Loading...
正在加载...
请稍候

RWKV模型深度研究报告 | 2026年2月

✨步子哥 (steper) 2026年02月13日 02:57
<!DOCTYPE html><html lang="zh"><head> <meta charset="UTF-8"/> <meta name="viewport" content="width=device-width, initial-scale=1.0"/> <title>RWKV模型深度研究报告 | 2026年2月</title> <script src="https://cdn.tailwindcss.com"></script> <script> tailwind.config = { theme: { extend: { colors: { primary: '#1e293b', secondary: '#64748b', accent: '#0f172a', neutral: '#f8fafc', 'base-100': '#ffffff', 'base-200': '#f1f5f9', 'base-300': '#e2e8f0' }, fontFamily: { 'sans': ['Inter', 'system-ui', 'sans-serif'], 'display': ['Playfair Display', 'serif'], 'mono': ['JetBrains Mono', 'monospace'] } } } } </script> <link href="https://fonts.googleapis.com/css2?family=Playfair+Display:ital,wght@0,400;0,600;0,700;1,400;1,600&amp;family=Inter:wght@300;400;500;600;700&amp;family=JetBrains+Mono:wght@400;500&amp;display=swap" rel="stylesheet"/> <link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.4.0/css/all.min.css"/> <script src="https://cdn.jsdelivr.net/npm/mermaid@10.6.1/dist/mermaid.min.js"></script> <style> .gradient-overlay { background: linear-gradient(135deg, rgba(30, 41, 59, 0.9) 0%, rgba(15, 23, 42, 0.7) 100%); } .text-shadow { text-shadow: 2px 2px 4px rgba(0, 0, 0, 0.3); } .hero-pattern { background-image: radial-gradient(circle at 25% 25%, rgba(100, 116, 139, 0.1) 0%, transparent 50%), radial-gradient(circle at 75% 75%, rgba(100, 116, 139, 0.05) 0%, transparent 50%); } .mermaid-container { display: flex; justify-content: center; min-height: 100px; max-height: 800px; background: #ffffff; border: 2px solid #e5e7eb; border-radius: 12px; padding: 30px; margin: 30px 0; box-shadow: 0 8px 25px rgba(0, 0, 0, 0.08); position: relative; overflow: hidden; } .mermaid-container .mermaid { width: 100%; max-width: 100%; height: 100%; cursor: grab; transition: transform 0.3s ease; transform-origin: center center; display: flex; justify-content: center; align-items: center; touch-action: none; -webkit-user-select: none; -moz-user-select: none; -ms-user-select: none; user-select: none; } .mermaid-container .mermaid svg { max-width: 100%; height: 100%; display: block; margin: 0 auto; } .mermaid-container .mermaid:active { cursor: grabbing; } .mermaid-container.zoomed .mermaid { height: 100%; width: 100%; cursor: grab; } .mermaid-controls { position: absolute; top: 15px; right: 15px; display: flex; gap: 10px; z-index: 20; background: rgba(255, 255, 255, 0.95); padding: 8px; border-radius: 8px; box-shadow: 0 2px 8px rgba(0, 0, 0, 0.1); } .mermaid-control-btn { background: #ffffff; border: 1px solid #d1d5db; border-radius: 6px; padding: 10px; cursor: pointer; transition: all 0.2s ease; color: #374151; font-size: 14px; min-width: 36px; height: 36px; text-align: center; display: flex; align-items: center; justify-content: center; } .mermaid-control-btn:hover { background: #f8fafc; border-color: #3b82f6; color: #3b82f6; transform: translateY(-1px); } .mermaid-control-btn:active { transform: scale(0.95); } /* Mermaid node styling for better contrast */ .mermaid .node rect, .mermaid .node circle, .mermaid .node ellipse, .mermaid .node polygon { stroke-width: 2px !important; stroke: #475569 !important; } .mermaid .node .label { color: #334155 !important; font-weight: 500 !important; font-size: 13px !important; text-shadow: none !important; } .mermaid .edgePath .path { stroke: #64748b !important; stroke-width: 2px !important; } .mermaid .edgeLabel { background-color: rgba(255, 255, 255, 0.9) !important; color: #475569 !important; font-size: 11px !important; font-weight: 500 !important; padding: 2px 6px !important; border-radius: 4px !important; } /* Ensure text readability in all mermaid elements */ .mermaid text { font-family: 'Inter', system-ui, sans-serif !important; text-shadow: none !important; } /* Enhanced contrast for different node colors */ .mermaid .node[style*="fill:#e3f2fd"] .label, .mermaid .node[style*="fill:#e8f5e8"] .label, .mermaid .node[style*="fill:#f3e5f5"] .label, .mermaid .node[style*="fill:#fff3e0"] .label, .mermaid .node[style*="fill:#e0f2f1"] .label, .mermaid .node[style*="fill:#f1f8e9"] .label, .mermaid .node[style*="fill:#fbe9e7"] .label { color: #1e293b !important; font-weight: 600 !important; } /* Prevent horizontal overflow on small screens */ <span class="mention-invalid">@media</span> (max-width: 767px) { body { overflow-x: hidden; } .mermaid-container { padding: 15px; } } </style> <base target="_blank"> </head> <body class="bg-base-100 text-primary font-sans leading-relaxed"> <!-- Main Content --> <div class="max-w-6xl mx-auto px-4 md:px-8 py-16"> <!-- Backdrop for mobile TOC --> <div id="toc-backdrop" class="fixed inset-0 bg-black/50 z-40 hidden lg:hidden"></div> <!-- Mobile TOC Toggle --> <button id="mobile-toc-toggle" class="fixed top-4 left-4 z-50 lg:hidden bg-primary text-white p-3 rounded-full shadow-lg"> <i class="fas fa-bars"></i> </button> <!-- Executive Summary --> <section id="executive-summary" class="mb-20"> <div class="bg-gradient-to-r from-blue-50 to-indigo-50 rounded-2xl p-4 md:p-8 border border-blue-100"> <h2 class="font-display text-3xl font-bold text-primary mb-6">执行摘要</h2> <div class="prose prose-lg max-w-none"> <p class="text-lg text-secondary leading-relaxed mb-6"> RWKV是一种创新的RNN-Transformer混合架构,通过线性注意力机制实现了<strong>O(n)的时间复杂度</strong>和<strong>O(1)的内存占用</strong>。截至2026年1月,RWKV-7-G1系列在保持同等模型质量水平的同时,相比传统Transformer架构展现出数量级的效率优势。 </p> <div class="grid grid-cols-1 md:grid-cols-3 gap-6 mt-8"> <div class="bg-white rounded-lg p-6 shadow-sm border border-blue-100"> <div class="text-xl font-semibold text-primary mb-2">架构突破</div> <p class="text-sm text-secondary">完全摒弃O(n²)自注意力机制,采用线性递归状态更新</p> </div> <div class="bg-white rounded-lg p-6 shadow-sm border border-blue-100"> <div class="text-xl font-semibold text-primary mb-2">性能优势</div> <p class="text-sm text-secondary">RTX 4090上115+ tokens/s,显存占用低至2.4GB</p> </div> <div class="bg-white rounded-lg p-6 shadow-sm border border-blue-100"> <div class="text-xl font-semibold text-primary mb-2">应用前景</div> <p class="text-sm text-secondary">特别适合长文档处理、实时交互和边缘部署场景</p> </div> </div> </div> </div> </section> <!