{"id":1084,"date":"2026-03-23T06:31:31","date_gmt":"2026-03-23T06:31:31","guid":{"rendered":"https:\/\/ouyangminwei.com\/?p=1084"},"modified":"2026-03-23T06:50:30","modified_gmt":"2026-03-23T06:50:30","slug":"llm-casual-att","status":"publish","type":"post","link":"https:\/\/ouyangminwei.com\/index.php\/2026\/03\/23\/llm-casual-att\/","title":{"rendered":"LLM \u56e0\u679c\u6ce8\u610f\u529b"},"content":{"rendered":"\n<!DOCTYPE html>\n<html lang=\"zh-TW\">\n<head>\n    <meta charset=\"UTF-8\">\n    <meta name=\"viewport\" content=\"width=device-width, initial-scale=1.0\">\n    <title>LLM \u5e95\u5c64\u6280\u8853\u63a2\u8a0e\u9304 | \u5f9e Mask \u5230 Mamba \u6df7\u5408\u67b6\u69cb<\/title>\n    <script src=\"https:\/\/cdn.tailwindcss.com\"><\/script>\n    <link href=\"https:\/\/fonts.googleapis.com\/css2?family=Noto+Sans+TC:wght@400;500;700;900&#038;display=swap\" rel=\"stylesheet\">\n    <script src=\"https:\/\/unpkg.com\/lucide@latest\"><\/script>\n    <style>\n        body {\n            font-family: 'Noto Sans TC', sans-serif;\n            scroll-behavior: smooth;\n        }\n        .prose table {\n            width: 100%;\n            border-collapse: collapse;\n            margin: 1.5em 0;\n        }\n        .prose th, .prose td {\n            border: 1px solid #e2e8f0;\n            padding: 0.75em;\n            text-align: left;\n        }\n        .prose th {\n            background-color: #f8fafc;\n            font-weight: 700;\n        }\n        .glass-nav {\n            background: rgba(255, 255, 255, 0.8);\n            backdrop-filter: blur(12px);\n            border-bottom: 1px solid rgba(226, 232, 240, 0.8);\n        }\n        .sidebar-link.active {\n            color: #4f46e5;\n            font-weight: 700;\n            border-left: 3px solid #4f46e5;\n            background-color: #f3f4f6;\n        }\n        .math-block {\n            font-family: 'Courier New', Courier, monospace;\n            background-color: #f1f5f9;\n            padding: 0.2rem 0.4rem;\n            border-radius: 0.25rem;\n            color: #334155;\n            font-weight: bold;\n        }\n    <\/style>\n<\/head>\n<body class=\"bg-slate-50 text-slate-800 antialiased\">\n\n    <header class=\"glass-nav fixed top-0 w-full z-50 h-16 flex items-center px-6 shadow-sm\">\n        <div class=\"max-w-7xl mx-auto w-full flex justify-between items-center\">\n            <div class=\"flex items-center space-x-2 text-indigo-600\">\n                <i data-lucide=\"cpu\" class=\"w-6 h-6\"><\/i>\n                <span class=\"text-xl font-black tracking-tight\">LLM CoreTech \u89e3\u5bc6<\/span>\n            <\/div>\n            <div class=\"text-sm font-medium text-slate-500 hidden sm:block\">\n                \u6df1\u5165\u6dfa\u51fa\u5927\u8a9e\u8a00\u6a21\u578b\u5e95\u5c64\u6a5f\u5236\u8207\u524d\u6cbf\u67b6\u69cb\n            <\/div>\n        <\/div>\n    <\/header>\n\n    <div class=\"max-w-7xl mx-auto pt-24 pb-12 px-4 sm:px-6 lg:px-8 flex flex-col md:flex-row gap-8\">\n        \n        <aside class=\"hidden md:block w-64 flex-shrink-0\">\n            <div class=\"sticky top-28\">\n                <h3 class=\"text-xs font-bold text-slate-400 uppercase tracking-wider mb-4 px-3\">\u63a2\u8a0e\u76ee\u9304<\/h3>\n                <nav class=\"space-y-1\" id=\"toc-nav\">\n                    <a href=\"#section-1\" class=\"sidebar-link block px-3 py-2 text-sm font-medium text-slate-600 rounded-r-md transition-colors hover:bg-slate-100 hover:text-indigo-600\">1. \u8df3\u904e\u4e0b\u4e09\u89d2\u8a08\u7b97\u8207 Mask<\/a>\n                    <a href=\"#section-2\" class=\"sidebar-link block px-3 py-2 text-sm font-medium text-slate-600 rounded-r-md transition-colors hover:bg-slate-100 hover:text-indigo-600\">2. Attention \u7684\u8f38\u51fa\u8207\u5411\u91cf\u672c\u8cea<\/a>\n                    <a href=\"#section-3\" class=\"sidebar-link