Files
llmgw/internal/web/templates/dashboard.html
T
Ilker ManapandClaude Opus 5.5 7d97c6b74f Install a node's engines on the node, instead of trusting what is there
A node ran whatever llama-server somebody had once unpacked on it. Onboarding
checked an engine only when told its path, installed none, and a node with
nothing on it registered as ready. The one time that went wrong it went wrong
invisibly: the binary was there, executable, and could not start for want of
libgomp1, while the node reported healthy.

The control plane now says which engines every node has, at which version —
`engines:` in server.yaml, defaulting to the llama.cpp build this release was
checked against — and each agent converges on it. What only the node can
know, the node works out: the build for its processor and driver (CUDA with
its runtime when an NVIDIA driver is loaded, CPU otherwise), the OS packages
that build needs from its own package sources, and whether the result runs.
Every archive is checked against the sha256 GitHub publishes for it and
unpacked refusing any path or link that leaves the install directory. ldd
names a missing library outright; running `--version` proves the rest. Only
then is a directory marked installed, and it is checked again every ten
minutes, because an unrelated upgrade can remove a library.

An engine is reported only once installed and seen to run, or when the
agent's configuration names it on purpose — never because something happened
to be on PATH. The agent now says when it has probed, so an empty list means
"has none" rather than "too old to say": a node rebuilt from nothing is no
longer a candidate for every model at the moment it can run none of them.
Placement's refusal carries the node's own words, "installing: downloading
llama-b10991-bin-ubuntu-x64.tar.gz (16 MB)", and so do `ctl nodes` and the
cluster page, along with a warning for a CPU without AVX2 — the VM CPU type
that makes everything five times slower and nothing look wrong.

Agents also report their OS, architecture, vector extensions and NVIDIA
driver. Onboarding installs llmgw itself from the CI release for the node's
architecture, checked against the release's checksums, rather than copying
whatever binary the operator's machine had.

vLLM and SGLang are not installed yet: they are Python environments of
several gigabytes, not an archive, and still have to be named in the agent's
configuration.

Also fixes a race on the weights fingerprint, which two heartbeat loops touch
during a failover.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-09-28 10:59:59 +02:00

256 lines
11 KiB
HTML
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
{{define "dashboard"}}<!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="{{.RefreshSec}}">
<title>llmgw — cluster</title><link rel="stylesheet" href="/_static/base.css">
</head><body><div class="wrap">
<header>
<h1>Cluster</h1>
<p class="sub">{{.NodeCount}} nodes · {{.ReadyReplicas}} of {{.TotalReplicas}} replicas ready · {{.ModelCount}} models</p>
<nav>
{{if .OnAdminListener}}
<a href="/metrics">metrics</a>
<a href="/debug/pprof/">pprof</a>
<a href="/api/v1/nodes">api</a>
{{else}}
<a href="/admin/users">accounts</a>
<a href="/admin/usage">usage</a>
{{if .Nav.Mirror}} <a href="/admin/mirror">mirror</a>
{{end}}
<a href="/profile">your keys</a>
<a href="/chat">chat</a>
{{end}}
</nav>
</header>
{{if .HasCharts}}
<div class="tiles">