-- Model Architecture --> <section id="architecture" class="mb-20"> <h2 class="font-display text-4xl font-bold text-primary mb-12">模型架构与技术原理</h2> <div class="mb-16"> <h3 class="font-display text-2xl font-semibold text-primary mb-8">核心架构设计</h3> <div class="grid grid-cols-1 lg:grid-cols-2 gap-12 mb-12"> <div> <h4 class="text-xl font-semibold text-primary mb-4">RNN-Transformer混合架构</h4> <p class="text-secondary mb-6 leading-relaxed"> RWKV(Receptance Weighted Key Value)通过线性注意力机制替代传统多头注意力,成功融合了RNN的高效推理特性与Transformer的并行训练优势。该架构由Bo Peng及其领导的RWKV开源社区开发,自2023年加入Linux基金会以来,已发展成为备受关注的非盈利开源项目。 </p> <div class="bg-blue-50 border-l-4 border-blue-400 p-4 rounded-r-lg"> <p class="text-sm text-blue-800"> <strong>关键创新:</strong>完全摒弃O(n²)自注意力机制,采用递归状态更新机制,将复杂度降至O(n)线性级别 </p> </div> </div> <div> <img src="https://kimi-web-img.moonshot.cn/img/pic4.zhimg.com/4127ed9565616dc5def74c2948a81c9a59ca1732.webp" alt="RNN和Transformer混合神经网络结构示意图" class="w-full rounded-lg shadow-md" size="medium" aspect="wide" style="linedrawing" query="RNN Transformer 混合架构" referrerpolicy="no-referrer" data-modified="1" data-score="0.00"/> </div> </div> <!-- Mermaid Diagram --> <div class="mermaid-container"> <div class="mermaid-controls"> <button class="mermaid-control-btn zoom-in" title="放大"> <i class="fas fa-search-plus"></i> </button> <button class="mermaid-control-btn zoom-out" title="缩小"> <i class="fas fa-search-minus"></i> </button> <button class="mermaid-control-btn reset-zoom" title="重置"> <i class="fas fa-expand-arrows-alt"></i> </button> <button class="mermaid-control-btn fullscreen" title="全屏查看"> <i class="fas fa-expand"></i> </button> </div> <div class="mermaid"> graph TD A[&#34;Input Sequence&#34;] --&gt; B[&#34;Token Embedding&#34;] B --&gt; C[&#34;Time-Mixing Block&#34;] B --&gt; D[&#34;Channel-Mixing Block&#34;] C --&gt; E[&#34;Receptance Gate R&#34;] C --&gt; F[&#34;Weight Decay W&#34;] C --&gt; G[&#34;Key K &amp; Value V&#34;] E --&gt; H[&#34;State Update&#34;] F --&gt; H G --&gt; H H --&gt; I[&#34;Next Layer&#34;] D --&gt; J[&#34;Feature Transformation&#34;] J --&gt; I style A fill:#e3f2fd,stroke:#1976d2,stroke-width:2px,color:#1e293b style B fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px,color:#1e293b style C fill:#e8f5e8,stroke:#388e3c,stroke-width:2px,color:#1e293b style D fill:#fff3e0,stroke:#f57c00,stroke-width:2px,color:#1e293b style H fill:#fce4ec,stroke:#c2185b,stroke-width:3px,color:#1e293b style I fill:#f1f8e9,stroke:#689f38,stroke-width:2px,color:#1e293b </div> </div> <div class="grid grid-cols-1 md:grid-cols-2 gap-8 mb-12"> <div class="bg-white rounded-lg p-6 shadow-sm border border-base-300"> <h5 class="font-semibold text-primary mb-4">Time-Mixing Block</h5> <p class="text-secondary text-sm mb-4"> 负责处理时序信息交互,通过RWKV机制实现历史信息的递归聚合。引入Token Shift机制增强局部上下文敏感性。 </p> <ul class="text-sm text-secondary space-y-2"> <li>• Receptance门控历史信息</li> <li>• Weighted时间衰减机制</li> <li>• Key-Value信息传递</li> </ul> </div> <div class="bg-white rounded-lg p-6 shadow-sm border border-base-300"> <h5 class="font-semibold text-primary mb-4">Channel-Mixing Block</h5> <p class="text-secondary text-sm mb-4"> 负责特征维度的非线性变换,功能类似Transformer的前馈网络,但完全并行化,不影响推理效率。 </p> <ul class="text-sm text-secondary space-y-2"> <li>• 特征空间线性变换</li> <li>• SiLU/GeLU激活函数</li> <li>• 时间步独立处理</li> </ul> </div> </div> </div> <div class="mb-16"> <h3 class="font-display text-2xl font-semibold text-primary mb-8">计算效率优势</h3> <div class="grid grid-cols-1 lg:grid-cols-3 gap-8 mb-12"> <div class="bg-green-50 rounded-lg p-6 border border-green-200"> <div class="flex items-center mb-4"> <i class="fas fa-clock text-green-600 text-2xl mr-3"></i> <h4 class="font-semibold text-primary">线性时间复杂度</h4> </div> <p class="text-sm text-secondary mb-4"> O(n) vs Transformer的O(n²),处理10K token时效率提升可达三个数量级 </p> <div class="text-xs text-green-700 bg-green-100 p-2 rounded"> 每token处理时间恒定,不受序列长度影响 </div> </div> <div class="bg-blue-50 rounded-lg p-6 border border-blue-200"> <div class="flex items-center mb-4"> <i class="fas fa-memory text-blue-600 text-2xl mr-3"></i> <h4 class="font-semibold text-primary">常数级内存</h4> </div> <p class="text-sm text-secondary mb-4"> 状态内存仅400KB,与序列长度无关,相比Transformer提升超100万倍 </p> <div class="text-xs text-blue-700 bg-blue-100 p-2 rounded"> 无需KV-Cache,内存占用与序列长度零相关 </div> </div> <div class="bg-purple-50 rounded-lg p-6 border border-purple-200"> <div class="flex items-center mb-4"> <i class="fas fa-microchip text-purple-600 text-2xl mr-3"></i> <h4 class="font-semibold text-primary">硬件友好</h4> </div> <p class="text-sm text-secondary mb-4"> 规则化内存访问模式,充分利用GPU计算单元,避免内存带宽瓶颈 </p> <div class="text-xs text-purple-700 bg-purple-100 p-2 rounded"> 消费级GPU即可实现高性能推理 </div> </div> </div> </div> <div class="mb-16"> <h3 class="font-display text-2xl font-semibold text-primary mb-8">架构演进历程</h3> <div class="overflow-x-auto"> <table class="w-full bg-white rounded-lg shadow-sm border border-base-300"> <thead class="bg-base-200"> <tr> <th class="px-6 py-4 text-left font-semibold text-primary">版本</th> <th class="px-6 py-4 text-left font-semibold text-primary">代号</th> <th class="px-6 py-4 text-left font-semibold text-primary">核心创新</th> <th class="px-6 py-4 text-left font-semibold text-primary">最大参数</th> </tr> </thead> <tbody class="divide-y divide-base-300"> <tr> <td class="px-6 py-4 text-sm">RWKV-4</td> <td class="px-6 py-4 text-sm">基础版</td> <td class="px-6 py-4 text-sm">验证混合架构可行性</td> <td class="px-6 py-4 text-sm">14B</td> </tr> <tr> <td class="px-6 py-4 text-sm">RWKV-5/6</td> <td class="px-6 py-4 text-sm">Eagle &amp; Finch</td> <td class="px-6 py-4 text-sm">矩阵值状态表示</td> <td class="px-6 py-4 text-sm">7B</td> </tr> <tr> <td class="px-6 py-4 text-sm">RWKV-7</td> <td class="px-6 py-4 text-sm">Goose</td> <td class="px-6 py-4 text-sm">动态状态演化、广义Delta规则</td> <td class="px-6 py-4 text-sm">13B</td> </tr> <tr class="bg-blue-50"> <td class="px-6 py-4 text-sm font-semibold">RWKV-7-G1</td> <td class="px-6 py-4 text-sm font-semibold">GooseOne</td> <td class="px-6 py-4 text-sm">推理专用优化版本</td> <td class="px-6 py-4 text-sm font-semibold">2.9B-13B</td> </tr> </tbody> </table> </div> </div> </section> <!