Thanks, good question. I tested it on Windows (RX 7900 XTX, Vulkan, 2026.45.3):
- Each model type has its own LRU slot, so an image request does not unload the chat model. A second LLM only evicts the first LLM.
- There is no global admission check before a load. If a load fails, Lemonade unloads all other models and retries once.
- There is also an opt-in global policy (auto_evict) that evicts idle models or the most "disposable" one under VRAM pressure. Pinning protects a model.
One thing I found while testing: on Windows the load does not fail when VRAM is full, the driver spills to shared memory. After an SDXL generation the chat model stayed loaded but dropped from ~103 to ~19 tok/s until I reloaded it. So "not cold" does not always mean "fast". In the four-model example everything fit, so this did not happen there.