A two-node DGX Spark serving recipe for DeepSeek-V4-Flash with vLLM tensor parallelism, speculative decoding and 1M context.