<div class="tile"><div class="v">{{.Requests24h}}</div><div class="k">requests · 24h</div>{{.RequestSpark}}</div>
<div class="tile"><div class="v">{{compact .Tokens24h}}</div><div class="k">tokens · 24h</div>{{.TokenSpark}}</div>
<div class="tile"><div class="v">{{.Errors24h}}</div><div class="k">errors · 24h</div></div>
<div class="tile"><div class="v">{{units .Spend24h}}</div><div class="k">spend · 24h</div></div>
</div>
{{end}}
<h2>Capacity</h2>
{{if .Summary.GPUs}}
<table><tr><th>Class</th><th class="num">In use</th><th class="num">Free</th><th class="num">Faulty</th><th>Usage</th></tr>
{{range .Summary.GPUs}}<tr>
<td><code>{{.Class}}</code></td>
<td class="num">{{.Used}}/{{.Total}}</td>
<td class="num">{{.Free}}</td>
<td class="num">{{if .Unhealthy}}<span class="bad">{{.Unhealthy}}</span>{{else}}—{{end}}</td>
<td class="meter"><span style="width:{{pct .Used .Total}}%"></span><em>{{pct .Used .Total}}%</em></td>
</tr>{{end}}</table>
{{else}}<p class="empty">No GPUs have been reported.</p>{{end}}
{{if .Summary.Fabrics}}
<h2>Fabrics</h2>
<table><tr><th>Fabric</th><th class="num">Nodes</th><th class="num">Fully free</th><th>Weights</th></tr>
{{range .Summary.Fabrics}}<tr>
<td><code>{{.Fabric}}</code></td>
<td class="num">{{.Nodes}}</td>
<td class="num">{{.LargestFreeBlock}}</td>
<td>{{if .NodesWithWeights}}{{.Models}} model(s) on {{.NodesWithWeights}}/{{.Nodes}} nodes{{else}}<span class="note">not reported</span>{{end}}</td>
</tr>{{end}}</table>
<p class="note">Fully free counts whole unused nodes, which is what decides whether a
replica spanning nodes can land: free cards on the wrong fabric are not capacity
for a model that needs this one.</p>
{{end}}
{{if .Summary.IdleReplicas}}
<h2>Holding hardware nothing is asking for</h2>
<table><tr><th>Replica</th><th>Model</th><th class="num">GPUs</th><th class="num">Unused for</th></tr>
{{range .Summary.IdleReplicas}}<tr>
<td><code>{{.ID}}</code></td>
<td>{{.Model}}</td>
<td class="num">{{.GPUs}}</td>
<td class="num">{{duration .IdleSec}}</td>
</tr>{{end}}</table>
{{end}}
{{if .HasPeople}}
<h2>People</h2>
{{.UserChart}}
<table><tr><th>Person</th><th class="num">Keys</th><th class="num">Requests</th><th class="num">Tokens</th><th class="num">Errors</th><th class="num">Spend</th><th>Last seen</th></tr>
{{range .Users}}<tr>
<td>{{if .Email}}{{.Email}}{{else}}<span class="note">operator-issued keys</span>{{end}}</td>
<td class="num">{{.Keys}}</td>
<td class="num">{{.Requests}}</td>
<td class="num">{{compact .Tokens}}</td>
<td class="num">{{if .Errors}}<span class="bad">{{.Errors}}</span>{{else}}—{{end}}</td>
<td class="num">{{units .CostMicros}}</td>
<td>{{ago .LastSeen}}</td>
</tr>{{end}}</table>
{{end}}
{{if .Models2}}
<h2>What is being run</h2>
{{.ModelChart}}
<table><tr><th>Model</th><th class="num">Requests</th><th class="num">Tokens</th><th class="num">Errors</th><th class="num">Mean latency</th></tr>
{{range .Models2}}<tr>
<td><code>{{.Model}}</code></td>
<td class="num">{{.Requests}}</td>
<td class="num">{{compact .Tokens}}</td>
<td class="num">{{if .Errors}}<span class="bad">{{.Errors}}</span>{{else}}—{{end}}</td>
<td class="num">{{printf "%.0f" .AvgLatency}} ms</td>
</tr>{{end}}</table>
<p class="note">Counted by the model that actually answered, which differs from the one
requested whenever a fallback served it.</p>
{{end}}
<h2>Nodes</h2>
{{if .Nodes}}
