GPUqueue: applications, coordinator and Ollama

GPUqueue: applications, coordinator and Ollama An architecture diagram generated by Archify. Local applications · Chat, indexing, automation · Architecture component Local applications Chat, indexing, automation GPUqueue coordinator · HTTP 127.0.0.1:11436 · Architecture component · PRIORITY + ADMISSION GPUqueue coordinator HTTP 127.0.0.1:11436 PRIORITY + ADMISSION Ollama · Internal HTTP :11434 · Architecture component Ollama Internal HTTP :11434 Watchdog · Windows scheduled task · Architecture component Watchdog Windows scheduled task Tray + web dashboard · GPU status / separate CPU tab · Architecture component Tray + web dashboard GPU status / separate CPU tab Durable local store · SQLite + Windows DPAPI · Architecture component Durable local store SQLite + Windows DPAPI Local GPU / CPU · Models execute on this host · Architecture component Local GPU / CPU Models execute on this host Requests / results Admitted inference Health / restart Read status Jobs + results Model execution Legend Frontend Backend Database Message bus External

A shared front door

  • • Integrated apps send inference to :11436.
  • • Results return through the coordinator.

Command workloads too

  • • client.py requests a lease, then launches its own process tree.
  • • Managed CUDA jobs share the same admission policy.

Explicit limits

  • • Direct backend / CUDA callers can bypass the queue.
  • • Uncertain cleanup stays blocked; no speculative replay.