DeepSeek-v4-Flash-DSpark-2x-DGX-Spark

Why we included this project

Running DeepSeek V4 Flash 0731 across two DGX Spark nodes is exactly the kind of setup that looks simple on paper and eats an afternoon in practice. This repo is a concrete recipe for it: vLLM with tensor parallelism across both hosts, DSpark speculative decoding, a 1M-token context ceiling, and a pinned container image and checkpoint so you are not chasing version drift. It also deals with the unglamorous parts of multi-node serving, like NCCL/RoCE fabric settings and keeping the Hugging Face cache warm on both machines. The KV-cache math is spelled out, including why six concurrent 1M-token requests will not all fit. You get start, stop, status, and smoke-test scripts plus dated benchmark tables, so you can reproduce the setup and see how your numbers compare. If you already run DGX Sparks and want a known-good DeepSeek deployment, this will save you real time.