Ilker ManapandClaude Opus 5.5
ac7f5063b6
The usage and organization pages drew their choices as filled green chips, the newer pages as outlined pills. They are all the outlined pill now. The NVLink page's GPU squares were styled as .chip, the class those choices already used, which shrank the usage page's filters into 1.7rem squares wherever the stylesheet loaded. They are .gpu-chip. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
55 lines
2.8 KiB
HTML
55 lines
2.8 KiB
HTML
{{define "nvlink"}}<!doctype html>
|
|
<html lang="en"><head>
|
|
<meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
|
|
<meta http-equiv="refresh" content="30">
|
|
<title>NVLink domains · llmgw</title><link rel="stylesheet" href="{{stylesheet}}">
|
|
</head><body><div class="wrap wide">
|
|
<header>
|
|
<h1>NVLink domains</h1>
|
|
<p class="sub">{{len .Domains}} multi-node domain{{if ne (len .Domains) 1}}s{{end}} — racks whose GPUs reach each other over NVLink across nodes</p>
|
|
{{if .OnAdminListener}}<nav><a href="/">cluster</a></nav>{{else}}{{template "nav" .Nav}}{{end}}
|
|
</header>
|
|
|
|
{{if not .Domains}}
|
|
<p class="empty">No GPU reports a multi-node NVLink domain. An eight-GPU machine's NVSwitch ends at its
|
|
chassis and is shown on its node's page; a domain appears here once the fabric manager has joined the
|
|
GPUs of an NVL72 rack, or an NVLink Switch system, to one.</p>
|
|
{{end}}
|
|
|
|
<div class="racks">
|
|
{{range .Domains}}
|
|
<section class="rack{{if .Problems}} troubled{{end}}">
|
|
<header class="rack-head">
|
|
<div><h2>{{.Short}}</h2><p class="note"><code>{{.ClusterUUID}}</code> · clique {{.CliqueID}}</p></div>
|
|
<div class="rack-stats">
|
|
<div><b>{{.Usable}}/{{len .Members}}</b><span>trays usable</span></div>
|
|
<div><b>{{.Joined}}/{{.GPUs}}</b><span>GPUs joined</span></div>
|
|
</div>
|
|
</header>
|
|
<div class="rack-body">
|
|
<div class="nvswitch"><span>NVLink Switch</span><i></i></div>
|
|
<div class="rack-trays">
|
|
<div class="backplane" aria-hidden="true"></div>
|
|
{{range .Members}}
|
|
<div class="tray{{if not .Usable}} out{{end}}">
|
|
<a class="tray-name" href="{{$.Nodes}}/{{.Node}}">{{.Node}}</a>
|
|
<div class="chips">{{range .GPUs}}<span class="gpu-chip{{if .Joined}} joined{{else}} lost{{end}}{{if and .Bandwidth (ne .Bandwidth "Full")}} slow{{end}}"
|
|
title="GPU {{.Index}} · {{.Name}} · {{or .State "no fabric"}}{{if .Bandwidth}} · {{.Bandwidth}} bandwidth{{end}}">{{.Index}}</span>{{end}}</div>
|
|
<span class="imex{{if and (eq .IMEX "active") .IMEXChannels}} up{{end}}" title="nvidia-imex {{or .IMEX "not installed"}}, {{.IMEXChannels}} channel(s)">IMEX</span>
|
|
</div>
|
|
{{end}}
|
|
</div>
|
|
</div>
|
|
{{if .Problems}}<ul class="fabric-problems rack-problems">{{range .Problems}}
|
|
<li class="{{.Level}}"><span class="tag{{if eq (printf "%s" .Level) "fail"}} bad{{end}}">{{.Level}}</span> {{.Message}}</li>{{end}}</ul>{{end}}
|
|
</section>
|
|
{{end}}
|
|
</div>
|
|
|
|
<p class="note">A replica that spans nodes is placed inside one of these first: tensor parallelism between its
|
|
nodes then runs over NVLink rather than InfiniBand. Only trays with every GPU joined and IMEX up are used —
|
|
a tray short of either would leave the replica a card, or its memory, short.</p>
|
|
|
|
<footer><span>llmgw {{.Version}} · refreshes every 30s</span><span>{{.Now}}</span></footer>
|
|
</div></body></html>{{end}}
|