RWKV-7 "Goose" 截至 2026 年初 RWKV 模型性能总结

✨步子哥 (steper) • 2026年02月12日 14:06

                        <!DOCTYPE html>
<html lang="zh-CN">
<head>
    <meta charset="UTF-8">
    <meta name="viewport" content="width=device-width, initial-scale=1.0">
    <title>RWKV 模型性能总结</title>
    <link href="https://fonts.googleapis.com/css2?family=Inter:wght@300;400;600;700&family=JetBrains+Mono:wght@400;700&display=swap" rel="stylesheet">
    <link href="https://fonts.googleapis.com/icon?family=Material+Icons" rel="stylesheet">
    <style>
        :root {
            --bg-color: #0f1115;
            --card-bg: #1a1d24;
            --primary: #00e5ff; /* Cyan */
            --secondary: #ff9100; /* Orange */
            --text-main: #ffffff;
            --text-muted: #b0b3b8;
            --accent-green: #00e676;
            --border-radius: 12px;
        }

        * {
            box-sizing: border-box;
            margin: 0;
            padding: 0;
        }

        body {
            background-color: var(--bg-color);
            font-family: 'Inter', sans-serif;
            color: var(--text-main);
            display: flex;
            justify-content: center;
            align-items: center;
            min-height: 100vh;
        }

        .poster-container {
            width: 720px;
            min-height: 1280px;
            background: linear-gradient(135deg, #0f1115 0%, #1a1f2e 100%);
            padding: 40px;
            display: flex;
            flex-direction: column;
            gap: 24px;
            position: relative;
            overflow: hidden;
        }

        /* Background Decoration */
        .bg-decoration {
            position: absolute;
            top: -100px;
            right: -100px;
            width: 400px;
            height: 400px;
            background: radial-gradient(circle, rgba(0, 229, 255, 0.1) 0%, rgba(0, 0, 0, 0) 70%);
            border-radius: 50%;
            z-index: 0;
        }
        
        .bg-decoration-2 {
            position: absolute;
            bottom: 0;
            left: 0;
            width: 100%;
            height: 300px;
            background: linear-gradient(to top, rgba(255, 145, 0, 0.05), transparent);
            z-index: 0;
        }

        header {
            z-index: 1;
            border-bottom: 2px solid rgba(255, 255, 255, 0.1);
            padding-bottom: 20px;
        }

        h1 {
            font-family: 'JetBrains Mono', monospace;
            font-size: 48px;
            font-weight: 700;
            color: var(--primary);
            line-height: 1.1;
            margin-bottom: 8px;
            text-shadow: 0 0 10px rgba(0, 229, 255, 0.3);
        }

        .subtitle {
            font-size: 18px;
            color: var(--text-muted);
            display: flex;
            align-items: center;
            gap: 8px;
        }

        .tag {
            background: rgba(0, 229, 255, 0.15);
            color: var(--primary);
            padding: 4px 8px;
            border-radius: 4px;
            font-size: 12px;
            font-weight: 600;
            font-family: 'JetBrains Mono', monospace;
        }

        .card {
            background: var(--card-bg);
            border-radius: var(--border-radius);
            padding: 20px;
            border: 1px solid rgba(255, 255, 255, 0.05);
            box-shadow: 0 4px 20px rgba(0, 0, 0, 0.2);
            z-index: 1;
            position: relative;
        }

        .section-title {
            font-size: 20px;
            font-weight: 700;
            margin-bottom: 16px;
            display: flex;
            align-items: center;
            gap: 8px;
            color: var(--text-main);
        }
        
        .section-title i {
            color: var(--secondary);
        }

        /* Core Advantages */
        .advantages-grid {
            display: grid;
            grid-template-columns: repeat(2, 1fr);
            gap: 12px;
        }

        .adv-item {
            background: rgba(255, 255, 255, 0.03);
            padding: 12px;
            border-radius: 8px;
            border-left: 3px solid var(--primary);
        }

        .adv-item h3 {
            font-size: 14px;
            color: var(--primary);
            margin-bottom: 4px;
            font-weight: 600;
        }

        .adv-item p {
            font-size: 12px;
            color: var(--text-muted);
            line-height: 1.4;
        }

