Donato Capitella
65 lines
3.0 KiB
HTML
65 lines
3.0 KiB
HTML
<!doctype html>
|
|
<html lang="en">
|
|
|
|
<head>
|
|
<meta charset="utf-8">
|
|
<meta name="viewport" content="width=device-width, initial-scale=1">
|
|
<title>AMD Strix Halo — MTP Benchmark Results</title>
|
|
<link rel="stylesheet" href="assets/index2.css">
|
|
<link rel="stylesheet" href="assets/mtp.css?v=2">
|
|
<script defer data-domain="kyuz0.github.io/amd-strix-halo-toolboxes" src="https://plausible.skybound.link/js/plausible.js"></script>
|
|
</head>
|
|
|
|
<body>
|
|
<header>
|
|
<div class="mtp-layout-inner">
|
|
<h1>AMD Ryzen AI MAX+ 395 “Strix Halo” — MTP Benchmarks</h1>
|
|
<p>Framework Desktop · AMD Ryzen AI MAX 395+ · 128GB unified RAM</p>
|
|
<p class="description">
|
|
Multi-Token Prediction (MTP) is an experimental speculative decoding feature for `llama.cpp`
|
|
(see <a href="https://github.com/ggml-org/llama.cpp/pull/22673" target="_blank" rel="noreferrer">PR #22673</a>).
|
|
It allows supported models to predict multiple tokens per forward pass, significantly increasing generation speed.
|
|
These benchmarks compare the baseline generation speed against MTP with 2-token and 3-token drafts.
|
|
</p>
|
|
</div>
|
|
</header>
|
|
|
|
<section class="panel compact">
|
|
<div class="panel-split mtp-layout-inner">
|
|
<div class="stats-box" style="margin-left: 0;">
|
|
<div class="stat-line" id="stats-line">Loading results...</div>
|
|
</div>
|
|
<div class="actions" style="margin-left: auto;">
|
|
<a href="index.html" class="chip small" style="text-decoration: none;">← Back to Main Benchmarks</a>
|
|
</div>
|
|
</div>
|
|
</section>
|
|
|
|
<section class="panel compact" id="tables-panel" style="border-bottom: none; background: transparent;">
|
|
<div class="table-wrap mtp-layout-inner" style="margin-top: 16px; margin-bottom: 32px; background: var(--card);">
|
|
<div class="table-scroll">
|
|
<table id="mtp-table" class="mtp-table hidden">
|
|
<thead>
|
|
<tr>
|
|
<th class="model">Model</th>
|
|
<th>Toolbox</th>
|
|
<th class="metric-col">Baseline<br><span class="sub">tok/s</span></th>
|
|
<th class="metric-col">MTP-2<br><span class="sub">tok/s</span></th>
|
|
<th class="metric-col">Speedup<br><span class="sub">MTP-2</span></th>
|
|
<th class="metric-col">MTP-3<br><span class="sub">tok/s</span></th>
|
|
<th class="metric-col">Speedup<br><span class="sub">MTP-3</span></th>
|
|
</tr>
|
|
</thead>
|
|
<tbody id="mtp-tbody">
|
|
<!-- Rows populated by JS -->
|
|
</tbody>
|
|
</table>
|
|
</div>
|
|
</div>
|
|
</section>
|
|
|
|
<script src="assets/mtp.js" type="module"></script>
|
|
</body>
|
|
|
|
</html>
|