uc774\ud130\uc640 \uc804\ubb38\uac00 \uc5f0\uc0b0 \uacb0\uacfc<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"Hot expert \uac00\uc911\uce58\ub294 GPU HBM\uc5d0, Cold expert \uac00\uc911\uce58\ub294 Host Memory\uc5d0 \ub454\ub2e4.\" style=\"left:4.000000%;top:92.321429%;width:92.000000%;height:3.035714%;font-size:1.56250cqi;font-weight:700;color:#235C3A;justify-content:center;text-align:center;\"><span>Hot expert \uac00\uc911\uce58\ub294 GPU HBM\uc5d0, Cold expert \uac00\uc911\uce58\ub294 Host Memory\uc5d0 \ub454\ub2e4.<\/span><\/div>\n<div class=\"moe-hc-label\" data-label=\"\ub17c\ub9ac\uc801 \uc2e4\ud589 \uad6c\uc870 \u00b7 \uac15\uc870 \uc21c\uc11c\uc640 \uc560\ub2c8\uba54\uc774\uc158 \uc18d\ub3c4\ub294 \uc2e4\uc81c \uc5f0\uc0b0 \uc2dc\uac04\uc774\ub098 \uc2e4\ud589 \uc911\ucca9\uc758 \uce21\uc815\uac12\uc774 \uc544\ub2c8\ub2e4.\" style=\"left:2.625000%;top:96.517857%;width:94.375000%;height:2.410714%;font-size:1.09375cqi;font-weight:400;color:#737373;justify-content:flex-start;text-align:left;\"><span>\ub17c\ub9ac\uc801 \uc2e4\ud589 \uad6c\uc870 \u00b7 \uac15\uc870 \uc21c\uc11c\uc640 \uc560\ub2c8\uba54\uc774\uc158 \uc18d\ub3c4\ub294 \uc2e4\uc81c \uc5f0\uc0b0 \uc2dc\uac04\uc774\ub098 \uc2e4\ud589 \uc911\ucca9\uc758 \uce21\uc815\uac12\uc774 \uc544\ub2c8\ub2e4.<\/span><\/div>\n<div aria-hidden=\"true\" class=\"moe-route\" data-route=\"checkpoint-convert\" data-source=\"fp8-checkpoint\" data-target=\"amxint4-checkpoint\" style=\"--moe-route-color:#A363CC;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:13.125000%;top:38.214286%;width:2.120000%;height:0.125000cqi;transform:translateY(-50%) rotate(90.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-head moe-route-head-down\" data-head-at=\"target\" data-tip-x=\"210.0\" data-tip-y=\"472.0\" style=\"left:13.125000%;top:42.142857%;width:0.717500cqi;height:0.875000cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"load-cold-weights\" data-source=\"amxint4-checkpoint\" data-target=\"cold-expert-weights\" style=\"--moe-route-color:#A363CC;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:20.750000%;top:45.178571%;width:2.562500%;height:0.125000cqi;transform:translateY(-50%) rotate(0.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-segment\" data-segment=\"1\" style=\"left:23.312500%;top:45.178571%;width:3.812500%;height:0.125000cqi;transform:translateY(-50%) rotate(-90.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-segment\" data-segment=\"2\" style=\"left:23.312500%;top:39.732143%;width:1.682500%;height:0.125000cqi;transform:translateY(-50%) rotate(0.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-head moe-route-head-right\" data-head-at=\"target\" data-tip-x=\"410.0\" data-tip-y=\"445.0\" style=\"left:25.625000%;top:39.732143%;width:0.875000cqi;height:0.717500cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"numa-0-link\" data-source=\"kt-cpuinfer\" data-target=\"numa-0\" style=\"--moe-route-color:#8D96A0;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:69.875000%;top:41.696429%;width:1.562500%;height:0.081250cqi;transform:translateY(-50%) rotate(90.000deg);\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"numa-1-link\" data-source=\"kt-cpuinfer\" data-target=\"numa-1\" style=\"--moe-route-color:#8D96A0;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:86.750000%;top:41.696429%;width:1.562500%;height:0.081250cqi;transform:translateY(-50%) rotate(90.000deg);\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"read-cold-weights\" data-source=\"cold-expert-weights\" data-target=\"kt-cpuinfer\" style=\"--moe-route-color:#A363CC;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:54.500000%;top:35.267857%;width:7.245000%;height:0.125000cqi;transform:translateY(-50%) rotate(0.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-head moe-route-head-right\" data-head-at=\"target\" data-tip-x=\"998.0\" data-tip-y=\"395.0\" style=\"left:62.375000%;top:35.267857%;width:0.875000cqi;height:0.717500cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"input-attention\" data-source=\"input-hidden-state\" data-target=\"attention\" style=\"--moe-route-color:#697B90;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:14.750000%;top:66.964286%;width:1.307500%;height:0.187500cqi;transform:translateY(-50%) rotate(0.