Experimental account
dgx-spark-step37-dual. Step 3.7 Flash (198B MoE, NVFP4) across two DGX Sparks — vLLM TP=2, RoCE RDMA, 262K context, 18.5 tok/s
dgx-spark-ds4-benchmark. DeepSeek V4 Flash distributed inference benchmark on dual NVIDIA DGX Sparks using ds4 (DwarfStar)
