Local AI experiments

Running models on my own hardware

Home lab inference, fine-tuning experiments, llama.cpp optimization, and practical AI tooling. No cloud required — everything runs locally on home lab hardware.

Experiments & Setup

Notes on models, quantization, server configs, and tooling for local AI workflows.

Inference

llama.cpp Server

Running GGUF models via llama.cpp with CUDA offload. Config tuning for speed vs quality trade-offs.

GGUFCUDAQ4_K_M
Coding

AI-Assisted Development

VS Code integration with local models for code completion, refactoring, and explanation. Faster than waiting on cloud APIs.

VS CodeLocalCopilot alt
Optimization

Quantization Benchmarks

Testing different quant methods — Q4, Q5, Q6, Q8 — to find the sweet spot between quality and inference speed.

Q4_K_MQ5_K_MBenchmarks
Model

Mistral 7B

Primary coding and general-purpose model. Strong instruction following, fast inference, and great with local fine-tunes.

Mistral7BCoding
Future

Fine-Tuning

Planning and testing LoRA fine-tunes on custom datasets. Domain-specific models for project notes, code review, and documentation.

LoRACustom dataQ-LoRA
Stack

Local AI Stack

llama.cpp, Ollama, text-generation-webui, and custom Python tooling for the local AI pipeline.

llama.cppOllamaPython

Setup Notes

Hardware and software details for running local AI on a home lab.

Hardware

Running on home lab hardware with dedicated GPU for CUDA inference. Proxmox host with GPU passthrough to the AI LXC container. Models stored on fast NVMe for quick loading.

ProxmoxGPU PassthroughNVMeCUDA

Software Stack

Ubuntu LXC with llama.cpp, GGUF format models, and custom Python scripts for model management and inference. VS Code extensions for local AI-assisted coding workflows.

llama.cppGGUFUbuntuPython