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-0.000000cqi;\"><\/span><\/span><span class=\"moe-route-head moe-route-head-right\" data-head-at=\"target\" data-tip-x=\"267.0\" data-tip-y=\"750.0\" style=\"left:16.687500%;top:66.964286%;width:0.875000cqi;height:0.717500cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"attention-router\" data-source=\"attention\" data-target=\"router-dispatch\" style=\"--moe-route-color:#697B90;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:29.562500%;top:66.964286%;width:2.807500%;height:0.187500cqi;transform:translateY(-50%) rotate(0.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-0.000000cqi;\"><\/span><\/span><span class=\"moe-route-head moe-route-head-right\" data-head-at=\"target\" data-tip-x=\"528.0\" data-tip-y=\"750.0\" style=\"left:33.000000%;top:66.964286%;width:0.875000cqi;height:0.717500cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"hot-dispatch\" data-source=\"router-dispatch\" data-target=\"gpu-hot-compute\" style=\"--moe-route-color:#24873F;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:48.250000%;top:66.964286%;width:3.197500%;height:0.312500cqi;transform:translateY(-50%) rotate(0.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-0.000000cqi;\"><\/span><\/span><span class=\"moe-route-head moe-route-head-right\" data-head-at=\"target\" data-tip-x=\"839.0\" data-tip-y=\"750.0\" style=\"left:52.437500%;top:66.964286%;width:1.375000cqi;height:1.127500cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"hot-result\" data-source=\"gpu-hot-compute\" data-target=\"result-combine\" style=\"--moe-route-color:#24873F;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:67.250000%;top:66.964286%;width:7.260000%;height:0.312500cqi;transform:translateY(-50%) rotate(0.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-0.000000cqi;\"><\/span><\/span><span class=\"moe-route-head moe-route-head-right\" data-head-at=\"target\" data-tip-x=\"1208.0\" data-tip-y=\"750.0\" style=\"left:75.500000%;top:66.964286%;width:1.375000cqi;height:1.127500cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"cold-dispatch\" data-source=\"router-dispatch\" data-target=\"kt-cpuinfer\" style=\"--moe-route-color:#1688A4;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:40.687500%;top:63.392857%;width:6.750000%;height:0.312500cqi;transform:translateY(-50%) rotate(-90.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-0.000000cqi;\"><\/span><\/span><span class=\"moe-route-segment\" data-segment=\"1\" style=\"left:40.687500%;top:53.750000%;width:37.687500%;height:0.312500cqi;transform:translateY(-50%) rotate(0.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-6.750000cqi;\"><\/span><\/span><span class=\"moe-route-segment\" data-segment=\"2\" style=\"left:78.375000%;top:53.750000%;width:7.322500%;height:0.312500cqi;transform:translateY(-50%) rotate(-90.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-44.437500cqi;\"><\/span><\/span><span class=\"moe-route-head moe-route-head-up\" data-head-at=\"target\" data-tip-x=\"1254.0\" data-tip-y=\"469.0\" style=\"left:78.375000%;top:41.875000%;width:1.127500cqi;height:1.375000cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"cold-result\" data-source=\"kt-cpuinfer\" data-target=\"result-combine\" style=\"--moe-route-color:#357F64;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:94.250000%;top:37.232143%;width:2.375000%;height:0.312500cqi;transform:translateY(-50%) rotate(0.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-0.000000cqi;\"><\/span><\/span><span class=\"moe-route-segment\" data-segment=\"1\" style=\"left:96.625000%;top:37.232143%;width:20.812500%;height:0.312500cqi;transform:translateY(-50%) rotate(90.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-2.375000cqi;\"><\/span><\/span><span class=\"moe-route-segment\" data-segment=\"2\" style=\"left:96.625000%;top:66.964286%;width:1.885000%;height:0.312500cqi;transform:translateY(-50%) rotate(180.000deg);\"><span class=\"moe-route-packets\" style=\"--moe-route-phase:-23.187500cqi;\"><\/span><\/span><span class=\"moe-route-head moe-route-head-left\" data-head-at=\"target\" data-tip-x=\"1500.0\" data-tip-y=\"750.0\" style=\"left:93.750000%;top:66.964286%;width:1.375000cqi;height:1.127500cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"residual-input\" data-source=\"attention\" data-target=\"result-combine\" style=\"--moe-route-color:#8C99A7;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:31.312500%;top:66.964286%;width:4.437500%;height:0.125000cqi;transform:translateY(-50%) rotate(90.