Share one inference engine across every lilbee process on the machine#547

Open

tobocop2 wants to merge 67 commits into

mainfrom

fix/serve-data-dir-singleton

Commits on Jul 17, 2026

Commits on Jul 18, 2026

Commits on Jul 19, 2026

Add the 600-round endurance results to the shared-engine benchmarks
tobocop2
committed
Commit the raw sweep and soak data behind the shared-engine benchmarks
tobocop2
committed
Name the model-architecture and workload limits of the shared-engine benchmarks
tobocop2
committed
Plan the coding-agent stress scenarios and reels
tobocop2
committed
Pick current agentic coding models for the stress scenarios
tobocop2
committed
Tier the coding-agent tasks per model and fix the reel syntax theme
tobocop2
committed
Run the demos on consumer hardware and ground every task in real work
tobocop2
committed
Keep issue IDs out of the demo plan
tobocop2
committed
Offload MoE experts to system memory and stop emitting unloadable KV flags
tobocop2
committed
Report the expert-offload settings in the diagnostic payloads
tobocop2
committed
Make the setup self-check derive its flags from the fleet, not by hand
tobocop2
committed
Size the estimate against the experts the launch actually offloads
tobocop2
committed
Stop marking offloadable models unrunnable in the catalog
tobocop2
committed
Serve a tensor-split chat on as many slots as its cards' KV headroom holds
tobocop2
committed
Bind the shared engine against the placeable set, not the raw config
tobocop2
committed
Make expert-offload sizing MoE-aware and reload-triggering
tobocop2
committed
Never stop a warm engine the process cannot replace
tobocop2
committed
Bound the engine build lock and unify the acquisition ladder
tobocop2
committed

Commits on Jul 20, 2026