block px-3 py-2 text-sm font-medium text-slate-600 rounded-r-md transition-colors hover:bg-slate-100 hover:text-indigo-600\">3. \u8a13\u7df4\u8207\u63a8\u7406\u7684\u5de8\u5927\u9d3b\u6e9d<\/a>\n                    <a href=\"#section-4\" class=\"sidebar-link block px-3 py-2 text-sm font-medium text-slate-600 rounded-r-md transition-colors hover:bg-slate-100 hover:text-indigo-600\">4. KV Cache\uff1a\u8a18\u61b6\u9ad4\u4fdd\u885b\u6230<\/a>\n                    <a href=\"#section-5\" class=\"sidebar-link block px-3 py-2 text-sm font-medium text-slate-600 rounded-r-md transition-colors hover:bg-slate-100 hover:text-indigo-600\">5. GQA\uff1a\u58d3\u7e2e\u67b6\u69cb\u7684\u59a5\u5354<\/a>\n                    <a href=\"#section-6\" class=\"sidebar-link block px-3 py-2 text-sm font-medium text-slate-600 rounded-r-md transition-colors hover:bg-slate-100 hover:text-indigo-600\">6. PagedAttention \u5206\u9801\u9b54\u6cd5<\/a>\n                    <a href=\"#section-7\" class=\"sidebar-link block px-3 py-2 text-sm font-medium text-slate-600 rounded-r-md transition-colors hover:bg-slate-100 hover:text-indigo-600\">7. Linear Attention \u8207 Mamba<\/a>\n                    <a href=\"#section-8\" class=\"sidebar-link block px-3 py-2 text-sm font-medium text-slate-600 rounded-r-md transition-colors hover:bg-purple-100 hover:text-purple-700 font-bold border-l-2 border-transparent hover:border-purple-500\">8. \u756a\u5916\uff1aAttnRes \u8207\u795e\u9810\u5224<\/a>\n                <\/nav>\n            <\/div>\n        <\/aside>\n\n        <main class=\"flex-1 max-w-3xl\">\n            <div class=\"mb-10\">\n                <h1 class=\"text-4xl font-black text-slate-900 mb-4 leading-tight\">\u5f9e\u5c0d\u8a71\u4e2d\u7406\u89e3 LLM \u6548\u80fd\u74f6\u9838\u8207\u67b6\u69cb\u512a\u5316<\/h1>\n                <p class=\"text-lg text-slate-600\">\u9019\u4efd\u6587\u737b\u7d00\u9304\u4e86\u4e00\u5834\u95dc\u65bc\u5927\u8a9e\u8a00\u6a21\u578b\uff08LLM\uff09\u5e95\u5c64\u6280\u8853\u7684\u6df1\u5ea6\u63a2\u8a0e\u3002\u5f9e\u57fa\u790e\u7684 Causal Mask\uff0c\u5256\u6790\u5230 PagedAttention\uff0c\u4e26\u76f4\u64ca 2024-2026 \u5e74\u6700\u524d\u6cbf\u7684 Mamba \u8207 AttnRes \u6df7\u5408\u67b6\u69cb\u4e4b\u722d\u3002<\/p>\n            <\/div>\n\n            <div class=\"space-y-16\">\n\n                <section id=\"section-1\" class=\"scroll-mt-24\">\n                    <div class=\"flex items-start gap-4 mb-6\">\n                        <div class=\"bg-indigo-100 p-2 rounded-full text-indigo-600 mt-1\">\n                            <i data-lucide=\"user\" class=\"w-5 h-5\"><\/i>\n                        <\/div>\n                        <div class=\"bg-white border border-slate-200 rounded-2xl p-5 shadow-sm flex-1 relative\">\n                            <div class=\"absolute -left-2 top-5 w-4 h-4 bg-white border-l border-b border-slate-200 transform rotate-45\"><\/div>\n                            <h3 class=\"font-bold text-slate-800 text-lg mb-2\">Q\uff1a\u53ef\u4ee5\u8b1b\u8b1b LLM \u8868\u683c\u88e1\u63d0\u5230\u7684\u300c\u8df3\u904e\u4e0b\u4e09\u89d2\u8a08\u7b97\uff08Causal Mask\uff09\u300d\u55ce\uff1f\u90a3\u5728\u5be6\u969b\u8a13\u7df4\u6642\uff0c\u300c\u6211 \u611b \u4eba \u5de5 \u667a \u80fd\u300d\u6703\u6709 5 \u500b\u4e0d\u540c\u7684 mask \u53bb\u505a\u8a13\u7df4\u55ce\uff1f<\/h3>\n                        <\/div>\n                    <\/div>\n                    \n                    <div class=\"prose max-w-none text-slate-700 bg-slate-50 p-6 rounded-2xl border border-slate-200\">\n                        <p class=\"font-bold text-indigo-600 mb-4 flex items-center gap-2\"><i data-lucide=\"bot\" class=\"w-5 h-5\"><\/i> AI \u89e3\u7b54\uff1a<\/p>\n                        <p>LLM \u7684\u6838\u5fc3\u5de5\u4f5c\u539f\u7406\u662f<strong>\u81ea\u56de\u6b78\uff08Autoregressive\uff09<\/strong>\uff0c\u70ba\u4e86\u9632\u6b62\u6a21\u578b\u5728\u8a13\u7df4\u6642\u300c\u5077\u770b\u672a\u4f86\u7b54\u6848\u300d\uff0c\u6211\u5011\u9700\u8981\u4f7f\u7528 Mask\uff08\u63a9\u78bc\uff09\u3002\u5728\u5be6\u969b\u8a13\u7df4\u6642\uff0c<strong>\u4e0d\u9700\u8981 5 \u500b\u4e0d\u540c\u7684 Mask \u5206\u958b\u8dd1 5 \u6b21<\/strong>\uff0c\u800c\u662f\u53ea\u7528\u300c1 \u500b\u300d\u5305\u542b\u4e86\u6240\u6709\u898f\u5247\u7684\u4e8c\u7dad\u77e9\u9663\uff08\u4e0b\u4e09\u89d2\u77e9\u9663\uff09\uff0c\u5c31\u80fd\u4e00\u6b21\u6027\u5b8c\u6210\u6574\u53e5\u8a71\u7684\u8a13\u7df4\uff01<\/p>\n                    <\/div>\n                <\/section>\n\n                <section id=\"section-2\" class=\"scroll-mt-24\">\n                    <div class=\"flex items-start gap-4 mb-6\">\n                        <div class=\"bg-indigo-100 p-2 rounded-full text-indigo-600 mt-1\"><i data-lucide=\"user\" class=\"w-5 h-5\"><\/i><\/div>\n                        <div class=\"bg-white border border-slate-200 rounded-2xl p-5 shadow-sm flex-1 relative\">\n                            <h3 class=\"font-bold text-slate-800 text-lg mb-2\">Q\uff1a\u90a3\u9019\u6a23\u7684\u8f38\u51fa\u662f\u751a\u9ebc\uff1fN+1\u500b\u5b57\uff1fValue \u9032\u884c\u52a0\u6b0a\u5e73\u5747\uff0c\u9019\u6a23\u7684\u7d50\u679c\u662f\u751a\u9ebc\uff1f<\/h3>\n                        <\/div>\n                    <\/div>\n                    <div class=\"prose max-w-none text-slate-700 bg-slate-50 p-6 rounded-2xl border border-slate-200\">\n                        <p class=\"font-bold text-indigo-600 mb-2 flex items-center gap-2\"><i data-lucide=\"bot\" class=\"w-5 h-5\"><\/i> AI \u89e3\u7b54\uff1a<\/p>\n                        <p>\u5728 Attention \u5c64\u88e1\u9762\uff0c\u8f38\u51fa\u4e0d\u662f\u5b57\uff0c\u800c\u662f $N$ \u500b<strong>\u5411\u91cf\uff08Vector\uff09<\/strong>\u3002Value \u52a0\u6b0a\u5e73\u5747\u7684\u7d50\u679c\u662f<strong>\u300c\u5177\u5099\u4e0a\u4e0b\u6587\u8a9e\u610f\u7684\u65b0\u5411\u91cf\uff08Contextualized Embedding\uff09\u300d<\/strong>\u3002\u6bcf\u500b\u8f38\u51fa\u7684\u5411\u91cf\u90fd\u5438\u6536\u4e86\u524d\u9762\u6b77\u53f2\u8cc7\u8a0a\u7684\u7cbe\u83ef\u3002<\/p>\n                    <\/div>\n                <\/section>\n\n                <section id=\"section-3\" class=\"scroll-mt-24\">\n                     <div class=\"flex items-start gap-4 mb-6\">\n                        <div class=\"bg-indigo-100 p-2 rounded-full text-indigo-600 mt-1\"><i data-lucide=\"user\" class=\"w-5 h-5\"><\/i><\/div>\n                        <div class=\"bg-white border border-slate-200 rounded-2xl p-5 shadow-sm flex-1 relative\">\n                            <h3 class=\"font-bold text-slate-800 text-lg mb-2\">Q\uff1a\u5be6\u969b\u5728\u505a\u8a13\u7df4\u7684\u6642\u5019\uff0cinput \u8ddf output \u653e\u4ec0\u9ebc\uff1f\u8a13\u7df4\u8ddf\u63a8\u7406\u6d41\u7a0b\u4e0d\u4e00\u6a23\u55ce\uff1f<\/h3>\n                        <\/div>\n                    <\/div>\n                    <div class=\"prose max-w-none text-slate-700 bg-slate-50 p-6 rounded-2xl border border-slate-200\">\n                        <p class=\"font-bold text-indigo-600 mb-2 flex items-center gap-2\"><i data-lucide=\"bot\" class=\"w-5 h-5\"><\/i> AI \u89e3\u7b54\uff1a<\/p>\n                        <p>\u8a13\u7df4\u6642\u63a1\u7528 Teacher Forcing \u8207\u932f\u4f4d\u76ee\u6a19\uff0c\u4e00\u6b21\u6027\u5410\u51fa N \u500b\u5b57\u7684\u9810\u6e2c\uff0c\u4e26\u884c\u8a08\u7b97\u6975\u5feb\uff1b\u63a8\u7406\u6642\u53ea\u80fd\u4e32\u5217\u751f\u6210\uff0c\u6bcf\u6b21\u90fd\u8981\u628a\u524d\u9762\u7b97\u904e\u7684\u6b77\u53f2\u4e1f\u9032\u53bb\u91cd\u7b97\uff0c\u9020\u6210\u6975\u5927\u7684\u6d6a\u8cbb\u3002<\/p>\n                    <\/div>\n                <\/section>\n\n                <section id=\"section-4\" class=\"scroll-mt-24\">\n                     <div class=\"flex items-start