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-segment\" data-segment=\"1\" style=\"left:31.312500%;top:73.303571%;width:58.500000%;height:0.125000cqi;transform:translateY(-50%) rotate(0.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-segment\" data-segment=\"2\" style=\"left:89.812500%;top:73.303571%;width:1.182500%;height:0.125000cqi;transform:translateY(-50%) rotate(-90.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-head moe-route-head-up\" data-head-at=\"target\" data-tip-x=\"1437.0\" data-tip-y=\"792.0\" style=\"left:89.812500%;top:70.714286%;width:0.717500cqi;height:0.875000cqi;\"><\/span><\/div><div aria-hidden=\"true\" class=\"moe-route\" data-route=\"load-gpu-weights\" data-source=\"fp8-checkpoint\" data-target=\"gpu-0\" style=\"--moe-route-color:#A363CC;\"><span class=\"moe-route-segment\" data-segment=\"0\" style=\"left:5.750000%;top:35.446429%;width:2.437500%;height:0.125000cqi;transform:translateY(-50%) rotate(180.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-segment\" data-segment=\"1\" style=\"left:3.312500%;top:35.446429%;width:33.000000%;height:0.125000cqi;transform:translateY(-50%) rotate(90.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-segment\" data-segment=\"2\" style=\"left:3.312500%;top:82.589286%;width:2.057500%;height:0.125000cqi;transform:translateY(-50%) rotate(0.000deg);background:repeating-linear-gradient(90deg,var(--moe-route-color) 0 .4cqi,transparent .4cqi .7cqi);\"><\/span><span class=\"moe-route-head moe-route-head-right\" data-head-at=\"target\" data-tip-x=\"96.0\" data-tip-y=\"925.0\" style=\"left:6.000000%;top:82.589286%;width:0.875000cqi;height:0.717500cqi;\"><\/span><\/div><\/div><\/div><\/div><\/details><\/figure><div class=\"moe-prose\">\n<h3>\uac00\uc911\uce58 \uc801\uc7ac\uc640 \ucd94\ub860 \uc911 \ub370\uc774\ud130 \uc774\ub3d9\uc744 \uad6c\ubcc4\ud55c\ub2e4<\/h3>\n<p>\uae30\ub3d9\ud560 \ub54c\ub294 \uc6d0\ubcf8 \uccb4\ud06c\ud3ec\uc778\ud2b8\uc5d0\uc11c GPU\uc6a9 \uac00\uc911\uce58\ub97c, \ubcc0\ud658\ub41c \uc804\ubb38\uac00 \ud30c\uc77c\uc5d0\uc11c CPU\uc6a9 \uac00\uc911\uce58\ub97c \uc77d\ub294\ub2e4. \uc2e4\ud5d8\uc758 GPU \uc804\ubb38\uac00 \uacbd\ub85c\ub294 8\ube44\ud2b8 \ubd80\ub3d9\uc18c\uc218\uc810\uc778 <strong>FP8<\/strong>, CPU \uacbd\ub85c\ub294 4\ube44\ud2b8 \uc815\uc218 \uc591\uc790\ud654\uc778 <strong>INT4<\/strong>\ub97c \uc0ac\uc6a9\ud55c\ub2e4. \uc815\ubc00\ub3c4 \ud45c\uae30\ub294 \uc804\ubb38\uac00 \uac00\uc911\uce58 \uacbd\ub85c\ub97c \uac00\ub9ac\ud0a4\uba70, \ubaa8\ub4e0 \ubaa8\ub4c8\uc774 \uac19\uc740 \uc815\ubc00\ub3c4\ub85c \uacc4\uc0b0\ub41c\ub2e4\ub294 \ub73b\uc740 \uc544\ub2c8\ub2e4. <a aria-label=\"\uc2e4\ud5d8 \uad6c\uc131 \uadfc\uac70\" href=\"#moe-source-test\">[2]<\/a><\/p>\n<p>\uadf8\ub9bc\uc758 \uc810\uc120\uc740 \uac00\uc911\uce58 \uc801\uc7ac\u00b7\ucc38\uc870\uc640 \ubcc4\ub3c4\ub85c \ud45c\uc2dc\ud55c \uc794\ucc28 \uc785\ub825\uc774\uace0, \uc0c9 \uc2e4\uc120\uc740 \ucd94\ub860 \uc911 \ud65c\uc131\uac12\u00b7\uc120\ud0dd \uc815\ubcf4\u00b7\uacb0\uacfc\uc758 \uc774\ub3d9\uc774\ub2e4. CPU\uc640 GPU\ub294 \ubcc4\ub3c4 \uc11c\ubc84\uac00 \uc544\ub2c8\ub77c <strong>\ub3d9\uc77c\ud55c \ubb3c\ub9ac \uc11c\ubc84\uc758 \uad6c\uc131\uc694\uc18c<\/strong>\ub2e4.<\/p>\n<p><strong>\u2460 \uc785\ub825 \uc900\ube44.<\/strong> GPU\uac00 Attention\uc744 \uc218\ud589\ud558\uace0 \uc804\ubb38\uac00 \uc5f0\uc0b0\uc5d0 \ud544\uc694\ud55c \ud65c\uc131\uac12\uc744 \uc900\ube44\ud55c\ub2e4. Attention\uc758 KV cache\ub294 GPU \uba54\ubaa8\ub9ac\uc5d0\uc11c \uc811\uadfc\ud55c\ub2e4.<\/p>\n<p><strong>\u2461 \uc120\ud0dd\uacfc \ubd84\uae30.<\/strong> Router\uac00 \uc0ac\uc6a9\ud560 \uc804\ubb38\uac00 ID\uc640 \ub77c\uc6b0\ud305 \uac00\uc911\uce58\ub97c \uacc4\uc0b0\ud55c\ub2e4. \uc2e4\ud589\uae30\ub294 \ubc30\uce58\ud45c\ub97c \uc870\ud68c\ud574 \uc120\ud0dd\ub41c \uc804\ubb38\uac00\ub97c GPU \uacbd\ub85c\uc640 CPU \uacbd\ub85c\ub85c \ub098\ub208\ub2e4. Router\uc758 \uc120\ud0dd \uc815\ubcf4\uc640 \uc804\ubb38\uac00\uc5d0 \uc785\ub825\ud560 \ud65c\uc131\uac12\uc740 \uc11c\ub85c \ub2e4\ub978 \ub370\uc774\ud130\ub2e4.<\/p>\n<p><strong>\u2462 \uc804\ubb38\uac00 \uacc4\uc0b0.