        /* Benchmark Table */
        .table-container {
            width: 100%;
            overflow: hidden;
        }
        
        table {
            width: 100%;
            border-collapse: collapse;
            font-size: 12px;
        }

        th {
            text-align: left;
            color: var(--text-muted);
            padding: 8px 6px;
            border-bottom: 1px solid rgba(255, 255, 255, 0.1);
            font-weight: 600;
            white-space: nowrap;
        }
        
        td {
            padding: 8px 6px;
            border-bottom: 1px solid rgba(255, 255, 255, 0.05);
            font-family: 'JetBrains Mono', monospace;
            color: var(--text-main);
            white-space: nowrap;
        }

        tr:last-child td {
            border-bottom: none;
        }

        .highlight-text {
            color: var(--accent-green);
            font-weight: bold;
        }

        .model-name {
            color: var(--secondary);
            font-weight: 700;
        }

        /* Inference Stats */
        .stats-grid {
            display: grid;
            grid-template-columns: repeat(2, 1fr);
            gap: 16px;
        }

        .stat-box {
            background: rgba(0, 0, 0, 0.2);
            padding: 16px;
            border-radius: 8px;
            text-align: center;
        }

        .stat-value {
            font-size: 28px;
            font-weight: 700;
            color: var(--primary);
            font-family: 'JetBrains Mono', monospace;
            margin-bottom: 4px;
        }

        .stat-label {
            font-size: 12px;
            color: var(--text-muted);
            text-transform: uppercase;
            letter-spacing: 1px;
        }
        
        .stat-sub {
            font-size: 11px;
            color: var(--text-muted);
            margin-top: 4px;
        }

        .comparison-badge {
            display: inline-block;
            background: rgba(255, 145, 0, 0.15);
            color: var(--secondary);
            padding: 2px 6px;
            border-radius: 4px;
            font-size: 10px;
            margin-top: 4px;
        }

        /* Usage Section */
        .tips-list {
            display: flex;
            flex-direction: column;
            gap: 12px;
        }

        .tip-item {
            display: flex;
            align-items: flex-start;
            gap: 10px;
            font-size: 13px;
            color: var(--text-muted);
            line-height: 1.5;
        }

        .tip-item i {
            color: var(--primary);
            font-size: 16px;
            margin-top: 2px;
        }

        .tip-strong {
            color: var(--text-main);
            font-weight: 600;
        }

        footer {
            margin-top: auto;
            text-align: center;
            padding-top: 16px;
            border-top: 1px solid rgba(255, 255, 255, 0.05);
            z-index: 1;
        }

        .links-grid {
            display: flex;
            justify-content: center;
            gap: 24px;
            flex-wrap: wrap;
        }
        
        .link-item {
            font-size: 12px;
            color: var(--text-muted);
            display: flex;
            align-items: center;
            gap: 4px;
        }
        
        .link-item i {
            font-size: 14px;
            color: var(--secondary);
        }

    </style>
</head>
<body>
    <div class="poster-container">
        <div class="bg-decoration"></div>
        <div class="bg-decoration-2"></div>

        <header>
            <div class="subtitle">
                <span class="tag">RWKV-7 "Goose"</span>
                <span>截至 2026 年初</span>
            </div>
            <h1>RWKV 模型性能总结</h1>
            <div style="margin-top: 8px; font-size: 14px; color: rgba(255,255,255,0.7);">
                纯 RNN 架构 · 无注意力机制 · 线性推理
            </div>
        </header>

        <!-- Core Advantages -->
        <div class="card">
            <div class="section-title">
                <i class="material-icons">bolt</i>
                核心优势
            </div>
            <div class="advantages-grid">
                <div class="adv-item">
                    <h3>推理时间线性</h3>
                    <p>推理耗时与序列长度无关，无二次复杂度瓶颈。</p>
                </div>
                <div class="adv-item">
                    <h3>显存恒定</h3>
                    <p>无 KV Cache，显存占用极低，支持无限上下文。</p>
                </div>
                <div class="adv-item">
                    <h3>训练可并行</h3>
                    <p>像 Transformer 一样高效训练，打破 RNN 串行限制。</p>
                </div>
                <div class="adv-item">
                    <h3>极致高效</h3>
                    <p>手机/集显即可实时运行，功耗优势巨大。</p>
                </div>
            </div>
        </div>

