Bash · CPU/NVMe inference · systemd lifecycle

Colibri Local-AI Infrastructure

An operator-focused toolkit for a persistent CPU/NVMe-backed Colibri service that keeps the GPU available, verifies and surgically repairs model weights, derives adaptive resource profiles and integrates with Open WebUI.

Active source-reviewed infrastructure
Role
Systems architect, principal implementer and maintainer
Evidence
2 direct public links
  • Bash
  • C/C++ build tooling
  • systemd
  • Node.js
  • Docker networking
  • OpenAI-compatible APIs
  • NVMe

Purpose and value

Why this system exists

Colibri Setup demonstrates practical large-model operations under real resource constraints: exact revision control, checksum repair, RAM ceilings, CPU-only enforcement, multi-NVMe reads and recoverable service management.

Engineering role

Systems architect, principal implementer and maintainer

System structure

Architecture

  1. 01

    Install, validation, repair, service lifecycle and user-interface integration are separate operator commands.

  2. 02

    Model identity is pinned and verified through metadata and checksums before selective repair.

  3. 03

    Adaptive profiles cap memory use against available RAM while enforcing the CPU-only execution path.

  4. 04

    Physical multi-NVMe reads and service integration are explicit configuration rather than hidden host assumptions.

Boundaries

Principal interfaces

Operator-focused Bash command surface
Colibri runtime and model files
systemd service lifecycle
OpenAI-compatible API and Open WebUI integration

Judgement

Engineering decisions

Repair only mismatched model files instead of performing broad destructive redownloads.

Cap adaptive profiles at 80% of available RAM and keep GPU resources outside the service path.

Keep model data persistent across stop, uninstall and service reconfiguration.

Confidence model

Delivery, validation, and maintenance

  • 53-commit complete reviewed history
  • Checksum verification and surgical repair
  • CPU-only startup and resource-profile tests
  • Custom-profile hardening and lifecycle coverage

Direct sources

Inspect the evidence

Current verified scope

53 default-branch commits in the complete reviewed history; current source covers an 80% available-RAM ceiling, CPU-only enforcement, checksum repair and lifecycle tests.

Interpretation boundary

Performance depends on host resources and upstream model/runtime behaviour; no benchmark claims are invented.

Source reviewed to 2026-07-29

Rapid navigation

Go to

navigate Enter open