<\/strong> GPU\ub294 \uc790\uc2e0\uc758 \uba54\ubaa8\ub9ac\uc5d0 \uc788\ub294 Hot expert\ub97c \uacc4\uc0b0\ud55c\ub2e4. CPU\ub294 \uc804\ub2ec\ubc1b\uc740 \uc785\ub825\uc73c\ub85c Host Memory\uc758 Cold expert\ub97c \uacc4\uc0b0\ud55c\ub2e4. CPU \ucabd \uc804\ubb38\uac00 \uc2e4\ud589\uae30\ub294 96\uac1c \uc6cc\ucee4\ub97c \ub450 \uc2a4\ub808\ub4dc \ud480\uc5d0 \ub098\ub204\uc5b4 \uc0ac\uc6a9\ud55c\ub2e4. \uc18c\ucf13\ub9c8\ub2e4 \uac00\uae4c\uc6b4 \uba54\ubaa8\ub9ac\uc640 \ucf54\uc5b4\ub97c \ud65c\uc6a9\ud558\ub294 NUMA \ubc30\uce58\ub3c4 \uc774 \uacbd\ub85c\uc5d0 \uc18d\ud55c\ub2e4.<\/p>\n<p><strong>\u2463 \uacb0\uacfc \uacb0\ud569.<\/strong> CPU\uc5d0\uc11c \ub3cc\uc544\uc628 \ubd80\ubd84 \uacb0\uacfc\uc640 GPU\uc758 Hot expert \ubd80\ubd84 \uacb0\uacfc\ub97c GPU\uc5d0\uc11c \uacb0\ud569\ud55c\ub2e4. \uac01 \uc804\ubb38\uac00\uc758 \ub77c\uc6b0\ud305 \uac00\uc911\uce58\uc640 \uae30\uc5ec\ubd84\uc744 \uc911\ubcf5 \uc801\uc6a9\ud558\uc9c0 \uc54a\uc544\uc57c \ud558\uba70, \uc794\ucc28 \uc785\ub825\uc744 \ub354\ud55c \ub4a4 \ub2e4\uc74c \uce35\uc73c\ub85c \uc9c4\ud589\ud55c\ub2e4.<\/p>\n<div class=\"moe-callout\"><strong>Hot\u00b7Cold\uc758 \uc800\uc7a5 \uc704\uce58\ub9cc \ub098\ub204\uc5b4\uc11c\ub294 \uc131\ub2a5\uc774 \ub098\uc624\uc9c0 \uc54a\ub294\ub2e4.<\/strong><br>\ub450 \uc7a5\uce58\uc758 \ub3c5\ub9bd\uc801\uc778 \uacc4\uc0b0\uc744 \uacb9\uccd0 \uc2e4\ud589\ud560 \uc218 \uc788\uc5b4\ub3c4, \ub2e4\uc74c \uc5f0\uc0b0\uc774 \ud544\uc694\ub85c \ud558\ub294 \uacb0\uacfc\uac00 \ubaa8\uc774\ub294 \uc9c0\uc810\uc740 \ub0a8\ub294\ub2e4. \uc804\uc1a1\uc744 \ube44\ub3d9\uae30\ub85c \ub9cc\ub4e4\uc5c8\ub2e4\uace0 \uc774 \uc758\uc874 \uad00\uacc4\uac00 \uc5c6\uc5b4\uc9c0\ub294 \uac83\uc740 \uc544\ub2c8\ub2e4.<\/div>\n<\/div><div class=\"moe-prose\">\n<h2 id=\"moe-experiment\">4. Hot expert\ub97c GPU\uc5d0 \ub454 \uae30\ubcf8 \uad6c\uc131\uc758 \uc2e4\ud5d8 \uacb0\uacfc<\/h2>\n<p>\uce21\uc815 \ub300\uc0c1\uc740 <strong>Qwen3-Coder-480B-A35B-Instruct-FP8<\/strong>\uc774\ub2e4. \uac19\uc740 \uc11c\ubc84\uc5d0\uc11c H100 80 GB 4\uc7a5\uacfc Xeon Platinum 8480+ 2\uac1c, DDR5 2 TB\ub97c \uc0ac\uc6a9\ud588\ub2e4. <strong>\uac01 \uce35\uc758 160\uac1c \uc804\ubb38\uac00 \uac00\uc6b4\ub370 Hot expert 96\uac1c\ub97c GPU\uc5d0, \ub098\uba38\uc9c0 64\uac1c\ub97c Host Memory\uc5d0 \ubc30\uce58<\/strong>\ud588\ub2e4. GPU \ub2f4\ub2f9 \uc804\ubb38\uac00\uc758 \uac00\uc911\uce58\ub294 4\uc7a5\uc5d0 \ud150\uc11c \ubcd1\ub82c\ub85c \ub098\ub204\uace0, CPU\uac00 Cold expert\ub97c \uc9c1\uc811 \uacc4\uc0b0\ud55c\ub2e4. <a aria-label=\"\ubc30\uce58\uc640 \uce21\uc815 \uadfc\uac70\" href=\"#moe-source-test\">[2]<\/a><\/p>\n<p>\uc774 \uad6c\uc131\uc740 \uce35\ubcc4 \uc804\ubb38\uac00 \uc218 \uc7ac\ubc30\ubd84, \uc791\uc5c5 \uc804\ub2ec \ubc29\uc2dd \uac1c\uc120, \ucd94\uac00 CPU \ucee4\ub110 \ucd5c\uc801\ud654\uc758 \ucd9c\ubc1c\uc810\uc73c\ub85c \uce21\uc815\ud55c <strong>\uae30\ubcf8 \ubd80\ubd84 \uc0c1\uc8fc \uad6c\uc131<\/strong>\uc774\ub2e4. \uc774\ubbf8 \uc120\ud0dd \ube48\ub3c4\ub97c \ubc18\uc601\ud55c \uc804\ubb38\uac00 \ubc30\uce58\uc640 \uc544\ub798 \uc2e4\ud589 \uc124\uc815\uc744 \uc0ac\uc6a9\ud558\ubbc0\ub85c, \uc544\ubb34 \uc124\uc815\ub3c4 \uc870\uc815\ud558\uc9c0 \uc54a\uc740 \uc0c1\ud0dc\ub97c \ub73b\ud558\uc9c0\ub294 \uc54a\ub294\ub2e4.<\/p>\n<div class=\"moe-table-wrap\"><table class=\"moe-table\"><thead><tr><th style=\"width:31%\">\ud56d\ubaa9<\/th><th>\ubd80\ubd84 \uc0c1\uc8fc \uce21\uc815 \uc124\uc815<\/th><\/tr><\/thead><tbody>\n<tr><td>GPU\u00b7CPU \uc5ed\ud560<\/td><td>GPU 4\uc7a5: Hot expert\u00b7Attention\u00b7Router \/ CPU 2\uc18c\ucf13: Cold expert<\/td><\/tr>\n<tr><td>\uc804\ubb38\uac00 \ubc30\uce58<\/td><td>62\uac1c \uce35 \uac01\uac01 GPU 96\uac1c\u00b7CPU 64\uac1c \/ \ud1a0\ud070\ubcc4 \uc120\ud0dd 8\uac1c<\/td><\/tr>\n<tr><td>CPU \uc2e4\ud589<\/td><td>\uc6cc\ucee4 96\uac1c, \uc2a4\ub808\ub4dc \ud480 2\uac1c \/ CPU 2.0 GHz, \ud130\ubcf4 \ube44\ud65c\uc131<\/td><\/tr>\n<tr><td>\uc11c\ube59 \uc18c\ud504\ud2b8\uc6e8\uc5b4<\/td><td>SGLang 0.5.18 + kt-kernel 0.7.0.post1 \/ \uc2e4\ud5d8\uc6a9 \uc218\uc815 \ube4c\ub4dc<\/td><\/tr>\n<tr><td>KV cache\u00b7\uc2e4\ud589 \uc635\uc158<\/td><td>BF16 \uc815\ubc00\ub3c4, \ucd1d 40,960\ud1a0\ud070 \uc6a9\ub7c9 \/ CUDA graph \ubc30\uce58 \ud06c\uae30 \ucd5c\ub300 64 \/ \uc804\ubb38\uac00 \uc9c0\uc5f0 \ucc98\ub9ac \ud55c\ub3c4 \ud1a0\ud070\ub2f9 4<\/td><\/tr>\n<\/tbody><\/table><\/div>\n<p>\uc785\ub825\uc740 Sonnet \ub370\uc774\ud130\uc14b\uc73c\ub85c <strong>512\ud1a0\ud070 \uc785\ub825\u00b7128\ud1a0\ud070 \ucd9c\ub825, \ub3d9\uc2dc \uc694\uccad 64, \ucd1d 256\uc694\uccad<\/strong>\uc744 \uc124\uc815\ud588\ub2e4. \uacf5\ud1b5 \uc811\ub450\uc0ac \uae38\uc774\ub294 0\uc73c\ub85c \ub450\uace0 \ub9e4 \ubc18\ubcf5 \uc804\uc5d0 \uce90\uc2dc\ub97c \ucd08\uae30\ud654\ud588\uc73c\uba70, \uc885\ub8cc \ud1a0\ud070\uc5d0 \uc758\ud55c \uc870\uae30 \uc885\ub8cc \uc5c6\uc774 \ucd9c\ub825 \uae38\uc774\ub97c \uace0\uc815\ud588\ub2e4. \uac19\uc740 \uc870\uac74\uc5d0\uc11c 3\ud68c \ubc18\ubcf5\ud588\ub2e4.<\/p>\n<p>\ube44\uad50 \ub300\uc0c1\uc740 <strong>\ub3d9\uc77c \ubaa8\ub378\uc758 GPU 8\uc7a5 \uc804\uc6a9 \uc2e4\ud589<\/strong>\uc774\ub2e4. CPU \uc804\ubb38\uac00 \uacc4\uc0b0 \uc5c6\uc774 GPU\uc5d0\uc11c \ubaa8\ub4e0 \uc120\ud0dd \uc804\ubb38\uac00\ub97c \ucc98\ub9ac\ud558\uba70, \ud150\uc11c \ubcd1\ub82c\uacfc \uc804\ubb38\uac00 \ubcd1\ub82c\uc744 \uc0ac\uc6a9\ud588\ub2e4. \uce90\uc2dc \uc6a9\ub7c9\uc740 131,072\ud1a0\ud070\uc774\uc5c8\ub2e4. \ub450 \uad6c\uc131 \ubaa8\ub450 SGLang\uc73c\ub85c \uc11c\ube59\ud558\uace0, \uac19\uc740 \uc694\uccad \uc870\uac74\uc5d0\uc11c \uce21\uc815\ud588\ub2e4.<\/p>\n<div aria-label=\"Qwen3-Coder-480B \ucd9c\ub825 \ucc98\ub9ac\ub7c9. GPU 4\uc7a5\uacfc CPU\uc758 \ubd80\ubd84 \uc0c1\uc8fc \uae30\ubcf8 \uad6c\uc131\uc740 \ucd08\ub2f9 612.11\ud1a0\ud070, GPU 8\uc7a5 \uc804\uc6a9 \uc2e4\ud589\uc740 \ucd08\ub2f9 1703.56\ud1a0\ud070. \uac01 3\ud68c \ud3c9\uade0\uc774\uba70 \uc0ac\uc6a9 GPU \uc218\uac00 \ub2e4\ub974\ub2e4.\" class=\"moe-bar-chart\" id=\"moe-baseline-chart\" role=\"img\">\n<p style=\"margin:0 0 16px;font-weight:700;\">\ucd9c\ub825 \ucc98\ub9ac\ub7c9 \u00b7 \uac19\uc740 \ubaa8\ub378\uacfc \uc694\uccad \uc870\uac74<\/p>\n<div class=\"moe-bar-row\"><div>GPU 4\uc7a5 + CPU<br><small>Hot\u00b7Cold \ubd84\ub2f4<\/small><\/div><div class=\"moe-bar-track\"><span style=\"width:34.0061%;background:#28758b;\"><\/span><\/div><div class=\"moe-bar-value\">612.11<\/div><\/div>\n<div class=\"moe-bar-row\"><div>GPU 8\uc7a5<br><small>GPU \uc804\uc6a9 \uc2e4\ud589<\/small><\/div><div class=\"moe-bar-track\"><span style=\"width:94.6422%;background:#667a91;\"><\/span><\/div><div class=\"moe-bar-value\">1,703.56<\/div><\/div>\n<div class=\"moe-bar-scale\">\ub2e8\uc704: \ucd9c\ub825 \ud1a0\ud070\/\ucd08 \u00b7 \uac01 3\ud68c \ud3c9\uade0 \u00b7 \uc120\ud615 \ub208\uae08 0\u20131,800 \u00b7 GPU \uc218\uc640 \uc2e4\ud589 \uad6c\uc131\uc774 \ub2e4\ub978 \ube44\uad50<\/div>\n<\/div>\n<div class=\"moe-table-wrap\"><table class=\"moe-table\"><thead><tr><th style=\"width:38%\">\uc9c0\ud45c<\/th><th>GPU 4\uc7a5 + CPU<br>\ubd80\ubd84 \uc0c1\uc8fc \uae30\ubcf8 \uad6c\uc131<\/th><th>GPU 8\uc7a5<br>GPU \uc804\uc6a9 \uc2e4\ud589<\/th><\/tr><\/thead><tbody>\n<tr><td>\ucd9c\ub825 \ucc98\ub9ac\ub7c9 \u00b7 \ud3c9\uade0 \u00b1 \ud45c\ubcf8\ud45c\uc900\ud3b8\ucc28<\/td><td><strong>612.11 \u00b1 6.76 \ud1a0\ud070\/\ucd08<\/strong><\/td><td><strong>1,703.56 \u00b1 4.93 \ud1a0\ud070\/\ucd08<\/strong><\/td><\/tr>\n<tr><td>\uccab \ud1a0\ud070\uae4c\uc9c0\uc758 \uc2dc\uac04 \u00b7 95\ubc31\ubd84\uc704\uc218<\/td><td>4,195 ms<\/td><td>1,464 ms<\/td><\/tr>\n<tr><td>\uccab \ud1a0\ud070 \uc774\ud6c4 \ud1a0\ud070\ub2f9 \uc0dd\uc131 \uc2dc\uac04 \u00b7 95\ubc31\ubd84\uc704\uc218<\/td><td>91.0 ms<\/td><td>32.8 ms<\/td><\/tr>\n<tr><td>\uc644\ub8cc \/ \uc2e4\ud328 \uc694\uccad \u00b7 \ubc18\ubcf5\ub9c8\ub2e4<\/td><td>256 \/ 0<\/td><td>256 \/ 0<\/td><\/tr>\n<\/tbody><\/table><\/div>\n<p class=\"moe-note\">\ucc98\ub9ac\ub7c9\uc758 \u00b1\ub294 3\ud68c \ubc18\ubcf5\uc758 \ud45c\ubcf8\ud45c\uc900\ud3b8\ucc28\ub2e4. \uc9c0\uc5f0 \uc218\uce58\ub294 \uac01 \ubc18\ubcf5\uc5d0\uc11c \uad6c\ud55c \uc694\uccad\ubcc4 95\ubc31\ubd84\uc704\uc218\ub97c \ud3c9\uade0\ud55c \uac12\uc774\uba70, \uc138 \ubc18\ubcf5\uc758 \ubaa8\ub4e0 \uc694\uccad\uc744 \ud569\uccd0 \ub2e4\uc2dc \uacc4\uc0b0\ud55c \ubc31\ubd84\uc704\uc218\uac00 \uc544\ub2c8\ub2e4. 