        <!-- Benchmark -->
        <div class="card">
            <div class="section-title">
                <i class="material-icons">leaderboard</i>
                RWKV-7 基准测试表现
            </div>
            <div class="table-container">
                <table>
                    <thead>
                        <tr>
                            <th>模型规模</th>
                            <th>MMLU</th>
                            <th>GSM8K</th>
                            <th>MATH</th>
                            <th>IFEval</th>
                            <th style="color: var(--accent-green)">Uncheatable</th>
                        </tr>
                    </thead>
                    <tbody>
                        <tr>
                            <td class="model-name">13.3B (G0b)</td>
                            <td>76.5%</td>
                            <td>92.3%</td>
                            <td>76.8%</td>
                            <td>68.9%</td>
                            <td class="highlight-text">6.843 (Best)</td>
                        </tr>
                        <tr>
                            <td class="model-name">7.2B (G0a3)</td>
                            <td>65.1%</td>
                            <td>83.9%</td>
                            <td>67.8%</td>
                            <td>58.0%</td>
                            <td>7.222</td>
                        </tr>
                        <tr>
                            <td class="model-name">2.9B (G1a4)</td>
                            <td>61.3%</td>
                            <td>77.3%</td>
                            <td>48.2%</td>
                            <td>51.0%</td>
                            <td>7.486</td>
                        </tr>
                        <tr>
                            <td class="model-name">1.5B (G1b)</td>
                            <td>50.5%</td>
                            <td>58.5%</td>
                            <td>29.8%</td>
                            <td>42.1%</td>
                            <td>7.969</td>
                        </tr>
                    </tbody>
                </table>
            </div>
            <div style="margin-top: 10px; font-size: 11px; color: var(--text-muted);">
                * Uncheatable Eval 数值越低越好；13.3B 优于 Qwen3-14B。
            </div>
        </div>

        <!-- Inference Performance -->
        <div class="card">
            <div class="section-title">
                <i class="material-icons">speed</i>
                推理性能实测 (RWKV-7 2.9B)
            </div>
            <div class="stats-grid">
                <div class="stat-box">
                    <div class="stat-value">115<span style="font-size:14px"> t/s</span></div>
                    <div class="stat-label">RTX 4090</div>
                    <div class="stat-sub">nf4 量化 · 2.4 GB VRAM</div>
                </div>
                <div class="stat-box">
                    <div class="stat-value">86<span style="font-size:14px"> t/s</span></div>
                    <div class="stat-label">RTX 4060 Laptop</div>
                    <div class="stat-sub">nf4 量化 · 2.4 GB VRAM</div>
                </div>
                <div class="stat-box">
                    <div class="stat-value">30<span style="font-size:14px">+ t/s</span></div>
                    <div class="stat-label">手机 S8 Gen 3</div>
                    <div class="stat-sub">A16W4 量化 · 边缘可用</div>
                </div>
                <div class="stat-box">
                    <div class="stat-value">6.5<span style="font-size:14px"> t/s</span></div>
                    <div class="stat-label">RK3588 NPU</div>
                    <div class="stat-sub">W8A8 · 嵌入式设备</div>
                </div>
            </div>
            <div style="margin-top: 12px; text-align: center;">
                <span class="comparison-badge">对比 Transformer: 速度快 3-10 倍 · 显存仅需 1/3</span>
            </div>
        </div>

