Lemonade NPU deployment: Qwen3.6-35B-A3B

Same caveat as Gemma 4: systemd lemond cannot see the user FLM cache. Use FastFlowLM Qwen3.6.

lemonade unload 2>/dev/null || true
pkill -x flm 2>/dev/null || true
unset ROCM_PATH HIP_PATH LD_LIBRARY_PATH
flm serve qwen3.6-moe:35b-a3b --host 0.0.0.0 --port 8219 --ctx-len 32768 --q-len 20 --socket 20 --cors 1

Resident ~29 GiB — stop GPU-side large models first. lemonade load qwen3.6-moe:35b-a3b is only worth trying if lemond runs as your login user without ProtectHome and flm:npu is installed.