TP1, TP2, and the spare
TP2 (the live cluster) shards DS4F across Spark 1 and Spark 2. TP1 lives on one machine: helper 35B, Qwen 27B, Gemma, vision 4B. A third Spark is a spare TP1 slot — the cluster can keep running while something else sits on the extra 121 GB.
This is a budget toy, not a live probe. It will not start or stop anything. H3 rank 0 is the measured 89.4 GB on node 1; rank 1 is 41.4 GB on node 2. Qwen 27B uses the measured 36 GB KV pool. Gemma is the 19 GB Ollama line.