# Local AI Infrastructure Lab

> A dated field record by Deshimaru Sakaguchi: three 128GB Strix Halo nodes and ten local model services.

Baseline: September 23, 2026 (Asia/Tokyo). Language: English. Source revision: 1.0.

The HTML reports and their Markdown sources contain the same approved material. They distinguish directly captured state, construction notes, design intent, and unmeasured work. Private raw logs are not published.

## Reports

- [Local AI Infrastructure Lab](https://deshimarusakaguchi.com/lab/local-ai-cluster/): A dated field record of a three-node Strix Halo local AI cluster: architecture, memory, context, persistent services, and future workload studies.
  - [Markdown source](https://deshimarusakaguchi.com/lab/local-ai-cluster/index.md)
- [Building a Fully Local 120B / 72B / 31B AI Cluster with Three 128GB Strix Halo Mini PCs](https://deshimarusakaguchi.com/lab/local-ai-cluster/architecture/): Architecture, model roles, software versions, and memory observations from three 128GB Strix Halo nodes hosting ten local AI services.
  - [Markdown source](https://deshimarusakaguchi.com/lab/local-ai-cluster/architecture/index.md)
- [Memory, Vulkan, and AMD Strix Halo](https://deshimarusakaguchi.com/lab/local-ai-cluster/memory-vulkan-strix-halo/): A 96GB UMA field record: Vulkan/RADV deployment, TTM/GTT differences, memory snapshots, historical ROCm failures, and micro-batch observations.
  - [Markdown source](https://deshimarusakaguchi.com/lab/local-ai-cluster/memory-vulkan-strix-halo/index.md)
- [Model Context vs. Runtime Context](https://deshimarusakaguchi.com/lab/local-ai-cluster/context-runtime/): GGUF declarations, runtime-recognized limits, requested context, effective slots, and captured request rejections in a local llama.cpp cluster.
  - [Markdown source](https://deshimarusakaguchi.com/lab/local-ai-cluster/context-runtime/index.md)
- [Persistent AI Services and Operations](https://deshimarusakaguchi.com/lab/local-ai-cluster/persistent-services/): Ten local model endpoints: the actual systemd supervision boundary, API contracts, Dify registration, remote management, and operational limits.
  - [Markdown source](https://deshimarusakaguchi.com/lab/local-ai-cluster/persistent-services/index.md)

## Machine-readable resources

- [Document manifest](https://deshimarusakaguchi.com/lab/local-ai-cluster/manifest.json): page URLs and SHA-256 hashes of the approved Markdown files.
- [Full Markdown collection](https://deshimarusakaguchi.com/lab/local-ai-cluster/llms-full.txt): the five source documents in one UTF-8 text file.
- [Section sitemap](https://deshimarusakaguchi.com/lab/local-ai-cluster/sitemap.xml): canonical HTML URLs for this lab.

## Scope

Phase 1 is a construction/deployment field record, not a benchmark certification. Controlled throughput, wall power, thermal behavior, sustained-load stability, repeated-run variance, and the complete Sanapedia/Dify research workflow belong to a later phase.