-- Performance Analysis --> <section id="performance" class="mb-20"> <h2 class="font-display text-4xl font-bold text-primary mb-12">推理性能深度分析</h2> <div class="mb-16"> <h3 class="font-display text-2xl font-semibold text-primary mb-8">消费级GPU性能实测</h3> <div class="grid grid-cols-1 lg:grid-cols-3 gap-8 mb-12"> <div class="bg-white rounded-lg p-6 shadow-sm border border-base-300"> <h4 class="font-semibold text-primary mb-4">RTX 4090平台</h4> <div class="space-y-4"> <div class="flex justify-between items-center"> <span class="text-sm text-secondary">FP16精度</span> <span class="font-mono text-sm font-semibold">95.98 t/s</span> </div> <div class="flex justify-between items-center"> <span class="text-sm text-secondary">INT8精度</span> <span class="font-mono text-sm font-semibold text-blue-600">108.22 t/s</span> </div> <div class="flex justify-between items-center"> <span class="text-sm text-secondary">NF4精度</span> <span class="font-mono text-sm font-semibold text-green-600">115.46 t/s</span> </div> </div> </div> <div class="bg-white rounded-lg p-6 shadow-sm border border-base-300"> <h4 class="font-semibold text-primary mb-4">RTX 4060 Ti 8GB</h4> <div class="space-y-4"> <div class="flex justify-between items-center"> <span class="text-sm text-secondary">FP16精度</span> <span class="font-mono text-sm font-semibold">43.92 t/s</span> </div> <div class="flex justify-between items-center"> <span class="text-sm text-secondary">INT8精度</span> <span class="font-mono text-sm font-semibold text-blue-600">62.93 t/s</span> </div> <div class="flex justify-between items-center"> <span class="text-sm text-secondary">NF4精度</span> <span class="font-mono text-sm font-semibold text-green-600">86.03 t/s</span> </div> </div> </div> <div class="bg-white rounded-lg p-6 shadow-sm border border-base-300"> <h4 class="font-semibold text-primary mb-4">AMD RX 7900 XTX</h4> <div class="space-y-4"> <div class="flex justify-between items-center"> <span class="text-sm text-secondary">INT8精度</span> <span class="font-mono text-sm font-semibold text-green-600">137.36 t/s</span> </div> <div class="text-xs text-green-700 bg-green-100 p-2 rounded mt-4"> 比RTX 4090同配置快27%,展现优秀硬件无关性 </div> </div> </div> </div> </div> <div class="mb-16"> <h3 class="font-display text-2xl font-semibold text-primary mb-8">量化技术效果分析</h3> <div class="bg-gradient-to-r from-yellow-50 to-orange-50 rounded-lg p-8 border border-yellow-200 mb-8"> <h4 class="font-semibold text-primary mb-4">精度-速度-显存权衡</h4> <div class="grid grid-cols-1 md:grid-cols-3 gap-6"> <div class="text-center"> <div class="text-2xl font-bold text-blue-600 mb-2">INT8</div> <div class="text-sm text-secondary">速度+12.8%,显存-33.9%</div> <div class="text-xs text-blue-700 mt-2">质量与FP16相当</div> </div> <div class="text-center"> <div class="text-2xl font-bold text-green-600 mb-2">NF4</div> <div class="text-sm text-secondary">速度+20.3%,显存-59.3%</div> <div class="text-xs text-green-700 mt-2">推荐用于资源受限场景</div> </div> <div class="text-center"> <div class="text-2xl font-bold text-purple-600 mb-2">FP16</div> <div class="text-sm text-secondary">最高质量,无精度损失</div> <div class="text-xs text-purple-700 mt-2">质量敏感场景首选</div> </div> </div> </div> <div class="overflow-x-auto"> <table class="w-full bg-white rounded-lg shadow-sm border border-base-300"> <thead class="bg-base-200"> <tr> <th class="px-6 py-4 text-left font-semibold text-primary">推理引擎</th> <th class="px-6 py-4 text-left font-semibold text-primary">支持精度</th> <th class="px-6 py-4 text-left font-semibold text-primary">平台覆盖</th> <th class="px-6 py-4 text-left font-semibold text-primary">核心特点</th> </tr> </thead> <tbody class="divide-y divide-base-300"> <tr class="bg-blue-50"> <td class="px-6 py-4 text-sm font-semibold">web-rwkv</td> <td class="px-6 py-4 text-sm">FP16/INT8/NF4</td> <td class="px-6 py-4 text-sm">NVIDIA/AMD/Intel</td> <td class="px-6 py-4 text-sm">推荐方案,跨平台高性能</td> </tr> <tr> <td class="px-6 py-4 text-sm">llama.cpp</td> <td class="px-6 py-4 text-sm">Q4_0-Q8_0等GGUF</td> <td class="px-6 py-4 text-sm">跨平台</td> <td class="px-6 py-4 text-sm">生态成熟,格式丰富</td> </tr> <tr> <td class="px-6 py-4 text-sm">RWKV pip</td> <td class="px-6 py-4 text-sm">FP16/INT8(有限)</td> <td class="px-6 py-4 text-sm">NVIDIA (CUDA)</td> <td class="px-6 py-4 text-sm">官方实现,易用性高</td> </tr> <tr> <td class="px-6 py-4 text-sm">Ai00 Server</td> <td class="px-6 py-4 text-sm">INT8/NF4</td> <td class="px-6 py-4 text-sm">跨平台</td> <td class="px-6 py-4 text-sm">开箱即用API服务器</td> </tr> </tbody> </table> </div> </div> <div class="mb-16"> <h3 class="font-display text-2xl font-semibold text-primary mb-8">长序列处理能力</h3> <div class="grid grid-cols-1 lg:grid-cols-2 gap-12"> <div> <h4 class="text-xl font-semibold text-primary mb-6">理论无限上下文支持</h4> <p class="text-secondary mb-6 leading-relaxed"> RWKV采用递归状态机制而非KV缓存,处理序列的内存占用与序列长度完全无关。无论是100 token还是100万token,所需的状态内存都是固定的约400KB。 </p> <div class="bg-blue-50 border-l-4 border-blue-400 p-4 rounded-r-lg mb-6"> <p class="text-sm text-blue-800"> <strong>架构层面原生能力:</strong>&#34;无限上下文&#34;无需特殊技巧即可实现,而非通过滑动窗口等近似方法 </p> </div> <img src="https://kimi-web-img.moonshot.cn/img/www.rwkv.cn/139587058ebcb66896f4985e95cc11292b1ed46f" alt="RWKV模型状态内存固定示意图" class="w-full rounded-lg shadow-md" size="medium" aspect="wide" query="RWKV模型状态内存固定" referrerpolicy="no-referrer" data-modified="1" data-score="0.00"/> </div> <div> <h4 class="text-xl font-semibold text-primary mb-6">实际部署限制</h4> <div class="space-y-4"> <div class="bg-white rounded-lg p-4 shadow-sm border border-base-300"> <h5 class="font-semibold text-primary mb-2">训练上下文限制</h5> <p class="text-sm text-secondary">模型在特定长度上训练,超出范围性能可能衰减,但可通过微调扩展</p> </div> <div class="bg-white rounded-lg p-4 shadow-sm border border-base-300"> <h5 class="font-semibold text-primary mb-2">数值稳定性</h5> <p class="text-sm text-secondary">极长序列递归计算可能累积数值误差,需要特定处理策略</p> </div> <div class="bg-white rounded-lg p-4 shadow-sm border border-base-300"> <h5 class="font-semibold text-primary mb-2">应用层设计</h5> <p class="text-sm text-secondary">大多数应用不需要真正无限上下文,RWKV的能力更多作为安全网</p> </div> </div> <div class="mt-8 bg-yellow-50 rounded-lg p-4 border border-yellow-200"> <h5 class="font-semibold text-yellow-800 mb-2">性能对比(100K token)</h5> <div class="text-sm text-yellow-700"> <div class="flex justify-between mb-1"> <span>RWKV:</span> <span class="font-mono">~400KB</span> </div> <div class="flex justify-between mb-2"> <span>Transformer:</span> <span class="font-mono">~20GB+</span> </div> <div class="text-xs font-semibold">内存效率优势超过50,000倍</div> </div> </div> </div> </div> </div> </section> <!