Skip to content

Clusters (0.2)

Rondine only wraps upstream multi-node mechanisms. It does not invent a scheduler.

Requirements

  • Homogeneous nodes (same platform / memory class)
  • Passwordless SSH
  • Matching engine versions on every node
  • Private high-speed fabric (Thunderbolt / RDMA / RoCE) — never silent Ethernet fallback
  • Dual-node is the validated target; larger inventories are diagnostic-only

Mac (MLX)

rondine cluster init lab --kind mac --head mac1.local --worker mac2.local
rondine cluster doctor lab
rondine cluster plan lab --engine mlx --model-repo unsloth/Qwen3.6-27B-UD-MLX-4bit

Uses mlx.launch --backend jaccl --hostfile ….

Mac GGUF (experimental)

llama.cpp RPC (ggml-rpc-server + --rpc) is trusted-LAN only. Do not expose it.

DGX Spark

rondine cluster init sparks --kind spark --head spark-a --worker spark-b --interface enP2p1s0f1np1
rondine cluster doctor sparks
rondine cluster plan sparks --engine vllm --model-repo unsloth/Qwen3.6-35B-A3B-NVFP4-Fast

Follow NVIDIA connect-two-sparks + Ray multi-node serving. Pin NCCL to ConnectX RoCE interfaces before serving.

Serve workflow

cluster serve validates the inventory and prints the same upstream commands as cluster plan. It is dry-run only by default:

rondine cluster serve sparks --engine vllm \
  --model-repo unsloth/Qwen3.6-35B-A3B-NVFP4-Fast

Rondine intentionally refuses unsupervised remote orchestration, including with --no-dry-run. Review and run the printed commands on the appropriate nodes.