darkmatter2222

HomeAILab — AI skill for Claude Code

AI community

Home AI inference lab running Qwen3.8 across RTX 5090, RTX 3090, and DGX Spark with vLLM, llama.cpp, custom Go routing, benchmarking, monitoring, long-context inference, and Claude Code integration.

How to install HomeAILab

This entry records only its repository, not the path inside it, so there is no exact command to give. Open darkmatter2222/HomeAILab and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What HomeAILab does

Home AI inference lab running Qwen3.8 across RTX 5090, RTX 3090, and DGX Spark with vLLM, llama.cpp, custom Go routing, benchmarking, monitoring, long-context inference, and Claude Code integration.

Alternatives in AI

Repomix — 📦 Repomix is a powerful tool that packs your entire repository into a single, AI-friendly file 22.7k ★
Qwen Code — A command-line AI workflow tool adapted from Gemini CLI, optimized for Qwen3-Coder models with enhanced parser 20.8k ★
Inference Serving — ai-research-skills vLLM, SGLang, TensorRT-LLM, llama.cpp 5.4k ★

README

HomeAILab — Qwen3.8-27B GPU Inference Fleet

A working, single-user GPU inference fleet that serves **Qwen3.8-27B** (abliterated "Uncensored") and a **180B MoE** vision model across three heterogeneous machines, fronted by a custom **Go LLM router** that gives any OpenAI- or Anthropic-compatible client (including Claude Code) a single, deterministic endpoint. Everything is deployed through **Portainer** stacks — the YAMLs in this repo are the source of truth you maintain in the Portainer UI.

**If this helps, give it a star.** It's the difference between a project and a reference.

The fleet in one line

Host	GPU	Role
RedPCv2 (`.37`)	RTX 5090 (32 GB)	vLLM, fastest text decode
Databrick (`.48`)	RTX 3090 (24 GB)	llama.cpp, MTP spec decode, 262K verified
DGX Spark (`.39`)	NVIDIA GB10 (122 GB unified)	NVFP4 vLLM + 180B MoE vision (Flash-Next)

A deterministic router on Databrick (`:8001`) cascades **5090 → 3090 → Spark**, pins vision requests to the Spark, and gives each client a stable `local-coding` alias that transparently maps to the backend's `qwen3.8`.

How it works

High-level architecture

flowchart LR
    CC[Claude Code]
    API[Other OpenAI-Compatible Clients]

    R[Go LLM Router :8001
Data Plane]
    RT[(In-Memory
Routing Table)]

    CP[Control Plane]
    D[Discovery Scanner]
    H[Health Monitor]

    N5090[RTX 5090
vLLM :8201]
    N3090[RTX 3090
llama.cpp :8101]
    DGX[DGX Spark
Flash-Next :8401]

    CC -->|OpenAI / Anthropic API| R
    API -->|OpenAI / Anthropic API| R

    R <--> RT
    CP --> RT
    D --> CP
    H --> CP
    D -. scan + health .-> N5090
    D -. scan + health .-> N3090
    D -. scan + health .-> DGX

    R -->|selected request only| N5090
    R -->|selected request only| N3090
    R -->|selected request only| DGX