-- Benchmarks --> <section id="benchmarks" class="mb-20"> <h2 class="font-display text-4xl font-bold text-primary mb-12">基准测试与模型对比</h2> <div class="mb-16"> <h3 class="font-display text-2xl font-semibold text-primary mb-8">标准学术基准测试</h3> <div class="bg-orange-50 border-l-4 border-orange-400 p-6 rounded-r-lg mb-8"> <h4 class="font-semibold text-orange-800 mb-2">数据可用性说明</h4> <p class="text-orange-700 text-sm"> 截至2026年1月,RWKV在MMLU、HellaSwag、GSM8K等标准学术基准上的公开评估数据存在显著缺口。与Llama、Mistral等主流模型不同,RWKV项目未系统发布完整的基准测试报告。 </p> </div> <div class="overflow-x-auto mb-8"> <table class="w-full bg-white rounded-lg shadow-sm border border-base-300"> <thead class="bg-base-200"> <tr> <th class="px-6 py-4 text-left font-semibold text-primary">模型</th> <th class="px-6 py-4 text-left font-semibold text-primary">参数</th> <th class="px-6 py-4 text-left font-semibold text-primary">MMLU</th> <th class="px-6 py-4 text-left font-semibold text-primary">HellaSwag</th> <th class="px-6 py-4 text-left font-semibold text-primary">GSM8K</th> <th class="px-6 py-4 text-left font-semibold text-primary">训练数据</th> </tr> </thead> <tbody class="divide-y divide-base-300"> <tr> <td class="px-6 py-4 text-sm">Llama-2-7B</td> <td class="px-6 py-4 text-sm font-mono">7B</td> <td class="px-6 py-4 text-sm">~45%</td> <td class="px-6 py-4 text-sm">~75%</td> <td class="px-6 py-4 text-sm">~15%</td> <td class="px-6 py-4 text-sm">2T tokens</td> </tr> <tr class="bg-blue-50"> <td class="px-6 py-4 text-sm font-semibold">Mistral-7B-v0.1</td> <td class="px-6 py-4 text-sm font-mono">7B</td> <td class="px-6 py-4 text-sm">~60%</td> <td class="px-6 py-4 text-sm">~80%</td> <td class="px-6 py-4 text-sm">~37%</td> <td class="px-6 py-4 text-sm">~2T+ tokens</td> </tr> <tr class="bg-green-50"> <td class="px-6 py-4 text-sm font-semibold">Llama-3-8B</td> <td class="px-6 py-4 text-sm font-mono">8B</td> <td class="px-6 py-4 text-sm">~68%</td> <td class="px-6 py-4 text-sm">~78%</td> <td class="px-6 py-4 text-sm">~46%</td> <td class="px-6 py-4 text-sm">15T+ tokens</td> </tr> <tr> <td class="px-6 py-4 text-sm">Gemma-2-9B</td> <td class="px-6 py-4 text-sm font-mono">9B</td> <td class="px-6 py-4 text-sm">~65%</td> <td class="px-6 py-4 text-sm">~82%</td> <td class="px-6 py-4 text-sm">~50%</td> <td class="px-6 py-4 text-sm">6T tokens</td> </tr> <tr class="bg-yellow-50"> <td class="px-6 py-4 text-sm font-semibold">RWKV-5/6 (估计)</td> <td class="px-6 py-4 text-sm font-mono">7B</td> <td class="px-6 py-4 text-sm">55-65%?</td> <td class="px-6 py-4 text-sm">?</td> <td class="px-6 py-4 text-sm">?</td> <td class="px-6 py-4 text-sm">2.25T tokens</td> </tr> </tbody> </table> </div> <div class="bg-blue-50 rounded-lg p-6 border border-blue-200"> <h4 class="font-semibold text-blue-800 mb-3">开发者报告的性能定位</h4> <p class="text-blue-700 text-sm mb-4"> RWKV官方博客报告EagleX v2(RWKV-v5,14B,2.25T tokens)<strong>&#34;明显超过了训练了2万亿token的Llama-2-7B,并且接近Mistral-7B-v0.1和Llama-3-8B&#34;</strong> </p> <div class="text-xs text-blue-600"> 数据来源:<a href="https://blog.rwkv.com/p/eaglex-v2-soaring-past-llama2-7b" class="underline hover:text-blue-800">RWKV官方博客</a> </div> </div> </div> <div class="mb-16"> <h3 class="font-display text-2xl font-semibold text-primary mb-8">实际任务表现</h3> <div class="grid grid-cols-1 md:grid-cols-3 gap-8"> <div class="bg-white rounded-lg p-6 shadow-sm border border-base-300"> <div class="flex items-center mb-4"> <i class="fas fa-code text-green-600 text-xl mr-3"></i> <h4 class="font-semibold text-primary">代码生成</h4> </div> <p class="text-sm text-secondary mb-4"> RWKV-7-G1系列明确强化了代码和数学能力。递归状态机制有助于维护语法结构的长期一致性。 </p> <div class="text-xs text-green-700 bg-green-100 p-2 rounded"> 社区已开发专用代码补全模型 </div> </div> <div class="bg-white rounded-lg p-6 shadow-sm border border-base-300"> <div class="flex items-center mb-4"> <i class="fas fa-comments text-blue-600 text-xl mr-3"></i> <h4 class="font-semibold text-primary">多轮对话</h4> </div> <p class="text-sm text-secondary mb-4"> 恒定每步延迟确保对话长度增加时响应时间稳定,不会出现Transformer中常见的&#34;对话越久越慢&#34;现象。 </p> <div class="text-xs text-blue-700 bg-blue-100 p-2 rounded"> 支持真正的&#34;无限记忆&#34;助手 </div> </div> <div class="bg-white rounded-lg p-6 shadow-sm border border-base-300"> <div class="flex items-center mb-4"> <i class="fas fa-stream text-purple-600 text-xl mr-3"></i> <h4 class="font-semibold text-primary">流式生成</h4> </div> <p class="text-sm text-secondary mb-4"> 明确推荐用于&#34;real-time streaming, chaining short prompts&#34;场景,特别适合语音助手、实时翻译应用。 </p> <div class="text-xs text-purple-700 bg-purple-100 p-2 rounded"> 状态机制天然支持高效提示链接 </div> </div> </div> </div> </section> <!-- Applications --> <section id="applications" class="mb-20"> <h2 class="font-display text-4xl font-bold text-primary mb-12">应用场景与部署实践</h2> <div class="mb-16"> <h3 class="font-display text-2xl font-semibold text-primary mb-8">核心应用场景</h3> <div class="grid grid-cols-1 lg:grid-cols-2 gap-12 mb-12"> <div class="space-y-6"> <div class="bg-white rounded-lg p-6 shadow-sm border border-base-300"> <div class="flex items-center mb-4"> <i class="fas fa-file-alt text-green-600 text-xl mr-3"></i> <h4 class="font-semibold text-primary">长文档处理</h4> </div> <p class="text-sm text-secondary mb-4"> 法律合同审查、学术论文综述、金融报告摘要等需要处理数万至数十万token的场景。 </p> <div class="text-xs text-green-700 bg-green-100 p-2 rounded"> 完整摄入文档,保持全局连贯理解,无需分段截断 </div> </div> <div class="bg-white rounded-lg p-6 shadow-sm border border-base-300"> <div class="flex items-center mb-4"> <i class="fas fa-robot text-blue-600 text-xl mr-3"></i> <h4 class="font-semibold text-primary">实时交互系统</h4> </div> <p class="text-sm text-secondary mb-4"> 聊天机器人、虚拟助手等需要低延迟响应的应用场景。 </p> <div class="text-xs text-blue-700 bg-blue-100 p-2 rounded"> 恒定每token推理时间,确保无论对话历史多长,生成速度稳定 </div> </div> </div> <div class="space-y-6"> <div class="bg-white rounded-lg p-6 shadow-sm border border-base-300"> <div class="flex items-center mb-4"> <i class="fas fa-mobile-alt text-purple-600 text-xl mr-3"></i> <h4 class="font-semibold text-primary">边缘与端侧部署</h4> </div> <p class="text-sm text-secondary mb-4"> 量化后的轻量级推理需求,适合资源受限环境。 </p> <div class="text-xs text-purple-700 bg-purple-100 p-2 rounded"> 骁龙8 Gen 3平台A16W4配置达到31.3 tokens/s </div> </div> <div class="bg-white rounded-lg p-6 shadow-sm border border-base-300"> <div class="flex items-center mb-4"> <i class="fas fa-stream text-orange-600 text-xl mr-3"></i> <h4 class="font-semibold text-primary">流式内容生成</h4> </div> <p class="text-sm text-secondary mb-4"> 连续短提示的链式处理,Agent系统、工作流自动化等应用。 </p> <div class="text-xs text-orange-700 bg-orange-100 p-2 rounded"> 避免注意力矩阵爆炸,状态向量作为紧凑上下文表示 </div> </div> </div> </div> </div> <div class="mb-16"> <h3 class="font-display text-2xl font-semibold text-primary mb-8">模型版本选择指南</h3> <div class="overflow-x-auto"> <table class="w-full bg-white rounded-lg shadow-sm border border-base-300"> <thead class="bg-base-200"> <tr> <th class="px-6 py-4 text-left font-semibold text-primary">显存预算</th> <th class="px-6 py-4 text-left font-semibold text-primary">推荐配置</th> <th class="px-6 py-4 text-left font-semibold text-primary">预期性能</th> <th class="px-6 py-4 text-left font-semibold text-primary">典型应用</th> </tr> </thead> <tbody class="divide-y divide-base-300"> <tr> <td class="px-6 py-4 text-sm font-semibold text-red-600">&lt;4GB</td> <td class="px-6 py-4 text-sm">RWKV7-G1-2.9B <strong>nf4</strong></td> <td class="px-6 py-4 text-sm">20-40 t/s (GPU), 可用 (CPU)</td> <td class="px-6 py-4 text-sm">边缘设备,轻量助手,离线处理</td> </tr> <tr class="bg-blue-50"> <td class="px-6 py-4 text-sm font-semibold text-blue-600">8GB</td> <td class="px-6 py-4 text-sm">RWKV7-G1-2.9B <strong>int8</strong> 或 RWKV-6 7B nf4</td> <td class="px-6 py-4 text-sm">60-110 t/s</td> <td class="px-6 py-4 text-sm font-semibold">标准部署,日常对话,文档处理</td> </tr> <tr> <td class="px-6 py-4 text-sm font-semibold text-green-600">16GB</td> <td class="px-6 py-4 text-sm">RWKV-6/7 7B int8 或 14B nf4</td> <td class="px-6 py-4 text-sm">40-80 t/s</td> <td class="px-6 py-4 text-sm">高性能应用,长文本,代码生成</td> </tr> <tr> <td class="px-6 py-4 text-sm font-semibold text-purple-600">24GB+</td> <td class="px-6 py-4 text-sm">RWKV-6/7 14B <strong>fp16/int8</strong></td> <td class="px-6 py-4 text-sm">30-60 t/s</td> <td class="px-6 py-4 text-sm">服务器部署,质量优先,研究用途</td> </tr> <tr class="bg-yellow-50"> <td class="px-6 py-4 text-sm font-semibold text-yellow-600">推理专用</td> <td class="px-6 py-4 text-sm font-semibold">RWKV-7-G1系列</td> <td class="px-6 py-4 text-sm">依配置</td> <td class="px-6 py-4 text-sm">数学推理,代码生成,多步思考</td> </tr> </tbody> </table> </div> </div> <div class="mb-16"> <h3 class="font-display text-2xl font-semibold text-primary mb-8">部署工具与教程</h3> <div class="grid grid-cols-1 lg:grid-cols-2 gap-8"> <div class="bg-white rounded-lg p-6 shadow-sm border border-base-300"> <h4 class="font-semibold text-primary mb-4 flex items-center"> <i class="fas fa-cog text-green-600 mr-2"></i> web-rwkv引擎(推荐方案) </h4> <div class="text-sm text-secondary mb-4"> <strong>环境依赖:</strong>Rust工具链(cargo)、Vulkan驱动,无需CUDA Toolkit或PyTorch </div> <div class="bg-gray-100 rounded p-4 mb-4"> <div class="text-xs font-mono text-gray-700 mb-2"># 安装命令</div> <code class="text-xs"> git clone https://github.com/cryscan/web-rwkv<br/> cd web-rwkv &amp;&amp; cargo build --release </code> </div> <div class="text-xs text-green-700 bg-green-100 p-2 rounded"> 支持策略:&#34;cuda fp16&#34;(全精度)、&#34;cuda int8&#34;(推荐平衡)、&#34;cuda nf4&#34;(极致效率) </div> </div> <div class="bg-white rounded-lg p-6 shadow-sm border border-base-300"> <h4 class="font-semibold text-primary mb-4 flex items-center"> <i class="fas fa-terminal text-blue-600 mr-2"></i> llama.cpp生态 </h4> <div class="text-sm text-secondary mb-4"> <strong>优势:</strong>生态成熟,格式丰富,工具链统一,跨平台支持 </div> <div class="bg-gray-100 rounded p-4 mb-4"> <div class="text-xs font-mono text-gray-700 mb-2"># 转换与量化</div> <code class="text-xs"> python convert-rwkv-to-gguf.py --input model.pth<br/> ./quantize model.gguf model-Q8_0.gguf Q8_0 </code> </div> <div class="text-xs text-blue-700 bg-blue-100 p-2 rounded"> 支持多种GGUF格式,从Q4_0到Q8_0不等 </div> </div> </div> <div class="mt-8 bg-blue-50 rounded-lg p-6 border border-blue-200"> <h4 class="font-semibold text-blue-800 mb-4">其他部署方案</h4> <div class="grid grid-cols-1 md:grid-cols-3 gap-4 text-sm"> <div> <strong class="text-blue-800">RWKV pip</strong> <br/> <span class="text-blue-700">pip install rwkv,适合快速原型</span> </div> <div> <strong class="text-blue-800">Ai00 Server</strong> <br/> <span class="text-blue-700">开箱即用API服务器,无PyTorch依赖</span> </div> <div> <strong class="text-blue-800">云端Docker</strong> <br/> <span class="text-blue-700">容器化web-rwkv服务,配合FastAPI或Triton</span> </div> </div> </div> </div> </section> <!-- Advantages and Future --> <section id="advantages" class="mb-20"> <h2 class="font-display text-4xl font-bold text-primary mb-12">竞争优势与发展展望</h2> <div class="mb-16"> <h3 class="font-display text-2xl font-semibold text-primary mb-8">核心差异化优势</h3> <div class="grid grid-cols-1 md:grid-cols-3 gap-8 mb-12"> <div class="bg-white rounded-lg p-6 shadow-sm border border-base-300"> <div class="flex items-center mb-4"> <i class="fas fa-balance-scale text-green-600 text-2xl mr-3"></i> <h4 class="font-semibold text-primary">效率与性能平衡</h4> </div> <p class="text-sm text-secondary mb-4"> 重新定义效率与性能的权衡边界,在线性复杂度下仍可实现competitive的模型质量。 </p> <div class="text-xs text-green-700 bg-green-100 p-2 rounded"> 为AI能力的普惠化提供技术基础 </div> </div> <div class="bg-white rounded-lg p-6 shadow-sm border border-base-300"> <div class="flex items-center mb-4"> <i class="fas fa-desktop text-blue-600 text-2xl mr-3"></i> <h4 class="font-semibold text-primary">硬件友好性</h4> </div> <p class="text-sm text-secondary mb-4"> 消费级GPU上的卓越推理效率,RTX 4060 Ti即可流畅运行7B级别模型。 </p> <div class="text-xs text-blue-700 bg-blue-100 p-2 rounded"> 使高质量语言模型真正accessible到普通用户 </div> </div> <div class="bg-white rounded-lg p-6 shadow-sm border border-base-300"> <div class="flex items-center mb-4"> <i class="fas fa-code text-purple-600 text-2xl mr-3"></i> <h4 class="font-semibold text-primary">架构简洁性</h4> </div> <p class="text-sm text-secondary mb-4"> 无KV-Cache设计大幅简化推理引擎实现,核心代码可数百行内完成。 </p> <div class="text-xs text-purple-700 bg-purple-100 p-2 rounded"> 便于审计、定制和移植到新型硬件 </div> </div> </div> </div> <div class="mb-16"> <h3 class="font-display text-2xl font-semibold text-primary mb-8">当前局限性与挑战</h3> <div class="bg-red-50 rounded-lg p-8 border border-red-200 mb-8"> <h4 class="font-semibold text-red-800 mb-6">主要挑战领域</h4> <div class="grid grid-cols-1 md:grid-cols-3 gap-6"> <div class="bg-white rounded-lg p-4 border border-red-100"> <h5 class="font-semibold text-red-700 mb-2">生态系统成熟度</h5> <p class="text-sm text-red-600 mb-3"> 微调模型、应用案例、社区规模落后于Llama/Mistral </p> <div class="text-xs text-red-500"> 缓解:持续开源,社区建设 </div> </div> <div class="bg-white rounded-lg p-4 border border-red-100"> <h5 class="font-semibold text-red-700 mb-2">工具链完善度</h5> <p class="text-sm text-red-600 mb-3"> RLHF微调、多模态扩展、企业级工具待完善 </p> <div class="text-xs text-red-500"> 缓解:重点投入,生态合作 </div> </div> <div class="bg-white rounded-lg p-4 border border-red-100"> <h5 class="font-semibold text-red-700 mb-2">基准测试透明度</h5> <p class="text-sm text-red-600 mb-3"> 缺乏系统的MMLU/HellaSwag/GSM8K公开报告 </p> <div class="text-xs text-red-500"> 缓解:第三方评估,学术合作 </div> </div> </div> </div> </div> <div class="mb-16"> <h3 class="font-display text-2xl font-semibold text-primary mb-8">技术演进方向</h3> <div class="space-y-8"> <div class="bg-gradient-to-r from-blue-50 to-indigo-50 rounded-lg p-8 border border-blue-200"> <h4 class="font-semibold text-primary mb-4 flex items-center"> <i class="fas fa-rocket text-blue-600 mr-3"></i> RWKV-7及后续版本 </h4> <p class="text-secondary mb-4"> 动态状态演化机制为复杂推理、规划、工具使用奠定基础,能力边界持续拓展。 </p> <div class="grid grid-cols-1 md:grid-cols-2 gap-4 text-sm"> <div class="bg-white rounded p-3"> <strong class="text-blue-700">广义Delta规则</strong> <br/> <span class="text-blue-600">突破TC⁰复杂度限制</span> </div> <div class="bg-white rounded p-3"> <strong class="text-blue-700">动态状态演化</strong> <br/> <span class="text-blue-600">输入依赖的状态更新</span> </div> </div> </div> <div class="bg-gradient-to-r from-green-50 to-emerald-50 rounded-lg p-8 border border-green-200"> <h4 class="font-semibold text-primary mb-4 flex items-center"> <i class="fas fa-eye text-green-600 mr-3"></i> 多模态扩展 </h4> <p class="text-secondary mb-4"> VisualRWKV等视觉-语言融合方向,将架构优势延伸至跨模态场景。 </p> <div class="bg-white rounded p-4"> <div class="text-sm text-green-700"> <strong>核心思路:</strong>将线性复杂度优势从文本扩展到图像、音频等多模态数据处理 </div> </div> </div> <div class="bg-gradient-to-r from-purple-50 to-pink-50 rounded-lg p-8 border border-purple-200"> <h4 class="font-semibold text-primary mb-4 flex items-center"> <i class="fas fa-brain text-purple-600 mr-3"></i> 推理能力强化 </h4> <p class="text-secondary mb-4"> GooseOne等专用版本的迭代,针对数学、代码、科学推理等任务深度优化。 </p> <div class="grid grid-cols-1 md:grid-cols-3 gap-4 text-sm"> <div class="bg-white rounded p-3 text-center"> <div class="text-purple-700 font-semibold">数学推理</div> <div class="text-purple-600">符号计算、定理证明</div> </div> <div class="bg-white rounded p-3 text-center"> <div class="text-purple-700 font-semibold">代码生成</div> <div class="text-purple-600">结构化编程、调试</div> </div> <div class="bg-white rounded p-3 text-center"> <div class="text-purple-700 font-semibold">科学推理</div> <div class="text-purple-600">实验设计、数据分析</div> </div> </div> </div> </div> </div> <div class="bg-gradient-to-r from-slate-50 to-gray-50 rounded-2xl p-8 border border-slate-200"> <h3 class="font-display text-2xl font-semibold text-primary mb-6 text-center">长期愿景</h3> <div class="text-center mb-6"> <p class="text-lg text-secondary mb-4"> RWKV的长期愿景是成为<strong class="text-primary">高效AI的默认选择</strong> </p> <div class="text-sm text-secondary"> 当应用场景对效率敏感时,开发者首先考虑RWKV架构 </div> </div> <div class="grid grid-cols-1 md:grid-cols-3 gap-6"> <div class="text-center"> <div class="bg-blue-100 rounded-full w-16 h-16 flex items-center justify-center mx-auto mb-3"> <i class="fas fa-cogs text-blue-600 text-xl"></i> </div> <div class="font-semibold text-primary mb-2">技术创新</div> <div class="text-sm text-secondary">持续架构优化,能力边界拓展</div> </div> <div class="text-center"> <div class="bg-green-100 rounded-full w-16 h-16 flex items-center justify-center mx-auto mb-3"> <i class="fas fa-users text-green-600 text-xl"></i> </div> <div class="font-semibold text-primary mb-2">生态建设</div> <div class="text-sm text-secondary">工具链完善,社区规模扩大</div> </div> <div class="text-center"> <div class="bg-purple-100 rounded-full w-16 h-16 flex items-center justify-center mx-auto mb-3"> <i class="fas fa-seedling text-purple-600 text-xl"></i> </div> <div class="font-semibold text-primary mb-2">社区培育</div> <div class="text-sm text-secondary">开发者支持,应用场景丰富</div> </div> </div> </div> </section> </div> <!