reederey87/dgx-spark-2x-deepseek-v4-flash

This project is a step by step deployment kit for running a large AI language model called DeepSeek-V4-Flash-DSpark across two NVIDIA DGX Spark computers connected together as one machine. It does not contain the AI model itself or the serving software, instead it provides scripts, configuration files, and documentation that pull those pieces from other sources and wire them together correctly on this specific two machine hardware setup. The two computers are linked by a single fast network cable and work together using a technique called tensor parallelism, splitting the model's work across both machines so they act as one inference engine. One machine, called the head, exposes an interface that looks like the well known OpenAI API, while the other, the worker, does not accept outside connections and only helps process requests behind the scenes. Setting this up is a multi stage process run mostly from a separate control computer over a remote connection. The steps include preparing each machine, verifying the network connection between them works correctly at full speed, building the software image both machines will run, downloading the AI model's weight files, distributing everything to both machines, and finally starting the service. After that, background system services keep the cluster running automatically and can restart it if something goes wrong. The README is explicit that this is an experimental, mostly one person maintained project built on fast changing, prebuilt components, and that its own performance measurements only reflect one specific pair of machines rather than a general guarantee. It documents a number of adjustable settings that control things like how much conversation history the model can consider and how many requests it can handle at once, along with guidance on which combinations reduce the risk of running out of GPU memory. This project is aimed at engineers with access to this exact expensive dual DGX Spark hardware who want a documented, repeatable way to serve this specific large model. It is released under the Apache 2.0 license, which allows free use, modification, and redistribution, including commercially.