Skip to content

Skylark-Software/FlightControl

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

12 Commits
 
 
 
 
 
 
 
 

Repository files navigation

FlightControl

FlightControl

Note: FlightControl is being rewritten in Rust. The Speedloader backend is available at Skylark-Software/Speedloader. This repository preserves the project overview and screenshots from the PyQt5 prototype.

A PyQt5 dashboard for managing GPU inference servers, RAM-tiered model caching, and distributed block storage across multi-node clusters.

Dashboard


Features

Dashboard

  • Real-time CPU, GPU, and RAM monitoring with per-core/per-device breakdown
  • Model lifecycle management (start, stop, health checks)
  • Docker container status and control
  • Live log viewer with throughput parsing
  • RDMA/InfiniBand resource panel

Speedloader (RAM Tier Management)

  • Storage Tiers — Multiple tmpfs-backed RAM tiers with priority-based model placement
  • Model Placement — Strategies: fastest-fit, fill-first, round-robin with per-model tier pinning
  • Remote Hosts — NVMe-oF over RDMA block devices from remote machines (swap, pool, or raw)
  • Local Raw Blocks — Loop-backed block devices from local RAM for RAID membership
  • Block Aggregation — Combine blocks across hosts into mdadm RAID 0 arrays for high-throughput model loading

The Speedloader backend has been extracted and rewritten in Rust as a standalone tool: Skylark-Software/Speedloader

Speedloader

Storage

  • BeeGFS/NFS/local storage monitoring
  • Hot (RAM) and Cold (disk) model inventory
  • One-click promote/demote between tiers

Storage

Models

  • Pull models from Ollama registry
  • Import local GGUF files
  • Search and download from HuggingFace

Models

Scripts

  • Script editor with syntax highlighting
  • Configurable launch parameters (GPU layers, context size, threads, batch size)
  • GPU expert pinning for MoE models
  • One-click run with output capture

Scripts


Related Projects

Project Description
Speedloader Rust rewrite of the Speedloader backend — fast hybrid RAM/storage management for LLM models
EagleBranch EAGLE speculative decoding for Mistral Large 3 and DeepSeek (llama.cpp fork)

License

Copyright (c) 2025-2026 Skylark Software LLC. All rights reserved.

This software is provided for viewing and reference purposes only. No permission is granted to use, copy, modify, merge, publish, distribute, sublicense, or sell copies without explicit written permission. See LICENSE.


Skylark Software

About

GPU inference dashboard with RAM-tiered model caching and distributed block storage. PyQt5 prototype — Speedloader backend rewritten in Rust.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors