Hacktoberfest 2026: le issue che i maintainer hanno segnato per ottobre, aperte e adatte ai principianti. Sfoglia le issue Hacktoberfest

Local node's GPU/CPU/memory enrichment is never refreshed after startup

Aperta
  • Qualcosa non va?
#4 1 commento 0 reazioni 1 assegnatario Vedi su GitHub

I maintainer di solito rispondono entro 4 giorni

@ckelseynv ci sta già lavorando.

Dal 11/9/2026.

  • #5 di @canja006 — aperta

Valutazione

Questa issue non è ancora stata valutata.

Descrizione

The discovery directory's enrichment for the local node is captured once and then frozen. Peers converge; self does not.

What I measured

Ubuntu node with an RTX 3060, paired with a macOS node, v0.1.1-463 from the .deb.

I loaded a 7 GB model into LM Studio with full GPU offload. nvidia-smi went from 121 MiB to 9329 MiB. node-info served the new figure within its usual cadence — curl 127.0.0.1:14318/v1/node-info returned vram_used_bytes: 9782165504 with telemetryValid: true, msSince: 280. The desktop UI's VRAM graph showed the step too, because it reads the node directly.

The discovery snapshot did not move. discovery:get-nodes kept returning vram_used_bytes: 126877696 — the value from daemon start — for as long as I watched, checked at 20 s intervals over several minutes. Restarting the broker made it correct again, until the next change.

Why

onBrowse deliberately skips our own entry (daemon.go: "Never let a browse event touch our own entry. Self is registry-driven"), so the path that re-enriches every peer on every browse never runs for self. publishSelf — the only thing that enriches self — is called at startup, on service (un)register, and after an identity/address change. refreshPeersLoop refreshes advertised addresses, the mesh, cluster identity and models on its 15 s tick, but nothing re-reads self's node-info.

Worth noting that the comment above noteNodeInfo already states the intended behaviour:

The enrichment sweep repeats every peerRefreshInterval for as long as a node is advertised

That holds for peers, not for self.

Who sees it

Only consumers of the discovery snapshot — which includes nvpair-job-scheduler, since the broker fans the same snapshot to it. The UI is unaffected, which is why this is easy to miss.

Repro

  1. Pair two nodes; note the local node's gpus[].vram_used_bytes via discovery:get-nodes.
  2. Load a large model with GPU offload on that node.
  3. curl 127.0.0.1:14318/v1/node-info — fresh.
  4. discovery:get-nodes — unchanged, indefinitely.

A fix that works

Calling d.publishSelf() on the refreshPeersLoop tick converges the directory within 20 s. Verified live in both directions, loading and unloading. It does mean one self node-updated per tick, since publishSelf emits unconditionally; the tidier alternative is an applyInfo path with a changed check, mirroring applyModels. Happy to open a PR for either shape if useful.

Lingua principale
Go
Stelle
1.6k
Fork
266
Merge medio
3g 18h
PR unite (30g)
28

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di NVIDIA/Personal-AI-Router

Tutte le issue di NVIDIA/Personal-AI-Router

Issue simili

Altre issue su Go

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.