Hardware validation records¶

Single-node (0.1.0) and dual-node (0.2.0) gates. Fill rows on real hardware.

Interactive and discovery¶

[ ] First bare rondine run performs doctor, recommendation, and selection
[ ] A later bare rondine run offers the active plan/presets without repeating doctor
[ ] Loading a preset updates the active last.json plan
[ ] rondine suggest --no-hub works offline
[ ] Hub API failure falls back to catalog suggestions with a note
[ ] rondine suggest --context 65536 excludes catalog models below 65,536 tokens

[ ] rondine doctor reports VRAM + GPU name + free disk
[ ] rondine suggest matches cuda-* target and fits against VRAM when resident
[ ] rondine plan <moe> --memory-mode hybrid selects hybrid when VRAM alone is short but RAM+VRAM is enough
[ ] Hybrid dry-run emits --cpu-moe and -ngl auto (or equivalent) for MoE GGUFs
[ ] rondine setup builds/installs llama.cpp with CUDA (or vLLM on Linux)
[ ] rondine serve + rondine verify --profile coding pass

[ ] rondine setup installs llama.cpp and/or MLX
[ ] rondine plan auto --profile coding selects a fitting Qwen/Gemma variant
[ ] rondine plan glm-5.2 rejects on ≤48GB hosts without --allow-oversize
[ ] rondine plan glm-5.2 --memory-mode mmap --allow-oversize is experimental, fits=false, and forces -ngl 0
[ ] Insufficient free disk rejects mmap plans / pulls before download
[ ] rondine serve + rondine verify --profile coding pass for a fitting small/medium model