wtf is dspark-vllm-gx10?

This project is a technical adaptation that lets a large AI language model called DeepSeek V4 Flash run across two connected NVIDIA DGX Spark, or ASUS GX10, machines at once. These are specialized computers built for running big AI models, and this repository modifies a popular AI serving framework called vLLM so the model can split its work across both machines using a fast direct connection between them. The model itself is very large, more than 150 gigabytes of data, and the README explains it will not fit or run on a single one of these machines because of memory limits. By splitting the work in two, using a compressed numeric format called NVFP4 to shrink memory use, and adding custom code to handle how the model processes attention and routes requests between its many internal expert components, this project makes it possible to actually run the model and serve responses to users. The repository comes with setup scripts to install and start the two machines in the correct order, since one machine must be ready before the other connects to it. It also includes a live dashboard, a small tool with no extra dependencies, that shows real time statistics like how fast the model is generating text, memory and power use, temperature, and the status of each machine while it runs. The README documents specific benchmark results comparing this version against an earlier one, showing double digit percentage speed gains in most tests. It notes the exact software versions and hardware configuration used so results can be reproduced, and includes scripts to rerun the same benchmarks. This is not a beginner tool. It assumes access to expensive specialized hardware and existing familiarity with running large AI models, and the model weights themselves are not included, only licensed separately. Repository code is released under the MIT license, while some files adapted from the underlying vLLM project remain under the Apache 2.0 license.