-- Footer --> <footer class="bg-primary text-white py-12"> <div class="max-w-6xl mx-auto px-8"> <div class="grid grid-cols-1 md:grid-cols-3 gap-8"> <div> <h4 class="font-display text-xl font-semibold mb-4">RWKV资源</h4> <ul class="space-y-2 text-sm text-gray-300"> <li> <a href="https://github.com/BlinkDL/RWKV-LM" class="hover:text-white transition-colors">GitHub仓库</a> </li> <li> <a href="https://wiki.rwkv.com" class="hover:text-white transition-colors">技术文档</a> </li> <li> <a href="https://blog.rwkv.com" class="hover:text-white transition-colors">官方博客</a> </li> <li> <a href="https://www.rwkv.cn" class="hover:text-white transition-colors">中文社区</a> </li> </ul> </div> <div> <h4 class="font-display text-xl font-semibold mb-4">性能数据参考</h4> <ul class="space-y-2 text-sm text-gray-300"> <li> <a href="https://wiki.rwkv.com/basic/rwkv-performance-data.html" class="hover:text-white transition-colors">性能基准测试</a> </li> <li> <a href="https://wiki.rwkv.com/basic/how-to-play.html" class="hover:text-white transition-colors">部署指南</a> </li> <li> <a href="https://blog.rwkv.com/p/eaglex-v2-soaring-past-llama2-7b" class="hover:text-white transition-colors">EagleX v2评估</a> </li> </ul> </div> <div> <h4 class="font-display text-xl font-semibold mb-4">研究参考</h4> <ul class="space-y-2 text-sm text-gray-300"> <li> <a href="https://deepfa.ir/en/blog/rwkv-architecture-rnn-transformer-hybrid" class="hover:text-white transition-colors">架构深度解析</a> </li> <li> <a href="https://www.techradar.com/pro/100x-less-compute-with-gpt-level-llm-performance" class="hover:text-white transition-colors">TechRadar报道</a> </li> <li> <a href="https://wiki.rwkv.com/basic/architecture.html" class="hover:text-white transition-colors">架构演进</a> </li> </ul> </div> </div> <div class="border-t border-gray-700 mt-8 pt-8 text-center text-sm text-gray-400"> <p>本报告基于截至2026年1月的公开信息整理,所有数据来源均已标注引用链接。</p> </div> </div> </footer> <script> // Mermaid configuration mermaid.initialize({ startOnLoad: true, theme: 'base', themeVariables: { primaryColor: '#e3f2fd', primaryTextColor: '#1e293b', primaryBorderColor: '#475569', lineColor: '#64748b', secondaryColor: '#f8fafc', tertiaryColor: '#f1f5f9', background: '#ffffff', mainBkg: '#ffffff', secondBkg: '#f8fafc', tertiaryBkg: '#f1f5f9', nodeBorder: '#475569', clusterBkg: '#f8fafc', defaultLinkColor: '#64748b', titleColor: '#1e293b', edgeLabelBackground: 'rgba(255, 255, 255, 0.9)', nodeTextColor: '#1e293b', // Enhanced contrast colors cScale0: '#e3f2fd', cScale1: '#f3e5f5', cScale2: '#e8f5e8', cScale3: '#fff3e0', cScale4: '#fce4ec', cScale5: '#f1f8e9', cScale6: '#e0f2f1', cScale7: '#fbe9e7' }, flowchart: { useMaxWidth: true, htmlLabels: true, curve: 'basis', padding: 20 }, fontSize: 13, fontFamily: 'Inter, system-ui, sans-serif' }); // Initialize Mermaid Controls for zoom and pan function initializeMermaidControls() { const containers = document.querySelectorAll('.mermaid-container'); containers.forEach(container => { const mermaidElement = container.querySelector('.mermaid'); let scale = 1; let isDragging = false; let startX, startY, translateX = 0, translateY = 0; // 触摸相关状态 let isTouch = false; let touchStartTime = 0; let initialDistance = 0; let initialScale = 1; let isPinching = false; // Zoom controls const zoomInBtn = container.querySelector('.zoom-in'); const zoomOutBtn = container.querySelector('.zoom-out'); const resetBtn = container.querySelector('.reset-zoom'); const fullscreenBtn = container.querySelector('.fullscreen'); function updateTransform() { mermaidElement.style.transform = `translate(${translateX}px, ${translateY}px) scale(${scale})`; if (scale > 1) { container.classList.add('zoomed'); } else { container.classList.remove('zoomed'); } mermaidElement.style.cursor = isDragging ? 'grabbing' : 'grab'; } if (zoomInBtn) { zoomInBtn.addEventListener('click', () => { scale = Math.min(scale * 1.25, 4); updateTransform(); }); } if (zoomOutBtn) { zoomOutBtn.addEventListener('click', () => { scale = Math.max(scale / 1.25, 0.3); if (scale <= 1) { translateX = 0; translateY = 0; } updateTransform(); }); } if (resetBtn) { resetBtn.addEventListener('click', () => { scale = 1; translateX = 0; translateY = 0; updateTransform(); }); } if (fullscreenBtn) { fullscreenBtn.addEventListener('click', () => { if (container.requestFullscreen) { container.requestFullscreen(); } else if (container.webkitRequestFullscreen) { container.webkitRequestFullscreen(); } else if (container.msRequestFullscreen) { container.msRequestFullscreen(); } }); } // Mouse Events mermaidElement.addEventListener('mousedown', (e) => { if (isTouch) return; // 如果是触摸设备,忽略鼠标事件 isDragging = true; startX = e.clientX - translateX; startY = e.clientY - translateY; mermaidElement.style.cursor = 'grabbing'; updateTransform(); e.preventDefault(); }); document.addEventListener('mousemove', (e) => { if (isDragging && !isTouch) { translateX = e.clientX - startX; translateY = e.clientY - startY; updateTransform(); } }); document.addEventListener('mouseup', () => { if (isDragging && !isTouch) { isDragging = false; mermaidElement.style.cursor = 'grab'; updateTransform(); } }); document.addEventListener('mouseleave', () => { if (isDragging && !isTouch) { isDragging = false; mermaidElement.style.cursor = 'grab'; updateTransform(); } }); // 获取两点之间的距离 function getTouchDistance(touch1, touch2) { return Math.hypot( touch2.clientX - touch1.clientX, touch2.clientY - touch1.clientY ); } // Touch Events - 触摸事件处理 mermaidElement.addEventListener('touchstart', (e) => { isTouch = true; touchStartTime = Date.now(); if (e.touches.length === 1) { // 单指拖动 isPinching = false; isDragging = true; const touch = e.touches[0]; startX = touch.clientX - translateX; startY = touch.clientY - translateY; } else if (e.touches.length === 2) { // 双指缩放 isPinching = true; isDragging = false; const touch1 = e.touches[0]; const touch2 = e.touches[1]; initialDistance = getTouchDistance(touch1, touch2); initialScale = scale; } e.preventDefault(); }, { passive: false }); mermaidElement.addEventListener('touchmove', (e) => { if (e.touches.length === 1 && isDragging && !isPinching) { // 单指拖动 const touch = e.touches[0]; translateX = touch.clientX - startX; translateY = touch.clientY - startY; updateTransform(); } else if (e.touches.length === 2 && isPinching) { // 双指缩放 const touch1 = e.touches[0]; const touch2 = e.touches[1]; const currentDistance = getTouchDistance(touch1, touch2); if (initialDistance > 0) { const newScale = Math.min(Math.max( initialScale * (currentDistance / initialDistance), 0.3 ), 4); scale = newScale; updateTransform(); } } e.preventDefault(); }, { passive: false }); mermaidElement.addEventListener('touchend', (e) => { // 重置状态 if (e.touches.length === 0) { isDragging = false; isPinching = false; initialDistance = 0; // 延迟重置isTouch,避免鼠标事件立即触发 setTimeout(() => { isTouch = false; }, 100); } else if (e.touches.length === 1 && isPinching) { // 从双指变为单指,切换为拖动模式 isPinching = false; isDragging = true; const touch = e.touches[0]; startX = touch.clientX - translateX; startY = touch.clientY - translateY; } updateTransform(); }); mermaidElement.addEventListener('touchcancel', (e) => { isDragging = false; isPinching = false; initialDistance = 0; setTimeout(() => { isTouch = false; }, 100); updateTransform(); }); // Enhanced wheel zoom with better center point handling container.addEventListener('wheel', (e) => { e.preventDefault(); const rect = container.getBoundingClientRect(); const centerX = rect.width / 2; const centerY = rect.height / 2; const delta = e.deltaY > 0 ? 0.9 : 1.1; const newScale = Math.min(Math.max(scale * delta, 0.3), 4); // Adjust translation to zoom towards center if (newScale !== scale) { const scaleDiff = newScale / scale; translateX = translateX * scaleDiff; translateY = translateY * scaleDiff; scale = newScale; if (scale <= 1) { translateX = 0; translateY = 0; } updateTransform(); } }); // Initialize display updateTransform(); }); } // Initialize mermaid controls after page loads document.addEventListener('DOMContentLoaded', function() { initializeMermaidControls(); }); // Smooth scrolling for anchor links document.querySelectorAll('a[href^="#"]').forEach(anchor => { anchor.addEventListener('click', function (e) { e.preventDefault(); const target = document.querySelector(this.getAttribute('href')); if (target) { target.scrollIntoView({ behavior: 'smooth', block: 'start' }); } }); }); // Highlight active section in TOC const sections = document.querySelectorAll('section[id]'); const tocLinks = document.querySelectorAll('#toc a[href^="#"]'); function highlightTOC() { let current = ''; sections.forEach(section => { const sectionTop = section.offsetTop; if (scrollY >= sectionTop - 200) { current = section.getAttribute('id'); } }); tocLinks.forEach(link => { link.classList.remove('bg-blue-100', 'text-blue-800', 'font-semibold'); if (link.getAttribute('href') === '#' + current) { link.classList.add('bg-blue-100', 'text-blue-800', 'font-semibold'); } }); } window.addEventListener('scroll', highlightTOC); // Mobile TOC toggle const mobileToggle = document.getElementById('mobile-toc-toggle'); const toc = document.getElementById('toc'); const tocBackdrop = document.getElementById('toc-backdrop'); const tocClose = document.getElementById('toc-toggle'); function openToc() { toc.classList.remove('-translate-x-full'); tocBackdrop.classList.remove('hidden'); document.body.classList.add('overflow-hidden'); } function closeToc() { toc.classList.add('-translate-x-full'); tocBackdrop.classList.add('hidden'); document.body.classList.remove('overflow-hidden'); } mobileToggle.addEventListener('click', openToc); tocClose.addEventListener('click', closeToc); tocBackdrop.addEventListener('click', closeToc); // Close TOC when clicking on links tocLinks.forEach(link => { link.addEventListener('click', closeToc); }); // Reset TOC on window resize window.addEventListener('resize', () => { if (window.innerWidth >= 1024) { closeToc(); } }); </script> </body></html>

讨论回复

2 条回复
✨步子哥 (steper) #1
02-13 03:00
<html><body> <!-- Hero Section --> <section class="relative min-h-[70vh] flex items-center hero-pattern"> <div class="absolute inset-0 gradient-overlay"></div> <div class="relative z-10 max-w-7xl mx-auto px-4 md:px-8 py-12"> <div class="grid grid-cols-1 lg:grid-cols-12 gap-6 md:gap-8 items-center"> <!-- Main Title Column --> <div class="lg:col-span-7"> <h1 class="font-display text-3xl sm:text-4xl lg:text-5xl xl:text-6xl font-bold text-white mb-4 sm:mb-6 text-shadow"> <em class="italic">RWKV模型</em> <br/> <span class="text-blue-200">深度研究报告</span> </h1> <p class="text-base sm:text-lg lg:text-xl text-gray-200 mb-6 sm:mb-8 leading-relaxed"> 突破Transformer效率瓶颈的线性复杂度架构革新 </p> <div class="flex flex-wrap gap-2 sm:gap-4 text-xs sm:text-sm text-gray-300"> <span class="bg-white/10 px-3 py-1 sm:px-4 sm:py-2 rounded-full backdrop-blur">RNN-Transformer混合架构</span> <span class="bg-white/10 px-3 py-1 sm:px-4 sm:py-2 rounded-full backdrop-blur">O(n)线性复杂度</span> <span class="bg-white/10 px-3 py-1 sm:px-4 sm:py-2 rounded-full backdrop-blur">消费级GPU部署</span> </div> </div> <!-- Visual Element Column --> <div class="lg:col-span-5 mt-8 lg:mt-0"> <div class="grid grid-cols-1 sm:grid-cols-2 gap-3 sm:gap-4"> <div class="bg-white/10 backdrop-blur-sm rounded-lg p-4 sm:p-6 border border-white/20"> <div class="text-2xl sm:text-3xl font-bold text-white mb-1 sm:mb-2">115+</div> <div class="text-xs sm:text-sm text-gray-200">tokens/s 推理速度</div> <div class="text-xs text-gray-300 mt-1">RTX 4090 nf4量化</div> </div> <div class="bg-white/10 backdrop-blur-sm rounded-lg p-4 sm:p-6 border border-white/20"> <div class="text-2xl sm:text-3xl font-bold text-white mb-1 sm:mb-2">2.4GB</div> <div class="text-xs sm:text-sm text-gray-200">显存占用</div> <div class="text-xs text-gray-300 mt-1">RWKV7-G1-2.9B</div> </div> <div class="bg-white/10 backdrop-blur-sm rounded-lg p-4 sm:p-6 border border-white/20"> <div class="text-2xl sm:text-3xl font-bold text-white mb-1 sm:mb-2">1M+</div> <div class="text-xs sm:text-sm text-gray-200">上下文长度</div> <div class="text-xs text-gray-300 mt-1">理论无限支持</div> </div> <div class="bg-white/10 backdrop-blur-sm rounded-lg p-4 sm:p-6 border border-white/20"> <div class="text-2xl sm:text-3xl font-bold text-white mb-1 sm:mb-2">O(n)</div> <div class="text-xs sm:text-sm text-gray-200">时间复杂度</div> <div class="text-xs text-gray-300 mt-1">vs Transformer O(n²)</div> </div> </div> </div> </div> </div> </section>
✨步子哥 (steper) #2
02-13 04:53
ollama run mollysama/rwkv-7-g1d:2.9b