        <!-- Usage Tips -->
        <div class="card" style="flex: 1;">
            <div class="section-title">
                <i class="material-icons">lightbulb</i>
                使用建议与资源
            </div>
            <div class="tips-list">
                <div class="tip-item">
                    <i class="material-icons">stars</i>
                    <div>
                        <span class="tip-strong">追求最强性能:</span> 选择 RWKV-7 13.3B / 7.2B，性能接近/超过主流 Transformer。
                    </div>
                </div>
                <div class="tip-item">
                    <i class="material-icons">smartphone</i>
                    <div>
                        <span class="tip-strong">手机/笔记本部署:</span> 推荐 2.9B G1 系列 (GGUF)，普通硬件即可流畅运行。
                    </div>
                </div>
                <div class="tip-item">
                    <i class="material-icons">public</i>
                    <div>
                        <span class="tip-strong">多语言任务:</span> 优选 World 系列，多语言基准达 SOTA 水平。
                    </div>
                </div>
                <div class="tip-item">
                    <i class="material-icons">code</i>
                    <div>
                        <span class="tip-strong">推荐后端:</span> 
                        <span style="color:var(--primary)">web-rwkv</span> (最快), 
                        <span style="color:var(--primary)">llama.cpp</span> (通用)。
                    </div>
                </div>
            </div>
        </div>

        <footer>
            <div class="links-grid">
                <div class="link-item"><i class="material-icons">language</i> www.rwkv.com</div>
                <div class="link-item"><i class="material-icons">menu_book</i> wiki.rwkv.com</div>
                <div class="link-item"><i class="material-icons">cloud_download</i> Hugging Face: BlinkDL/rwkv7-g1</div>
            </div>
        </footer>
    </div>
</body>
</html>                    

讨论回复

1 条回复

✨步子哥 (steper) #1

02-12 14:27

                                        <!DOCTYPE html>
<html lang="zh-CN">
<head>
    <meta charset="UTF-8">
    <meta name="viewport" content="width=device-width, initial-scale=1.0">
    <title>Web-RWKV - 纯WebGPU推理引擎</title>
    <link href="https://fonts.googleapis.com/css2?family=Inter:wght@300;400;600;700&family=JetBrains+Mono:wght@400;600;700&display=swap" rel="stylesheet">
    <link href="https://fonts.googleapis.com/icon?family=Material+Icons" rel="stylesheet">
    <style>
        :root {
            --bg-color: #1e1e1e; /* VS Code Dark+ */
            --card-bg: #252526;
            --code-bg: #1e1e1e;
            --primary: #569cd6; /* Blue keyword */
            --secondary: #ce9178; /* Orange string */
            --accent: #4ec9b0; /* Cyan type */
            --text-main: #d4d4d4;
            --text-muted: #858585;
            --success: #6a9955; /* Green comment */
            --warning: #dcdcaa; /* Yellow function */
            --purple: #c586c0; /* Purple control */
            --border-radius: 8px;
        }

        * {
            box-sizing: border-box;
            margin: 0;
            padding: 0;
        }

        body {
            background-color: #121212;
            font-family: 'Inter', sans-serif;
            color: var(--text-main);
            display: flex;
            justify-content: center;
            align-items: center;
            min-height: 100vh;
        }

        .poster-container {
            width: 720px;
            min-height: 1280px;
            background: var(--bg-color);
            padding: 32px;
            display: flex;
            flex-direction: column;
            gap: 24px;
            position: relative;
            overflow: hidden;
            border: 1px solid #333;
        }

        /* Editor Header */
        .editor-header {
            display: flex;
            align-items: center;
            gap: 12px;
            padding-bottom: 16px;
            border-bottom: 1px solid #333;
        }

        .window-controls {
            display: flex;
            gap: 8px;
        }

        .window-btn {
            width: 12px;
            height: 12px;
            border-radius: 50%;
            background: #333;
        }
        .window-btn.close { background: #f5475b; }
        .window-btn.min { background: #fbd84b; }
        .window-btn.max { background: #37d283; }

        .file-path {
            font-family: 'JetBrains Mono', monospace;
            font-size: 13px;
            color: var(--text-muted);
        }

        header {
            border-bottom: 1px solid #333;
            padding-bottom: 16px;
        }

        h1 {
            font-family: 'JetBrains Mono', monospace;
            font-size: 42px;
            font-weight: 700;
            color: var(--warning); /* Function color */
            margin-bottom: 12px;
        }

        .subtitle {
            font-size: 16px;
            color: var(--accent); /* Type color */
            margin-bottom: 8px;
        }

        .badges {
            display: flex;
            gap: 8px;
            flex-wrap: wrap;
        }