<table><tr><th>Node</th><th>State</th><th class="num">GPUs</th><th class="num">RAM</th><th class="num">Disk</th><th>Engines</th><th>Classes</th><th>Fabrics</th><th class="num">Models</th><th class="num">Replicas</th><th>Last seen</th></tr>
{{range .Nodes}}<tr>
<td><code>{{.ID}}</code>{{if .MgmtAddr}} <span class="note">{{.MgmtAddr}}</span>{{end}}</td>
<td class="state state-{{.State}}">{{.State}}</td>
<td class="num">{{.GPUFree}}/{{.GPUTotal}}</td>
<td class="num">{{if .MemoryGB}}{{.MemoryGB}}G{{else}}—{{end}}</td>
{{/* Free out of total, because the first is meaningless without the
second, and an em dash where nothing was reported: an agent that
was never asked about its disk must not read as a full one. */}}
<td class="num">{{if .DiskTotalGB}}{{.DiskFreeGB}}G/{{.DiskTotalGB}}G{{else}}—{{end}}</td>
<td>{{if .Engines}}{{join .Engines}}{{else}}<span class="note">none yet</span>{{end}}</td>
<td>{{join .Classes}}</td>
<td>{{join .Fabrics}}</td>
<td class="num">{{if .Models}}{{len .Models}}{{else}}—{{end}}</td>
<td class="num">{{.Replicas}}</td>
<td>{{or .LastSeenAgo "—"}}</td>
</tr>{{if .Unhealthy}}<tr><td colspan="11" class="note bad">{{join .Unhealthy}}</td></tr>{{end}}
{{range .Provision}}{{if ne .State "ready"}}<tr><td colspan="11" class="note{{if or (eq .State "failed") (eq .State "unsupported")}} bad{{end}}">{{.Engine}} {{.Version}}{{if .Variant}} ({{.Variant}}){{end}} {{.State}}: {{.Detail}}</td></tr>{{end}}{{end}}
{{range .Warnings}}<tr><td colspan="11" class="note bad">{{.}}</td></tr>{{end}}
{{end}}</table>
{{else}}<p class="empty">No agents have registered.</p>{{end}}
<h2>Models</h2>
{{with .Fleet}}
<h2>What the cluster is doing <span class="opt">last {{.Window}}</span></h2>
{{if .Err}}
{{/* Said outright. An empty panel reads as a healthy cluster with nothing
happening, which is the opposite of not knowing. */}}
<p class="note error-row">No history: {{.Err}}</p>
{{else}}
<div class="chart-row">
{{template "tschart" dict "t" .Saturation "title" "Pool saturation, by model"
"note" "The worse of queue depth and cache pressure. 1 is the limit measured as healthy, not the point of failure."}}
{{template "tschart" dict "t" .Ready "title" "Replicas ready, by model"
"note" "A model deliberately stopped has no line at all, which is why an on-demand catalogue looks quiet here."}}
</div>
<div class="chart-row">
{{template "tschart" dict "t" .Queue "title" "Engine queue depth, by replica"
"note" "Requests accepted and not yet started. The first thing to move when a replica is falling behind."}}
{{template "tschart" dict "t" .KV "title" "KV cache occupancy, by replica"
"note" "How full the attention cache is. Long contexts fill it with an empty queue, which is why saturation reads both."}}
</div>
{{end}}
{{end}}
{{if .Models}}
<table><tr><th>Model</th><th>Aliases</th><th class="num">Ready</th><th class="num">Wanted</th><th>Profiles</th><th>Fallbacks</th></tr>
{{range .Models}}<tr>
<td><code>{{.Name}}</code></td>
<td>{{join .Aliases}}</td>
<td class="num">{{.Ready}}</td>
<td class="num">{{.Min}}–{{.Max}}</td>
<td>{{join .Profiles}}</td>
<td>{{join .Fallbacks}}</td>
</tr>
{{/* Under the row rather than in a column: a sentence in a table cell either
wraps every other column into a stripe or gets cut off. */}}
{{if .Description}}<tr><td colspan="6" class="url-row">{{.Description}}</td></tr>{{end}}
{{end}}</table>
{{else}}<p class="empty">The catalog is empty.</p>{{end}}
{{if .Unplaced}}
<h2>Why nothing was placed</h2>
{{range .Unplaced}}