gap-4 mb-6\">\n                        <div class=\"bg-indigo-100 p-2 rounded-full text-indigo-600 mt-1\"><i data-lucide=\"user\" class=\"w-5 h-5\"><\/i><\/div>\n                        <div class=\"bg-white border border-slate-200 rounded-2xl p-5 shadow-sm flex-1 relative\">\n                            <h3 class=\"font-bold text-slate-800 text-lg mb-2\">Q\uff1a\u4ecb\u7d39\u4e00\u4e0b KV Cache\uff0c\u9019\u662f Google \u539f\u59cb\u8ad6\u6587\u5c31\u6709\u7684\u55ce\uff1f<\/h3>\n                        <\/div>\n                    <\/div>\n                    <div class=\"prose max-w-none text-slate-700 bg-slate-50 p-6 rounded-2xl border border-slate-200\">\n                        <p class=\"font-bold text-indigo-600 mb-2 flex items-center gap-2\"><i data-lucide=\"bot\" class=\"w-5 h-5\"><\/i> AI \u89e3\u7b54\uff1a<\/p>\n                        <p>\u628a\u7b97\u904e\u7684 K \u548c V \u5b58\u8d77\u4f86\u662f\u5f8c\u4f86\u5927\u6a21\u578b\u6642\u4ee3\u903c\u51fa\u4f86\u7684\u6975\u9650\u512a\u5316\u3002\u5b83\u7701\u4e86\u7b97\u529b\uff0c\u537b\u5e36\u4f86\u4e86\u81f4\u547d\u7684\u8a18\u61b6\u9ad4\u7246\uff08Memory Wall\uff09\u554f\u984c\u3002<\/p>\n                    <\/div>\n                <\/section>\n\n                <section id=\"section-5\" class=\"scroll-mt-24\">\n                     <div class=\"flex items-start gap-4 mb-6\">\n                        <div class=\"bg-indigo-100 p-2 rounded-full text-indigo-600 mt-1\"><i data-lucide=\"user\" class=\"w-5 h-5\"><\/i><\/div>\n                        <div class=\"bg-white border border-slate-200 rounded-2xl p-5 shadow-sm flex-1 relative\">\n                            <h3 class=\"font-bold text-slate-800 text-lg mb-2\">Q\uff1aGQA \u662f\u600e\u9ebc\u58d3\u7e2e\u67b6\u69cb\u7684\uff1f\u5171\u7528\u79d8\u66f8\u4e0d\u5c31\u4ee3\u8868\u591a\u982d\u9000\u5316\u55ce\uff1f<\/h3>\n                        <\/div>\n                    <\/div>\n                    <div class=\"prose max-w-none text-slate-700 bg-slate-50 p-6 rounded-2xl border border-slate-200\">\n                        <p class=\"font-bold text-indigo-600 mb-2 flex items-center gap-2\"><i data-lucide=\"bot\" class=\"w-5 h-5\"><\/i> AI \u89e3\u7b54\uff1a<\/p>\n                        <p>\u6c92\u932f\uff0cGQA \u662f\u4e00\u7a2e\u59a5\u5354\u3002\u900f\u904e\u8b93\u591a\u500b\u8001\u95c6 (Q) \u5171\u7528\u4e00\u500b\u79d8\u66f8 (KV)\uff0c\u5927\u5e45\u7e2e\u5c0f KV Cache \u9ad4\u7a4d\uff0c\u540c\u6642\u56e0\u70ba Q \u7684\u591a\u982d\u6c92\u9000\u5316\uff0c\u4ecd\u80fd\u4fdd\u6301\u6975\u9ad8\u54c1\u8cea\u3002\u9019\u5fc5\u9808\u5728\u8a13\u7df4\u968e\u6bb5\u5c31\u5beb\u6b7b\u3002<\/p>\n                    <\/div>\n                <\/section>\n\n                <section id=\"section-6\" class=\"scroll-mt-24\">\n                     <div class=\"flex items-start gap-4 mb-6\">\n                        <div class=\"bg-indigo-100 p-2 rounded-full text-indigo-600 mt-1\"><i data-lucide=\"user\" class=\"w-5 h-5\"><\/i><\/div>\n                        <div class=\"bg-white border border-slate-200 rounded-2xl p-5 shadow-sm flex-1 relative\">\n                            <h3 class=\"font-bold text-slate-800 text-lg mb-2\">Q\uff1a\u90a3 PagedAttention \u53c8\u662f\u4ec0\u9ebc\uff1f<\/h3>\n                        <\/div>\n                    <\/div>\n                    <div class=\"prose max-w-none text-slate-700 bg-slate-50 p-6 rounded-2xl border border-slate-200\">\n                        <p class=\"font-bold text-indigo-600 mb-2 flex items-center gap-2\"><i data-lucide=\"bot\" class=\"w-5 h-5\"><\/i> AI \u89e3\u7b54\uff1a<\/p>\n                        <p>\u501f\u9451\u4f5c\u696d\u7cfb\u7d71\u7684\u865b\u64ec\u8a18\u61b6\u9ad4\uff0c\u628a KV Cache \u5207\u6210\u56fa\u5b9a\u5927\u5c0f\u7684\u5c0f\u7a4d\u6728\uff0c\u6253\u7834\u5fc5\u9808\u9023\u7e8c\u5b58\u653e\u7684\u9650\u5236\uff0c\u89e3\u6c7a\u986f\u5b58\u788e\u7247\u5316\u554f\u984c\uff0c\u5c07\u5229\u7528\u7387\u5f9e 