Open source: vLLM tuning recipes for H100 clusters

Continuous batching, KV-cache quantization, and multi-node tensor parallel configs measured on GridNexa GPU Cloud.

This note summarizes reproducible vLLM flags for H100 clusters during Platform Preview.