Spark

Spark is the control plane and runtime for my private two-node NVIDIA DGX Spark inference cluster. The two machines serve one logical DeepSeek-V4-Flash-0731 model with tensor parallelism, a 1,048,576-token context window, and a dedicated high-speed link between them.

The Mac remains the control plane. Hosted Codex plans and reviews work, while an isolated local Codex worker can execute a bounded task on the cluster. A personal Spark plugin handles status, the private tunnel, delegation, job tracking, cancellation, and contribution reports. Every delegated job starts with an agreed split of responsibilities and explicit acceptance checks; the hosted agent then inspects the result before accepting it.

The network boundary is intentionally narrow. The vLLM API binds only to loopback on the head node and reaches the Mac through a repository-managed Tailscale SSH tunnel. The Sparks are inference-only: they hold no working repositories or GitHub credentials. Tensor-parallel traffic stays on the dedicated ConnectX-7 fabric, where startup checks make a missing RDMA path fail visibly rather than falling back silently.

The runtime pins the model revision, container image, serving configuration, and selected correctness patches. Independent boot launchers restore the two-rank service after a reboot, persist expensive compilation caches, and warm the request shapes used in normal work. The same cluster also provides the local model behind my self-hosted OpenClaw assistant.

Commissioning is part of the project rather than an afterthought. The acceptance suite exercises the native Responses API, streaming, reasoning, tool calls, structured output, concurrency, long-context behaviour, recovery, and the live Codex plugin. Raw reports and dated commissioning notes stay in the repository, so changes to performance or correctness have evidence behind them.

I built Spark to make local inference useful for real agent work without turning it into an opaque appliance. The result is private, inspectable, reproducible, and deliberately subordinate to a reviewed workflow: the cluster does the bounded work; hosted Codex remains responsible for judgment and delivery.