95\ubc31\ubd84\uc704\uc218\ub294 \ub290\ub9b0 \ucabd 5%\uc758 \uacbd\uacc4\ub97c \ub098\ud0c0\ub0b8\ub2e4. <a href=\"#moe-source-test\">[2]<\/a><\/p>\n<h3>\uc774 \uacb0\uacfc\uac00 \ubcf4\uc5ec \uc8fc\ub294 \uac83<\/h3>\n<p>Hot\u00b7Cold \ubd84\ub2f4 \uad6c\uc131\uc740 GPU 4\uc7a5\uc73c\ub85c \ubaa8\ub378\uc744 \uc2e4\ud589\ud574 \uc694\uccad\uc744 \uc644\ub8cc\ud588\uc9c0\ub9cc, <strong>GPU 8\uc7a5 \uc804\uc6a9 \uc2e4\ud589\ubcf4\ub2e4 \ucc98\ub9ac\ub7c9\uc774 \ub0ae\uace0 \uc694\uccad \uc9c0\uc5f0\uc774 \uae38\uc5c8\ub2e4.<\/strong> \ud544\uc694\ud55c GPU \uc218\ub97c \uc904\uc778 \uac83\uacfc \uac19\uc740 \uc751\ub2f5 \uc18d\ub3c4\ub97c \ud655\ubcf4\ud55c \uac83\uc740 \ubcc4\uac1c\uc758 \uacb0\uacfc\ub2e4.<\/p>\n<p>\uc774 \ucc28\uc774\ub97c CPU \uc624\ud504\ub85c\ub529\ub9cc\uc758 \uc190\uc2e4\ub85c \ub2e8\uc815\ud560 \uc218\ub294 \uc5c6\ub2e4. GPU \uc218, \uc804\ubb38\uac00 \uc815\ubc00\ub3c4, \ubcd1\ub82c \uad6c\uc131\uacfc \uce90\uc2dc \uc6a9\ub7c9\uc774 \ud568\uaed8 \ub2e4\ub974\ub2e4. \ub610\ud55c 612.11\ud1a0\ud070\/\ucd08\ub294 \ubaa8\ub4e0 \uc694\uccad\uc774 \uc0dd\uc131\ud55c \ud1a0\ud070\uc744 \ud569\uce5c \ucc98\ub9ac\ub7c9\uc774\uc9c0, \uc0ac\uc6a9\uc790 \ud55c \uba85\uc758 \uc0dd\uc131 \uc18d\ub3c4\uac00 \uc544\ub2c8\ub2e4. \uc694\uccad \uc644\ub8cc\ub9cc\uc73c\ub85c \ucd9c\ub825 \uc815\ud655\uc131\uc774\ub098 \uc7a5\uc2dc\uac04 \uc6b4\uc601 \uc548\uc815\uc131\uc774 \uac80\uc99d\ub418\ub294 \uac83\ub3c4 \uc544\ub2c8\ub2e4.<\/p>\n<p>\uc774 \uce21\uc815\uc740 Hot\u00b7Cold \ubd84\ub2f4\uc774 \uc791\ub3d9\ud558\ub294 \uae30\uc900\uc810\uc774\ub2e4. \ub2e4\uc74c \uc9c8\ubb38\uc740 <strong>\uac19\uc740 GPU 4\uc7a5\uacfc \ucd9c\ub825 \uc870\uac74\uc744 \uc720\uc9c0\ud558\uba74\uc11c \uc5b4\ub290 \ubd80\ubd84\uc758 \ube44\uc6a9\uc744 \uc904\uc77c \uac83\uc778\uac00<\/strong>\ub2e4.<\/p>\n<\/div><div class=\"moe-prose\">\n<h2 id=\"moe-measurement\">5. \uc0ac\uc6a9\ub960\ubcf4\ub2e4 \uc2e4\ud589 \uad6c\uac04\uc744 \ub098\ub204\uc5b4 \ubd10\uc57c \ud55c\ub2e4<\/h2>\n<p>CPU \uc0ac\uc6a9\ub960\uc774 \ub192\ub2e4\uace0 \uc720\ud6a8\ud55c \uc804\ubb38\uac00 \uc5f0\uc0b0\uc73c\ub85c \ud3ec\ud654\ub41c \uac83\uc740 \uc544\ub2c8\ub2e4. \uc791\uc5c5 \uc804\ub2ec\uacfc \uc644\ub8cc \ud655\uc778, \ud3f4\ub9c1\uc5d0\ub3c4 CPU \uc2dc\uac04\uc774 \uc0ac\uc6a9\ub41c\ub2e4. \ub17c\ub9ac \uc2a4\ub808\ub4dc \uc804\uccb4\uc758 \ud3c9\uade0\uc740 \uc2e4\uc81c\ub85c \uacc4\uc0b0\ud558\ub294 \ucf54\uc5b4\uc758 \uc0c1\ud0dc\ub97c \uac00\ub9b4 \uc218 \uc788\ub2e4. GPU \uc0ac\uc6a9\ub960 \uc5ed\uc2dc Cold \uacb0\uacfc\ub97c \uae30\ub2e4\ub9b0 \uc2dc\uac04\uacfc GPU \uc790\uccb4 \uc5f0\uc0b0 \uc2dc\uac04\uc744 \uc9c1\uc811 \ubd84\ub9ac\ud574 \uc8fc\uc9c0\ub294 \uc54a\ub294\ub2e4.<\/p>\n<div class=\"moe-formula\"><strong>\ubd84\ub9ac\ud574\uc11c \uce21\uc815\ud560 \uacbd\ub85c<\/strong><br>\uacf5\ud1b5: GPU \uc785\ub825 \uc900\ube44\u00b7Router \uc120\ud0dd<br>Hot: GPU \uc804\ubb38\uac00 \uacc4\uc0b0<br>Cold: \uc785\ub825 \uc804\ub2ec \u2192 CPU \uc804\ubb38\uac00 \uacc4\uc0b0 \u2192 \uacb0\uacfc \ubc18\ud658<br>\ud569\ub958: \ub450 \uacbd\ub85c \uc644\ub8cc \ud655\uc778 \u2192 \uacb0\uacfc \uacb0\ud569\u00b7\ub2e4\uc74c \uce35 \uc9c4\ud589<\/div>\n<p>\ud655\uc778\ud560 \uac83\uc740 CPU\uac00 \uba54\ubaa8\ub9ac\ub97c \uc77d\ub290\ub77c \ub2a6\ub294\uc9c0, \uc791\uc740 \uc791\uc5c5\uc744 \ub118\uae30\ub294 \uad00\ub9ac \ube44\uc6a9\uc774 \ud070\uc9c0, \uc774\ubbf8 \uc900\ube44\ub41c \uacb0\uacfc\ub97c GPU\uac00 \ub2a6\uac8c \ubc1b\uc544 \uac00\ub294\uc9c0\ub2e4. \uac19\uc740 \uc774\uc720\ub85c GPU \uba54\ubaa8\ub9ac \ucd1d\uc0ac\uc6a9\ub7c9\ub9cc \ubcf4\uc9c0 \ub9d0\uace0 Hot expert \uac00\uc911\uce58, \uce90\uc2dc, \uc2e4\ud589 \ubc84\ud37c\uc758 \uc810\uc720\ub97c \uad6c\ubd84\ud574\uc57c \ud55c\ub2e4. GPU \uc804\ub825\uc774 \ub0ae\uc544\uc9c0\ub354\ub77c\ub3c4 CPU \uc5f0\uc0b0\uacfc \uc804\uccb4 \uc2e4\ud589 \uc2dc\uac04\uc774 \ub298\uc5c8\ub2e4\uba74 \uc5d0\ub108\uc9c0 \ud6a8\uc728\uc774 \uc88b\uc544\uc84c\ub2e4\uace0 \ub2e8\uc815\ud560 \uc218 \uc5c6\ub2e4.<\/p>\n<div aria-label=\"\uce21\uc815 \ubc0f \ubaa8\ub378 \uad6c\uc131 \ucd9c\ucc98\" class=\"moe-source\">\n<p id=\"moe-source-model\" style=\"margin:0 0 8px;\">[1] Qwen3-Coder-480B \ubaa8\ub378 \uc124\uc815 \ubc0f \uc804\ubb38\uac00 \ubc30\uce58 \uae30\ub85d, 2026-09-15. 62\uac1c \uce35\u00b7\uce35\ub2f9 160\uac1c \uc804\ubb38\uac00\u00b7\ud1a0\ud070\ubcc4 \uc120\ud0dd 8\uac1c.<\/p>\n<p id=\"moe-source-test\" style=\"margin:0;\">[2] Qwen3-Coder-480B \ubd80\ubd84 \uc0c1\uc8fc\u00b7GPU \uc804\uc6a9 \ube44\uad50 \uc2e4\ud5d8 \uae30\ub85d, 2026-09-16. \uc2e4\ud589 \uba85\ub839, \ubc18\ubcf5\ubcc4 \uc131\ub2a5\uacfc \uc9d1\uacc4, \uce90\uc2dc \uc815\ucc45 \ubc0f \uc9c0\ud45c \uacc4\uc0b0 \uaddc\uce59. \ub9e4 \ubc18\ubcf5\uc758 \uc2e4\uc81c \uc785\ub825\u00b7\ucd9c\ub825 \ud569\uacc4\ub294 \ub450 \uad6c\uc131 \ubaa8\ub450 129,237 \/ 32,768\ud1a0\ud070\uc774\ub2e4.<\/p>\n<\/div>\n<h2 id=\"moe-optimization\">6. \ucd5c\uc801\ud654\ub294 \ubc30\uce58\u00b7\uc804\ub2ec\u00b7\uacc4\uc0b0\uc744 \ud568\uaed8 \ub2e4\ub8ec\ub2e4<\/h2>\n<p>\ubaa9\ud45c\ub294 CPU\ub97c \ub354 \ubc14\uc058\uac8c \ub9cc\ub4dc\ub294 \uac83\uc774 \uc544\ub2c8\ub77c, <strong>\ucd9c\ub825\uc758 \uc815\ud655\uc131\uc744 \uc720\uc9c0\ud558\uba74\uc11c Cold \uacbd\ub85c\uc758 \uacc4\uc0b0\ub7c9\uacfc \uacb0\uacfc \ub300\uae30\ub97c \uc904\uc774\ub294 \uac83<\/strong>\uc774\ub2e4. \uae30\ubcf8 \uad6c\uc131\uc5d0\uc11c \ud655\uc778\ud560 \ucd5c\uc801\ud654 \uc9c0\uc810\uc740 \ub2e4\uc74c\uacfc \uac19\ub2e4.<\/p>\n<h3>\uac19\uc740 GPU \uacf5\uac04\uc73c\ub85c \ub354 \ud6a8\uacfc\uc801\uc778 \uc804\ubb38\uac00\ub97c \ubc30\uce58\ud55c\ub2e4<\/h3>\n<p>\ubaa8\ub4e0 \uce35\uc5d0 96\uac1c\uc529 \ub3d9\uc77c\ud558\uac8c \ub0a8\uae30\ub294 \ub300\uc2e0, \uc2e4\uc81c \uc694\uccad\uc758 \uce35\ubcc4 \uc120\ud0dd \ube48\ub3c4\uc640 Cold \uacbd\ub85c \uc9c0\uc5f0\uc744 \uc0b4\ud3b4\ubcfc \uc218 \uc788\ub2e4. \uc790\uc8fc \uc0ac\uc6a9\ub418\uac70\ub098 CPU\uc5d0\uc11c \uacc4\uc0b0\ud560 \ub54c \ub2e4\uc74c \ub2e8\uacc4\ub97c \uc624\ub798 \uae30\ub2e4\ub9ac\uac8c \ud558\ub294 \uc804\ubb38\uac00\uc5d0 GPU \uacf5\uac04\uc744 \uc6b0\uc120 \ubc30\ubd84\ud558\ub294 \ubc29\uc2dd\uc774\ub2e4. \ub2e8\uc21c \uc801\uc911\ub960\ubcf4\ub2e4 <strong>\uc2e4\uc81c \uacb0\uacfc \ub300\uae30\uc2dc\uac04\uc774 \uc5bc\ub9c8\ub098 \uc904\uc5c8\ub294\uc9c0<\/strong>\ub97c \ud655\uc778\ud574\uc57c \ud55c\ub2e4.<\/p>\n<h3>Hot expert\uc640 KV cache\uc758 \uacf5\uac04\uc744 \ud568\uaed8 \uc870\uc815\ud55c\ub2e4<\/h3>\n<p>Hot expert \uc218\ub97c \ub298\ub9ac\uba74 CPU \uacc4\uc0b0\uc740 \uc904\uc5b4\ub4e4 \uc218 \uc788\uc9c0\ub9cc \uce90\uc2dc\uc640 \uc2e4\ud589 \ubc84\ud37c\uc758 \uc5ec\uc720\ub3c4 \uc904\uc5b4\ub4e0\ub2e4. \ud544\uc694\ud55c \ubb38\ub9e5 \uae38\uc774\uc640 \ub3d9\uc2dc \uc694\uccad\uc744 \uc720\uc9c0\ud560 \uc218 \uc788\ub294 \ubc94\uc704\uc5d0\uc11c \uc870\uc815\ud558\uace0, \ucc98\ub9ac\ub7c9\ubfd0 \uc544\ub2c8\ub77c \uccab \ud1a0\ud070 \uc9c0\uc5f0\uacfc \uc0dd\uc131 \uc911 \uc9c0\uc5f0\uc744 \ud568\uaed8 \ube44\uad50\ud55c\ub2e4. \uce90\uc2dc \ubd80\uc871\uc73c\ub85c \ub300\uae30\ub098 \uc7ac\ucc98\ub9ac\uac00 \ub298\uba74 \uc804\ubb38\uac00\ub97c \ub354 \uc62c\ub9b0 \uc774\ub4dd\uc774 \uc0ac\ub77c\uc9c8 \uc218 \uc788\ub2e4.<\/p>\n<h3>\uc791\uc5c5 \uc804\ub2ec\uacfc \uc644\ub8cc \ud655\uc778\uc758 \ubc18\ubcf5 \ube44\uc6a9\uc744 \uc904\uc778\ub2e4<\/h3>\n<p>CPU\uc5d0 \uc791\uc740 \uc804\ubb38\uac00 \uc791\uc5c5\uc744 \uc694\uccad\ud558\ub294 \uacfc\uc815\uc774 \uc5ec\ub7ec \uce35\uc5d0\uc11c \ubc18\ubcf5\ub418\uba74 \uad00\ub9ac \ube44\uc6a9\ub3c4 \ub204\uc801\ub41c\ub2e4. \uc785\ub825\u00b7\uacb0\uacfc \ubc84\ud37c \uc7ac\uc0ac\uc6a9, \ubd88\ud544\uc694\ud55c \ucf5c\ubc31\uacfc \ud3f4\ub9c1 \uac10\uc18c, \uc2e4\uc81c \uacc4\uc0b0\ud560 \ud56d\ubaa9\uc774 \uc5c6\ub294 \uc791\uc5c5\uc758 \uc0dd\ub7b5\uc744 \uac80\ud1a0\ud560 \uc218 \uc788\ub2e4. \ub2e4\ub9cc \ube48 \uc791\uc5c5\uc744 \uc0dd\ub7b5\ud558\ub294 \uac83\uacfc \uc120\ud0dd\ub41c \uc804\ubb38\uac00 \uacc4\uc0b0\uc744 \uc0dd\ub7b5\ud558\ub294 \uac83\uc740 \ub2e4\ub974\ub2e4. \uc644\ub8cc \uc2e0\ud638\uc640 \uacb0\uacfc \ubc84\ud37c\uc758 \uc218\uba85\uc774 \uc815\ud655\ud788 \uc720\uc9c0\ub3fc\uc57c \ud55c\ub2e4.<\/p>\n<h3>CPU \ucee4\ub110\uacfc \uba54\ubaa8\ub9ac \ubc30\uce58\ub97c \uc2e4\uc81c \uc791\uc5c5 \ud06c\uae30\uc5d0 \ub9de\ucd98\ub2e4<\/h3>\n<p>\ub3d9\uc2dc \uc694\uccad\uc774 64\uac1c\uc5ec\ub3c4 \ud1a0\ud070\uc774 \uc5ec\ub7ec \uc804\ubb38\uac00\ub85c \ud769\uc5b4\uc9c0\uba74 \uac01 \uc804\ubb38\uac00\uac00 \ubc1b\ub294 \uc785\ub825 \ud589 \uc218\ub294 \uc791\uc744 \uc218 \uc788\ub2e4. \uc774 \ubd84\ud3ec\uc5d0 \ub9de\ucdb0 CPU\uc758 \ud589\ub82c\u00b7\ubca1\ud130 \uc5f0\uc0b0 \uacbd\ub85c, \uc791\uc5c5 \ubb36\uc74c\uacfc \uc6cc\ucee4 \uc218\ub97c \uc120\ud0dd\ud574\uc57c \ud55c\ub2e4. \uc18c\ucf13\ubcc4 \uba54\ubaa8\ub9ac \uc811\uadfc\uc744 \uc870\uc815\ud560 \ub54c\ub3c4 \uc6d0\uaca9 \uc811\uadfc\uc744 \uc904\uc774\ub294 \uc774\ub4dd\uacfc \uc0ac\uc6a9\ud560 \uc218 \uc788\ub294 \ucf54\uc5b4\u00b7\uba54\ubaa8\ub9ac \ucc44\ub110\uc774 \uc904\uc5b4\ub4dc\ub294 \uc190\ud574\ub97c \ud568\uaed8 \ubcf8\ub2e4.