r0b0tlab/deepseek-v4-flash-dspark-2x-dgx-spark

This repository is a set of scripts and configuration files for running a specific large AI language model, DeepSeek V4 Flash DSpark, on a pair of NVIDIA DGX Spark computers connected together. It is not the AI model itself, and it is not a general purpose tool for running any AI model. It is a very specific, carefully pinned setup for one exact model version on one exact hardware combination. The project builds a Docker container image with exact, locked versions of every important piece: the model itself, the underlying vLLM serving software, PyTorch, CUDA, and a component called FlashInfer that speeds up how the model processes requests. Every one of these pieces is pinned to a specific version number or file hash, and the software actively checks these pins before it will run, refusing to start if anything does not match what is expected. The authors describe this as a fail closed design, meaning that if something looks wrong, the system stops rather than continuing with a possibly broken or slower setup. The repository includes scripts to build the container image, configure the two connected machines to talk to each other over a fast network connection, launch the model server, and then run tests and benchmarks against it once it is running. These benchmarks measure things like how quickly the model starts generating a response, how fast it generates text once started, and how well it handles many requests at the same time. Results are saved as files that can be compared against each other later. Using this project requires two NVIDIA DGX Spark systems with compatible drivers, Docker, a specific type of fast network hardware connection between them, and access to download the exact DeepSeek model files. This is specialized infrastructure work aimed at people who already run large AI models on dedicated hardware, not something that could be tried casually on a laptop or a single cloud server.