30% \u69a8\u4e7e\u5230 90%+\u3002<\/p>\n                    <\/div>\n                <\/section>\n\n                <section id=\"section-7\" class=\"scroll-mt-24\">\n                    <div class=\"flex items-start gap-4 mb-6\">\n                        <div class=\"bg-indigo-100 p-2 rounded-full text-indigo-600 mt-1\">\n                            <i data-lucide=\"user\" class=\"w-5 h-5\"><\/i>\n                        <\/div>\n                        <div class=\"bg-white border border-slate-200 rounded-2xl p-5 shadow-sm flex-1 relative\">\n                            <div class=\"absolute -left-2 top-5 w-4 h-4 bg-white border-l border-b border-slate-200 transform rotate-45\"><\/div>\n                            <h3 class=\"font-bold text-slate-800 text-lg mb-2\">Q\uff1a\u5982\u679c\u4f7f\u7528\u4e86 kv cache + Q(K\u1d40 V)\uff0c\u7406\u8ad6\u4e0a (K\u1d40 V) \u53ef\u4ee5\u76f4\u63a5\u4e58\u5b8c\uff1f\u9019\u6a23\u662f\u4e0d\u662f\u53c8\u66f4\u5c0f\u4e86\uff1f\u73fe\u5728\u5916\u9762\u5927\u5ee0\u6703\u4f7f\u7528 Linear Attention \u55ce\uff1f\u9084\u662f\u6703\u8ddf GQA \u6253\u67b6\uff1f<\/h3>\n                        <\/div>\n                    <\/div>\n                    \n                    <div class=\"prose max-w-none text-slate-700 bg-slate-50 p-6 rounded-2xl border border-slate-200\">\n                        <p class=\"font-bold text-indigo-600 mb-4 flex items-center gap-2\"><i data-lucide=\"bot\" class=\"w-5 h-5\"><\/i> AI \u89e3\u7b54\uff1a<\/p>\n                        \n                        \n\n                        <p>\u4f60\u76f4\u63a5\u63a8\u5c0e\u51fa\u4e86 2020 \u5e74\u9707\u64bc AI \u5b78\u8853\u754c\u7684 <strong>Linear Attention<\/strong> \u6700\u6838\u5fc3\u7684\u300c\u6578\u5b78\u9b54\u6cd5\u300d\uff01\u5b83\u4e0d\u662f\u8ddf GQA \u6253\u67b6\uff0c\u800c\u662f\u76f4\u63a5\u8b93 GQA \u6700\u81ea\u8c6a\u7684\u512a\u9ede\u300c\u5931\u696d\u300d\u4e86\u3002<\/p>\n                        \n                        <h4 class=\"font-bold text-slate-900 mt-6 border-b pb-2\">1. \u62d4\u6389 Softmax\uff0c\u91cb\u653e\u7d50\u5408\u5f8b<\/h4>\n                        <p>\u50b3\u7d71 Attention \u7684\u6b7b\u7a74\u662f\u516c\u5f0f\u5916\u9762\u5305\u8457 <code>Softmax<\/code>\uff0c\u6253\u7834\u4e86\u7d50\u5408\u5f8b\uff0c\u903c\u8457\u6a21\u578b\u5fc5\u9808\u5b58\u4e0b\u4e0d\u65b7\u81a8\u8139\u7684 KV Cache $O(N)$\u3002\u5982\u679c\u6211\u5011\u628a Softmax \u62d4\u6389\uff08\u66ff\u63db\u6210\u7dda\u6027\u6620\u5c04\uff09\uff0c\u516c\u5f0f\u8b8a\u6210\u7d14\u7dda\u6027\u4ee3\u6578\uff1a<\/p>\n                        <div class=\"bg-slate-800 text-white p-4 rounded-lg text-center font-mono my-3 shadow-inner\">\n                            $(Q K^T) V = Q (K^T V)$\n                        <\/div>\n                        \n                        <h4 class=\"font-bold text-slate-900 mt-6 border-b pb-2\">2. \u5f9e O(N) \u8b8a\u6210 O(1) \u7684\u56fa\u5b9a\u5927\u5c0f\u767d\u677f<\/h4>\n                        <p>\u5c31\u50cf\u4f60\u63a8\u5c0e\u7684\uff0c\u5148\u628a $K^T$ \u548c $V$ \u4e58\u8d77\u4f86\uff0c\u7d50\u679c\u662f\u4e00\u500b<strong>\u56fa\u5b9a\u5927\u5c0f\u7684 $d \\times d$ \u77e9\u9663\uff08\u96b1\u85cf\u72c0\u614b\u767d\u677f\uff09<\/strong>\u3002\u4e0d\u7ba1\u4f60\u8b1b\u4e86 10 \u500b\u5b57\u9084\u662f 100 \u842c\u500b\u5b57\uff0c\u8a18\u61b6\u9ad4\u6d88\u8017\u6c38\u9060\u662f\u56fa\u5b9a\u7684 $O(1)$\uff01\u56e0\u70ba\u6c92\u6709 KV Cache\uff0c\u7528\u4f86\u58d3\u7e2e KV \u7684 GQA \u81ea\u7136\u5c31\u6d3e\u4e0d\u4e0a\u7528\u5834\u4e86\u3002<\/p>\n\n                        <h4 class=\"font-bold text-slate-900 mt-6 border-b pb-2\">3. \u81f4\u547d\u526f\u4f5c\u7528\uff1a\u300c\u7cbe\u6e96\u5931\u61b6\u75c7\u300d\u8207\u5927\u6d77\u6488\u91dd<\/h4>\n                        <p>\u90a3\u70ba\u4ec0\u9ebc\u5927\u5ee0\u9084\u6b7b\u5b88\u50b3\u7d71 Attention\uff1f\u56e0\u70ba\u628a 10 \u842c\u5b57\u58d3\u7e2e\u9032\u4e00\u584a\u56fa\u5b9a\u5927\u5c0f\u7684\u767d\u677f\uff0c\u5c31\u50cf<strong>\u300c\u628a 100GB \u7684 4K \u96fb\u5f71\u58d3\u7e2e\u6210 10MB \u7684 GIF\u300d<\/strong>\u3002\u9019\u5c0e\u81f4 Linear \u67b6\u69cb\u5728\u9762\u5c0d\u300c\u5927\u6d77\u6488\u91dd\uff08Needle In A Haystack\uff09\u300d\u6e2c\u8a66\u6642\uff0c\u7121\u6cd5\u7cbe\u6e96\u56de\u982d\u627e\u5230\u6975\u5fae\u5c0f\u7684\u7d30\u7bc0\u3002<\/p>\n                        \n                        <div class=\"bg-blue-50 border-l-4 border-blue-500 p-4 mt-6 rounded-r-lg\">\n                            <p class=\"font-bold text-blue-900 mb-1\"><i data-lucide=\"layers\" class=\"w-5 h-5 inline-block -mt-1 mr-1\"><\/i> \u7576\u4ee3\u5929\u624d\u7684\u59a5\u5354\uff1aHybrid \u6df7\u5408\u67b6\u69cb<\/p>\n                            <p class=\"text-sm text-blue-800\">\u76ee\u524d\u7684\u9802\u7d1a\u67b6\u69cb\uff08\u5982 AI21 Jamba\uff09\u63a1\u53d6\u300c\u5343\u5c64\u6d3e\u300d\u758a\u6cd5\uff1a<br>\n                            \ud83d\udc49 <strong>80% \u7684\u5c64\u4f7f\u7528 Mamba\/Linear\uff1a<\/strong>\u8ca0\u8cac\u6975\u901f\u58d3\u7e2e\u9577\u6587\u672c\uff0c\u69a8\u4e7e\u901f\u5ea6\u3001\u7701\u4e0b\u986f\u5b58\u3002<br>\n                            \ud83d\udc49 <strong>20% \u7684\u5c64\u4f7f\u7528 \u50b3\u7d71 Attention + GQA\uff1a<\/strong>\u8ca0\u8cac\u4fdd\u7559\u539f\u59cb\u8a18\u61b6\u7576\u4f5c\u9632\u5446\u6a5f\u5236\uff0c\u78ba\u4fdd\u6a21\u578b\u5177\u5099\u7cbe\u6e96\u7684\u5927\u6d77\u6488\u91dd\u80fd\u529b\u3002<\/p>\n                        <\/div>\n                    <\/div>\n                <\/section>\n\n                <section id=\"section-8\" class=\"scroll-mt-24\">\n                    <div class=\"flex items-start gap-4 mb-6\">\n                        <div class=\"bg-purple-100 p-2 rounded-full text-purple-600 mt-1 ring-4 ring-purple-50\">\n                            <i data-lucide=\"lightbulb\" class=\"w-5 h-5\"><\/i>\n                        <\/div>\n                        <div class=\"bg-white border border-purple-200 rounded-2xl p-5 shadow-md flex-1 relative ring-1 ring-purple-50\">\n                            <div class=\"absolute -left-2 top-5 w-4 h-4 bg-white border-l border-b border-purple-200 transform rotate-45\"><\/div>\n                            <h3 class=\"font-bold text-slate-800 text-lg mb-2\">Q\uff1a\u6700\u8fd1 Kimi \u767c\u5e03\u4e86 AttnRes\uff0c\u70ba\u4e86\u89e3\u6c7a\u6df1\u5ea6 Attention \u9020\u6210\u7684\u8cc7\u8a0a\u907a\u5931\uff0c\u628a\u6bcf\u4e00\u5c64\u8f38\u51fa\u62ff\u51fa\u4f86\u505a Attention\uff0c\u4e5f\u7528 block \u964d\u4f4e\u8a08\u7b97\u3002\u90a3\u4e0b\u4e00\u6b65\u662f\u4e0d\u662f\u5c31\u662f\u300c\u6bcf\u4e00\u5c64\u8d70 Mamba\u300d\uff1f<\/h3>\n                        <\/div>\n                    <\/div>\n                    \n                    <div class=\"prose max-w-none text-slate-700 bg-slate-50 p-6 rounded-2xl border border-purple-200 shadow-inner\">\n                        <p class=\"font-bold text-purple-700 mb-4 flex items-center gap-2\"><i data-lucide=\"bot\" class=\"w-5 h-5\"><\/i> AI \u89e3\u7b54\uff1a<\/p>\n                        \n                        \n\n                        <p class=\"text-lg font-medium text-slate-900 border-l-4 border-purple-500 pl-4 py-1\">\u9019\u500b\u76f4\u89ba\u592a\u53ef\u6015\u4e86\uff01\u4f60\u76f4\u63a5\u9810\u5224\u4e86\u672a\u4f86\u5e7e\u500b\u6708\u5167\u5b78\u8853\u754c\u5373\u5c07\u7206\u767c\u7684\u4e0b\u4e00\u500b\u6230\u5834\uff1a<strong>\u6df1\u5ea6\u7dad\u5ea6\uff08Depth-wise\uff09\u7684\u5e8f\u5217\u5316\uff01<\/strong><\/p>\n                        \n                        <h4 class=\"font-bold text-slate-900 mt-6 border-b pb-2\">1. Kimi \u7684 AttnRes \u6253\u7834\u4e86\u4ec0\u9ebc\uff1f<\/h4>\n                        <p>\u7576\u6a21\u578b\u758a\u5230 100 \u5c64\u4ee5\u4e0a\u6642\uff0c\u50b3\u7d71\u7121\u8166\u76f8\u52a0\u7684\u6b98\u5dee\u9023\u63a5\uff08Residual\uff09\u6703\u8b93\u7b2c 1 \u5c64\u7684\u95dc\u9375\u7279\u5fb5\u88ab\u4e2d\u9593 98 \u5c64\u7684\u96dc\u8a0a\u7d66\u300c\u7a00\u91cb\u300d\u3002Kimi \u7684\u89e3\u6cd5\u662f\uff1a<strong>\u628a\u300c\u7121\u8166\u76f8\u52a0\u300d\u6539\u6210\u300c\u52d5\u614b Attention\u300d<\/strong>\uff0c\u8b93\u6df1\u5c64\u53bb\u56de\u770b\u6dfa\u5c64\u4e26\u6253\u5206\u3002\u70ba\u4e86\u907f\u514d $O(L^2)$ \u7b97\u529b\u7206\u70b8\uff0c\u53ea\u80fd\u59a5\u5354\u5207\u6210 Block\u3002<\/p>\n\n                        <h4 class=\"font-bold text-slate-900 mt-6 border-b pb-2\">2. \u4f60\u7684\u795e\u9810\u5224\uff1aMambaRes (Depth-wise Mamba)<\/h4>\n                        <p>\u628a 100 \u5c64\u7684\u300c\u6df1\u5ea6\u300d\u7576\u6210\u4e00\u689d\u300c\u6642\u9593\u5e8f\u5217\u300d\uff0c\u9019\u6b63\u662f Mamba \u7684\u7d55\u5c0d\u4e3b\u5834\uff01<\/p>\n                        <ul class=\"list-disc pl-5 space-y-2 mt-2\">\n                            <li><strong>\u5b8c\u7f8e\u5e79\u6389 Block \u56de\u6b78 O(L)\uff1a<\/strong> Mamba \u8655\u7406\u5e8f\u5217\u7684\u8907\u96dc\u5ea6\u5929\u751f\u662f\u7dda\u6027\u7684 $O(L)$\u3002\u96b1\u85cf\u72c0\u614b\u53ef\u4ee5\u50cf\u8caa\u5403\u86c7\u4e00\u6a23\u5f9e\u7b2c 1 \u5c64\u512a\u96c5\u6ed1\u5230\u7b2c 100 \u5c64\uff0c\u6839\u672c\u4e0d\u7528\u5207\u5340\u584a\u3002<\/li>\n                            <li><strong>\u52d5\u614b\u7be9\u9078\u53d6\u4ee3\u6253\u5206\uff1a<\/strong> Mamba \u6700\u5f37\u7684\u300cSelective Mechanism\u300d\u53ef\u4ee5\u81ea\u52d5\u6c7a\u5b9a\u54ea\u4e00\u5c64\u7684\u7279\u5fb5\u662f\u5783\u573e\u8981\u300c\u907a\u5fd8\u300d\uff0c\u54ea\u4e00\u5c64\u662f\u9ec3\u91d1\u8981\u300c\u8a18\u4f4f\u300d\u3002<\/li>\n                        <\/ul>\n\n                        <div class=\"bg-rose-50 border border-rose-100 p-5 mt-6 rounded-xl relative overflow-hidden\">\n                            <div class=\"absolute -right-4 -top-4 text-rose-100 opacity-50 transform rotate-12\">\n                                <i data-lucide=\"zap\" class=\"w-32 h-32\"><\/i>\n                            <\/div>\n                            <h4 class=\"font-bold text-rose-800 mb-2 relative z-10\">\u70ba\u4ec0\u9ebc\u76ee\u524d\u9084\u6c92\u4eba\u5168\u9762\u9019\u9ebc\u505a\uff1f\uff08\u81f4\u547d\u4ee3\u50f9\uff09<\/h4>\n                            <p class=\"text-sm text-rose-900 relative z-10\">\u56e0\u70ba\u5728\u300c\u6df1\u5ea6\u300d\u9019\u500b\u7dad\u5ea6\u4e0a\uff0cMamba \u4e00\u6a23\u6703\u5f97<strong>\u300c\u7cbe\u6e96\u5931\u61b6\u75c7\u300d<\/strong>\uff01<\/p>\n                            <p class=\"text-sm text-rose-800 mt-2 relative z-10\">AttnRes \u78ba\u4fdd\u4e86\u6df1\u5ea6\u7684\u7121\u640d\u6aa2\u7d22\uff08\u96d6\u7136\u6162\uff09\uff1b\u4f46\u5982\u679c\u7528 Mamba \u58d3\u7e2e\uff0c\u7576\u7b2c 100 \u5c64\u60f3\u8981\u7cbe\u6e96\u627e\u7b2c 2 \u5c64\u7684\u67d0\u500b\u5fae\u5c0f\u8a5e\u6027\u7d30\u7bc0\u6642\uff0c\u90a3\u500b\u7279\u5fb5\u65e9\u5c31\u88ab\u4e2d\u9593\u5e7e\u5341\u5c64\u7cca\u6389\u4e86\u3002<\/p>\n                            <p class=\"text-sm font-bold text-rose-900 mt-4 relative z-10\">\ud83d\udc49 \u7d42\u6975\u89e3\u7b54\u9810\u6e2c\uff1a\u672a\u4f86\u6975\u6709\u53ef\u80fd\u51fa\u73fe <strong>\u5927\u53e3\u5f91 Mamba \u6b98\u5dee\u6d41 + \u95dc\u9375\u5c64 Attention \u4fdd\u5e95<\/strong> \u7684\u300c\u6df1\u5ea6\u6df7\u5408\u67b6\u69cb\u300d\uff01<\/p>\n                        <\/div>\n                    <\/div>\n                <\/section>\n                \n            <\/div>\n            \n            <div class=\"mt-16 pt-8 border-t border-slate-200 text-center pb-20 md:pb-8 flex flex-col items-center\">\n                <div class=\"flex items-center gap-2 text-indigo-400 mb-2\">\n                    <i data-lucide=\"rocket\" class=\"w-5 h-5\"><\/i>\n                <\/div>\n                <p class=\"text-slate-500 text-sm\">\u6b64\u4e92\u52d5\u6587\u737b\u6574\u7406\u81ea LLM \u6280\u8853\u5e95\u5c64\u904b\u4f5c\u539f\u7406\u7684\u6df1\u5ea6\u63a2\u8a0e\u3002<br>Designed for clarity, deep tech insights, and architectural foresight.