<\/p>\n<h3>\uc785\ub825 \ucc98\ub9ac\uc640 \uc0dd\uc131, \uc815\ud655\uc131\uacfc \uc548\uc815\uc131\uc744 \ub530\ub85c \uac80\uc99d\ud55c\ub2e4<\/h3>\n<p>\uae34 \uc785\ub825\uc744 \ud55c\uaebc\ubc88\uc5d0 \ucc98\ub9ac\ud558\ub294 \ub2e8\uacc4\uc640 \ud55c \ud1a0\ud070\uc529 \uc0dd\uc131\ud558\ub294 \ub2e8\uacc4\ub294 \uc791\uc5c5 \ud06c\uae30\uac00 \ub2e4\ub974\ub2e4. \uc785\ub825 \ucc98\ub9ac \ubb36\uc74c, \ub3d9\uc2dc \uc694\uccad \uc81c\ud55c, GPU \uc2e4\ud589 \uadf8\ub798\ud504 \ubc94\uc704\ub97c \uc870\uc815\ud560 \ub54c \ub450 \ub2e8\uacc4\uc758 \uc9c0\uc5f0\uc744 \ub098\ub204\uc5b4 \ubd10\uc57c \ud55c\ub2e4. \uce90\uc2dc\ub97c \uc7ac\uc0ac\uc6a9\ud55c \uc694\uccad\uacfc \uce90\uc2dc\ub97c \ucd08\uae30\ud654\ud55c \uc694\uccad\ub3c4 \ubd84\ub9ac\ud55c\ub2e4. \ubc30\uce58\u00b7\uc815\ubc00\ub3c4\u00b7\ucee4\ub110\u00b7\uc9c0\uc5f0 \ucc98\ub9ac \ubc29\uc2dd\uc744 \ubc14\uafc0 \ub54c\ub294 \ub3d9\uc77c \uc785\ub825\uc758 \uc815\ub2f5\uacfc \ucd9c\ub825 \ubcc0\ud654, \uc218\uce58 \uc624\ub958\uc640 \uc7a5\uc2dc\uac04 \uc2e4\ud589 \uc2e4\ud328\ub97c \ud655\uc778\ud574\uc57c \ud55c\ub2e4.<\/p>\n<p class=\"moe-ending\"><strong>Hot expert\ub294 GPU\uc5d0, Cold expert\ub294 Host Memory\uc5d0 \ub454\ub2e4.<\/strong> \uadf8\ub2e4\uc74c\uc758 \uc131\ub2a5\uc740 \uc5b4\ub290 \uc804\ubb38\uac00\ub97c \ub0a8\uae30\uace0, \ub450 \uc2e4\ud589 \uacbd\ub85c\uac00 \uc5bc\ub9c8\ub098 \ud6a8\uc728\uc801\uc73c\ub85c \uacc4\uc0b0\ud558\uace0 \ud569\ub958\ud558\ub294\uc9c0\uc5d0 \ub2ec\ub824 \uc788\ub2e4. \ubc30\uce58\uc640 \uce90\uc2dc \uc608\uc0b0\uc744 \uc815\ud55c \ub4a4 \uc804\ub2ec\u00b7\uacc4\uc0b0\u00b7\ub300\uae30\ub97c \uad6c\ubd84\ud574 \uc904\uc774\ub294 \uac83\uc774 CPU MoE Offloading \ucd5c\uc801\ud654\uc758 \ucd9c\ubc1c\uc810\uc774\ub2e4.<\/p>\n<\/div><\/div><\/article>\n","protected":false},"excerpt":{"rendered":"<p>LLM INFERENCE \u00b7 CPU\u2013GPU HYBRID CPU MoE Offloading:GPU \uba54\ubaa8\ub9ac\uc758 \ud55c\uacc4\ub97c \ub118\uc73c\uba74 \ubb34\uc5c7\uc774 \ub2ec\ub77c\uc9c8\uae4c Hot expert\ub294 GPU\uc5d0, Cold expert\ub294 Host Memory\uc5d0: \uc2e4\ud589 \uad6c\uc870\uc640 \uae30\ubcf8 \uc131\ub2a5, \ucd5c\uc801\ud654\uc758 \ucd9c\ubc1c\uc810 \uadf8\ub9bc \uc6c0\uc9c1\uc784 \uba48\ucda4 \ud070<span class=\"more-button\"><a href=\"http:\/\/kyunam.com\/?p=1036\" class=\"more-link\">\uc77d\uc5b4 \ubcfc\uae4c?<span class=\"screen-reader-text\">CPU MoE Offloading:GPU \uba54\ubaa8\ub9ac\uc758 \ud55c\uacc4\ub97c \ub118\uc73c\uba74 \ubb34\uc5c7\uc774 \ub2ec\ub77c\uc9c8\uae4c<\/span><\/a><\/span><\/p>\n","protected":false},"author":1,"featured_media":1041,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[72,63],"tags":[179,394,388,386,390,75,168,392,314,393,385,389,340,351,98,391,387,395,306],"class_list":["post-1036","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-gpgpu","category-parallel-computing","tag-ai","tag-aiengineering","tag-aiinfra","tag-cpuoffloading","tag-deepseek","tag-gpgpu","tag-gpu","tag-gpumemory","tag-h100","tag-heterogeneouscompute","tag-llminference","tag-llmserving","tag-mlops","tag-moe","tag-programming","tag-qwen","tag-sglang","tag-tokeneconomy","tag-vllm"],"_links":{"self":[{"href":"http:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/posts\/1036","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/kyunam.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1036"}],"version-history":[{"count":2,"href":"http:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/posts\/1036\/revisions"}],"predecessor-version":[{"id":1040,"href":"http:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/posts\/1036\/revisions\/1040"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/kyunam.com\/index.php?rest_route=\/wp\/v2\/media\/1041"}],"wp:attachment":[{"href":"http:\/\/kyunam.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1036"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/kyunam.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1036"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/kyunam.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1036"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}