<p><code>{{.Model}}</code></p>
<pre>{{range .Rejections}}{{.Profile}} @ {{or .Pool "-"}}{{if .Node}}/{{.Node}}{{end}}: {{.Reason}}
{{end}}</pre>
{{end}}
{{end}}
<h2>Replicas</h2>
{{if .Replicas}}
<table><tr><th>Replica</th><th>Model</th><th>State</th><th>Pool</th><th>Endpoint</th><th class="num">Age</th><th>Nodes</th></tr>
{{range .Replicas}}<tr>
<td><code>{{.ID}}</code></td>
<td>{{.Model}}</td>
<td class="state state-{{.State}}">{{.State}}</td>
<td>{{or .Pool "—"}}</td>
<td>{{if .Endpoint}}<code>{{.Endpoint}}</code>{{else}}—{{end}}</td>
<td class="num">{{duration .AgeSec}}</td>
<td>{{range .Shards}}<code>{{.Node}}</code> {{end}}</td>
</tr>{{if .Err}}<tr><td colspan="7" class="note">{{.Err}}</td></tr>{{end}}
{{end}}</table>
{{else}}<p class="empty">No replicas.</p>{{end}}
<h2>Traffic <span class="note">— last 24 hours, hourly</span></h2>
{{if .HasCharts}}
<p class="note">Requests</p>{{.RequestChart}}
<div class="chart-row">
<div><p class="note">Tokens</p>{{.TokenChart}}</div>
<div><p class="note">Mean latency</p>{{.LatencyChart}}</div>
</div>
{{else if .AuthEnabled}}
<p class="empty">No traffic recorded yet.</p>
{{end}}
<h2>Usage <span class="note">— last 24 hours</span></h2>
{{if .AuthEnabled}}
{{if .Usage}}
<table><tr><th>Key</th><th>Project</th><th class="num">Requests</th><th class="num">Errors</th><th class="num">Prompt</th><th class="num">Completion</th><th class="num">Spend</th><th>Last</th></tr>
{{range .Usage}}<tr>
<td>{{if .KeyName}}{{.KeyName}} <span class="note">{{.KeyPrefix}}</span>{{else}}<code>{{.KeyID}}</code>{{end}}</td>
<td>{{.ProjectID}}</td>
<td class="num">{{.Requests}}</td>
<td class="num">{{if .Errors}}<span class="state-failed">{{.Errors}}</span>{{else}}0{{end}}</td>
<td class="num">{{.Prompt}}</td>
<td class="num">{{.Completion}}</td>
<td class="num">{{units .CostMicros}}</td>
<td><span class="note">{{ago .LastAt}}</span></td>
</tr>{{end}}</table>
<p class="note">Spend is in units, whatever a unit was configured to mean. Usage is
written in batches, so the last few seconds may be missing.</p>
{{else}}<p class="empty">No requests recorded in the last 24 hours.</p>{{end}}
{{else}}
<p class="empty">Authentication is off, so there is no identity to account against.</p>
{{end}}
{{if not .OnAdminListener}}
<p class="note">Metrics, pprof and the management API are on the admin listener,
which is bound to loopback and not reachable from here. pprof can dump the
process heap, where API keys sit in plaintext, which is why that port is not
exposed:
<code>ssh -L 9090:127.0.0.1:9090 &lt;control-plane&gt;</code></p>
{{end}}
<footer>
<span>llmgw {{.Version}} · refreshes every {{.RefreshSec}}s</span>
<span>{{.Now}}</span>
</footer>
</div></body></html>{{end}}
{{/* One chart: the SVG, its legend, and the sentence that says how to read it.
A legend is always present for two lines or more — colour alone is not an
encoding a reader can rely on — and a single line needs none, because the
title already names it. */}}
{{define "tschart"}}
<div class="chart-wrap">
<h3 class="chart-title">{{.title}}</h3>
{{if .t.Empty}}
<p class="empty">Nothing recorded yet.</p>
{{else}}
{{.t.SVG}}
<span class="chart-peak">peak {{.t.Peak}}</span>
{{if gt (len .t.Legend) 1}}
<div class="legend">
{{range .t.Legend}}<span><i style="background:var({{cssvar .Colour}})"></i>{{.Name}} <b>{{.Last}}</b></span>{{end}}
</div>
{{end}}
{{end}}
<p class="note">{{.note}}</p>
</div>
{{end}}