        .badge {
            background: rgba(86, 156, 214, 0.15);
            color: var(--primary);
            padding: 4px 8px;
            border-radius: 4px;
            font-size: 12px;
            font-family: 'JetBrains Mono', monospace;
            border: 1px solid rgba(86, 156, 214, 0.3);
        }

        /* Code Block Style */
        .code-block {
            background: var(--code-bg);
            border: 1px solid #333;
            border-radius: var(--border-radius);
            padding: 16px;
            font-family: 'JetBrains Mono', monospace;
            font-size: 13px;
            position: relative;
        }

        .code-title {
            position: absolute;
            top: 0;
            right: 0;
            background: #333;
            padding: 2px 8px;
            border-radius: 0 0 0 4px;
            font-size: 10px;
            color: var(--text-muted);
        }

        .section-title {
            font-size: 18px;
            font-weight: 600;
            color: var(--text-main);
            margin-bottom: 12px;
            display: flex;
            align-items: center;
            gap: 8px;
            font-family: 'JetBrains Mono', monospace;
        }

        .section-title::before {
            content: '//';
            color: var(--success);
            margin-right: 4px;
        }

        .feature-list {
            display: grid;
            grid-template-columns: repeat(2, 1fr);
            gap: 10px;
        }

        .feature-item {
            display: flex;
            align-items: flex-start;
            gap: 8px;
            font-size: 13px;
            line-height: 1.4;
        }

        .feature-item i {
            color: var(--accent);
            font-size: 16px;
            margin-top: 2px;
        }

        .keyword { color: var(--purple); }
        .function { color: var(--warning); }
        .string { color: var(--secondary); }
        .type { color: var(--accent); }
        .comment { color: var(--success); font-style: italic; }
        .operator { color: var(--text-main); }

        .provides-grid {
            display: grid;
            grid-template-columns: 1fr 1fr;
            gap: 12px;
        }

        .provide-card {
            background: rgba(255, 255, 255, 0.02);
            padding: 12px;
            border-radius: 6px;
            border-left: 3px solid var(--success);
        }
        
        .not-provide-card {
            background: rgba(255, 255, 255, 0.02);
            padding: 12px;
            border-radius: 6px;
            border-left: 3px solid #f5475b;
        }

        .card-header {
            font-weight: 600;
            font-size: 14px;
            margin-bottom: 6px;
            color: var(--text-main);
        }

        .card-content {
            font-size: 12px;
            color: var(--text-muted);
        }

        .example-command {
            padding: 8px 12px;
            background: rgba(255, 255, 255, 0.02);
            border-radius: 4px;
            margin-bottom: 6px;
            border-left: 2px solid var(--primary);
            font-family: 'JetBrains Mono', monospace;
            font-size: 12px;
            color: var(--text-main);
        }

        .hook-diagram {
            display: flex;
            gap: 8px;
            align-items: center;
            margin-top: 10px;
            font-size: 12px;
        }

        .hook-box {
            background: rgba(86, 156, 214, 0.1);
            border: 1px dashed var(--primary);
            padding: 6px;
            border-radius: 4px;
            text-align: center;
            font-size: 11px;
            color: var(--primary);
        }

        footer {
            margin-top: auto;
            padding-top: 16px;
            border-top: 1px solid #333;
            display: flex;
            justify-content: space-between;
            align-items: center;
            font-size: 12px;
            color: var(--text-muted);
        }

        .links {
            display: flex;
            gap: 16px;
        }

        .link {
            color: var(--accent);
            text-decoration: none;
            display: flex;
            align-items: center;
            gap: 4px;
        }

    </style>
</head>
<body>
    <div class="poster-container">
        <!-- Editor Header Decoration -->
        <div class="editor-header">
            <div class="window-controls">
                <div class="window-btn close"></div>
                <div class="window-btn min"></div>
                <div class="window-btn max"></div>
            </div>
            <div class="file-path">~/projects/web-rwkv/README.md</div>
        </div>

        <header>
            <h1>Web-RWKV</h1>
            <div class="subtitle">Inference engine for RWKV implemented in pure WebGPU</div>
            <div class="badges">
                <span class="badge">v0.10</span>
                <span class="badge">Rust</span>
                <span class="badge">WebGPU</span>
                <span class="badge">WASM</span>
                <span class="badge">Cross-Platform</span>
            </div>
        </header>

