What is dual-spark-optimizations?

dual-spark-optimizations is a configuration project for running a large language model quickly across two NVIDIA DGX Spark machines connected directly to each other. It combines two existing open source projects, a set of modified model weights and a fast two-node serving runtime built on vLLM, into one tested setup, rather than introducing new model code of its own. The main contribution is a serving option called the adaptive profile, which the README recommends as the default for most users. It automatically behaves like a fast single-request setup when only one request is active, and switches to handling up to four requests at once when traffic overlaps, without needing any special flag or model reload. Two other profiles are also included: a fast profile tuned purely for single-request speed, and a long-c4 profile that supports a much larger one-million-token context window at the cost of some speed. The README backs its claims with detailed benchmark tables measured on an actual two-Spark cluster, comparing decode speed and aggregate throughput across different prompt sizes and profiles, along with network fabric test results showing the direct high-speed connection between the two machines. It is upfront that these numbers are specific to this hardware and model combination and not directly comparable to results from other setups. Setting it up requires two DGX Spark machines connected by a single cable, Ubuntu 24.04, Docker and Docker Compose, matching Linux usernames on both machines, and approved access to the gated model weights on Hugging Face, since the model itself is not included in the repository. The README walks through cloning the repo, filling in a configuration file with machine specific values, and a script that checks and fixes duplicate machine IDs that can occur on cloned factory images before the two nodes are connected. This is a specialized project aimed at people who already own or have access to this specific two-node hardware setup.