GPUStack doubles GLM inference on four DGX Spark nodes to 51.3 tok/s

A four-node NVIDIA DGX Spark cluster managed by GPUStack and served through vLLM raised single-stream decode throughput for GLM-5.3-Flash-NVFP4 from 24.3 to 51.3 tok/s. RoCE and NCCL tuning lifted 256 MB AllReduce bandwidth from 4.0 to 22.9 GB/s, while aggregate throughput at four concurrent requests rose from 72.0 to 104.1 tok/s.

• Single-stream decode rose from 24.3 to 51.3 tok/s (2.11×)
• 256 MB AllReduce bandwidth improved from 4.0 to 22.9 GB/s (~5.7×)
• Aggregate throughput at four requests grew from 72.0 to 104.1 tok/s
• TP=4 across four nodes with 128 GB memory each and ConnectX-7