thread_id: 01a10985-1eed-7260-ad2d-49a2af9c96c4
updated_at: 1791176837

Documented expert arena definition levers in frank-primary vLLM fork: arena = budget - non-expert - KV_reserve(9190B/token × max_model_len) - graphs - overhead. Key levers: max_model_len (lower=increase), --expert-precise-memory (reduces overhead 7.60→5.25 GiB), VLLM_CUDAGRAPH_* env vars (graph term), gpu_memory_utilization (budget). Pool/KV_BYTES is NOT an arena lever. --expert-offload-mem controls host offload (120 GiB). Current shape G: 17.77 GiB / 7762 slots, UTIL must stay 0.982.