        <!-- Core Features -->
        <div class="code-block">
            <div class="section-title">Core Features</div>
            <div class="feature-list">
                <div class="feature-item"><i class="material-icons">check_circle</i>No CUDA/Python dependencies</div>
                <div class="feature-item"><i class="material-icons">check_circle</i>Support Nvidia/AMD/Intel GPUs</div>
                <div class="feature-item"><i class="material-icons">check_circle</i>Vulkan/Dx12/OpenGL backends</div>
                <div class="feature-item"><i class="material-icons">check_circle</i>WASM support (Browser ready)</div>
                <div class="feature-item"><i class="material-icons">check_circle</i>Batched inference</div>
                <div class="feature-item"><i class="material-icons">check_circle</i>Int8 and Float4 quantization</div>
                <div class="feature-item"><i class="material-icons">check_circle</i>Support RWKV V4 through V7</div>
                <div class="feature-item"><i class="material-icons">check_circle</i>LoRA merging & Model serialization</div>
            </div>
        </div>

        <!-- Scope -->
        <div class="code-block">
            <div class="section-title">Functional Scope</div>
            <div class="provides-grid">
                <div class="provide-card">
                    <div class="card-header">✅ Provides</div>
                    <div class="card-content">
                        • Tokenizer<br>
                        • Model Loading<br>
                        • State Creation & Updating<br>
                        • GPU-accelerated `run` & `softmax`<br>
                        • Model Quantization
                    </div>
                </div>
                <div class="not-provide-card">
                    <div class="card-header">❌ Does Not Provide</div>
                    <div class="card-content">
                        • OpenAI-compatible API<br>
                        • Built-in Samplers<br>
                        • State Caching System<br>
                        • Python Bindings
                    </div>
                </div>
            </div>
        </div>

        <!-- Usage Examples -->
        <div class="code-block">
            <div class="section-title">Usage Examples</div>
            
            <div class="example-command">
                <span class="comment"># Performance Test (500 tokens)</span><br>
                cargo run --release --example gen
            </div>
            
            <div class="example-command">
                <span class="comment"># Chat Demo</span><br>
                cargo run --release --example chat -- --model /path/to/model.st
            </div>
            
            <div class="example-command">
                <span class="comment"># Quantization Example (First 32 layers)</span><br>
                cargo run --release --example chat -- --quant 32
            </div>
        </div>

        <!-- Advanced Features -->
        <div class="code-block">
            <div class="section-title">Advanced Features</div>
            
            <div style="margin-bottom: 10px;">
                <span class="keyword">let</span> <span class="function">runtime</span> = TokioRuntime::new(bundle).await; <span class="comment">// Async runtime</span>
            </div>
            <div style="margin-bottom: 10px; font-size: 12px; color: var(--text-muted);">
                The asynchronous runtime API allows CPU and GPU to work in parallel, maximizing utilization.
            </div>

            <div class="hook-diagram">
                <div class="hook-box">Input Tokens</div>
                <span class="operator">→</span>
                <div class="hook-box" style="border-style: solid; border-color: var(--warning); color: var(--warning);">Hook Point</div>
                <span class="operator">→</span>
                <div class="hook-box">Output Logits</div>
            </div>
            <div style="margin-top: 8px; font-size: 12px;">
                <span class="type">Hooks</span>: Inject tensor ops into inference process for dynamic LoRA, control net, etc.
            </div>
        </div>

        <!-- Convert & Troubleshoot -->
        <div class="code-block">
            <div class="section-title">Model Conversion</div>
            <div class="example-command">
                python assets/scripts/convert_safetensors.py <span class="operator">--input</span> model.pth <span class="operator">--output</span> model.st
            </div>
        </div>

        <footer>
            <div>© 2024 Web-RWKV Project</div>
            <div class="links">
                <span class="link"><i class="material-icons" style="font-size:14px">code</i> GitHub</span>
                <span class="link"><i class="material-icons" style="font-size:14px">description</i> docs.rs/web-rwkv</span>
            </div>
        </footer>
    </div>
</body>
</html>                                    

需要登录才能发表回复

登录注册