№ 04 · /api/public/status · cache 30 s

Status — live.

last poll · Z

next in · 15s

FLEET

5/5

healthy

p50

252.8

req / 24h

—

all workers

uptime

—

30d window

WorkerHost · aliasLatency p50UptimeModelStatus

Mac Studio · vllm-mlx multi-model :8500

studio

174.9 ms

0 h

—

healthy

Mac Studio · Qwen3.6-35B multi-LoRA (hardware/EDA/math) :9360

studio

244.89999999999998 ms

0 h

—

healthy

Mac Studio · Qwen3.6-35B multi-LoRA (code/web/lang) :9361

studio

251.8 ms

0 h

—

healthy

macM1 · vllm-mlx FC hot-path (granite) :8520

macm1

94.2 ms

0 h

—

healthy

KXKM-AI · gpu-swap on-demand (RTX 4090) :8005

kxkm-ai (RTX 4090, autossh tunnel)

8024.4 ms

0 h

—

healthy

auto-router

multilingual-e5-large 1024d · classifier MLP

studio.tail

:9300

14 ms

814 h

88.9% macro-F1

healthy

Probe sequence

[gateway_probe.py] tick = 30s
  studio:8500      → 200 OK · 224 ms · vllm-mlx multi-modèle (catalogue chargé)
  studio:9360      → 200 OK · 188 ms · qwen36-35B multi-LoRA (hardware/EDA/math)
  studio:9361      → 200 OK · 196 ms · qwen36-35B multi-LoRA (code/web/lang)
  macm1:8520       → 200 OK · 142 ms · vllm-mlx granite FC hot-path
  kxkm-ai:8005     → 200 OK · 257 ms · gpu-swap RTX 4090 (FC failover, SchGen, vision)
  ----
  cache age:       12 s
  next refresh:    30 s

Incidents · 30 derniers jours

06 maikxkm-aiautossh restart · 4 min downtime
01 maistudioMLX model reload · 2 min
24 avriltowerOS kernel panic, replaced PSU
29 maistudioserving consolidé sur :8500
11 juinstudiomigration :8500 vers vllm-mlx (fork souverain)
12 avril—router v9 shipped