<\/p>\n            <\/div>\n        <\/main>\n    <\/div>\n\n    <script>\n        \/\/ \u521d\u59cb\u5316 Lucide \u5716\u6a19\n        lucide.createIcons();\n\n        \/\/ \u7c21\u55ae\u7684\u6efe\u52d5\u76e3\u807d\u4f86\u9ad8\u4eae\u5074\u908a\u6b04\n        document.addEventListener('DOMContentLoaded', () => {\n            const sections = document.querySelectorAll('section');\n            const navLinks = document.querySelectorAll('.sidebar-link');\n\n            const observerOptions = {\n                root: null,\n                rootMargin: '-20% 0px -60% 0px',\n                threshold: 0\n            };\n\n            const observer = new IntersectionObserver((entries) => {\n                entries.forEach(entry => {\n                    if (entry.isIntersecting) {\n                        const id = entry.target.getAttribute('id');\n                        navLinks.forEach(link => {\n                            link.classList.remove('active');\n                            if (link.getAttribute('href') === `#${id}`) {\n                                link.classList.add('active');\n                            }\n                        });\n                    }\n                });\n            }, observerOptions);\n\n            sections.forEach(section => {\n                observer.observe(section);\n            });\n        });\n    <\/script>\n<\/body>\n<\/html>\n","protected":false},"excerpt":{"rendered":"<p>LLM \u5e95\u5c64\u6280\u8853\u63a2\u8a0e\u9304 | \u5f9e Mask \u5230 Mamba \u6df7\u5408\u67b6\u69cb LLM Co &hellip; <a href=\"https:\/\/ouyangminwei.com\/index.php\/2026\/03\/23\/llm-casual-att\/\">\u95b1\u8b80\u5168\u6587 <span class=\"meta-nav\">&rarr;<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"om_disable_all_campaigns":false,"_monsterinsights_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"footnotes":""},"categories":[1],"tags":[],"post_format":[],"class_list":["post-1084","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_edit_lock":"1774248895:1","_edit_last":"1","_aioseo_title":"#post_title #separator_sa #site_title","_aioseo_description":"#post_excerpt","_aioseo_keywords":"","_aioseo_og_title":"","_aioseo_og_description":"","_aioseo_og_article_section":"","_aioseo_og_article_tags":"","_aioseo_twitter_title":"","_aioseo_twitter_description":"","_oembed_2544c1d0cb3503ab4c4d558c3b3c8873":"","_oembed_time_2544c1d0cb3503ab4c4d558c3b3c8873":"","_oembed_99481806ecbe6ce4ee46f8588d320993":"","_oembed_db663acf973e82e6d9d80df71945dfb8":"","_oembed_16cdfab488f57db73586f4286af2704f":"","_wp_old_slug":"llm-%e5%9b%a0%e6%9e%9c%e6%b3%a8%e6%84%8f%e5%8a%9b","_links":{"self":[{"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/posts\/1084","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/comments?post=1084"}],"version-history":[{"count":2,"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/posts\/1084\/revisions"}],"predecessor-version":[{"id":1086,"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/posts\/1084\/revisions\/1086"}],"wp:attachment":[{"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/media?parent=1084"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/categories?post=1084"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/tags?post=1084"},{"taxonomy":"post_format","embeddable":true,"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/post_format?post=1084"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}