FastFlowLM / Lemonade troubleshooting
Common NPU and Lemonade issues. For ROCm / HIP GPU problems, see 00-Environment verification.
| Symptom | Fix |
|---|---|
No flm | Follow FastFlowLM environment, then reboot |
No NPU device found | Check /dev/accel/accel0 and render group; sudo usermod -aG render,video $USER then re-login |
SVA bind -19 | cmdline must have amd_iommu=on iommu=pt; update GRUB and reboot |
No such device with index '0' | unset ROCM_PATH HIP_PATH LD_LIBRARY_PATH; check xrt-smi examine |
| NPU gone after firmware-only swap | driver + firmware must match; use PPA amdxdna-dkms |
modinfo amdxdna not under updates/dkms | no reboot yet, or still on in-tree module |
ulimit -l not unlimited | write limits.conf, re-login |
| NPU vs large-GTT recipe | GTT often uses amd_iommu=off, incompatible with FLM |
No lemonade | sudo apt install lemonade-server |
status cannot reach 13305 | process /usr/bin/lemond; ss -tlnp | grep 13305; sudo systemctl start lemond |
| 13305 will not stop | sudo systemctl stop lemond (unit is not lemonade-server) |
llamacpp:rocm = kernel missing support | use llamacpp:vulkan; see gfx1151 Linux |
flm:npu missing / GitHub timeout | use native FastFlowLM |
load gemma4-it:e4b 404 | systemd lemond ProtectHome hides user FLM cache; use 8219 |
Lemonade and flm serve fight over NPU | lemonade unload + stop lemond, or pkill -x flm; keep one client |
| 503 | NPU queue full; lower concurrency or raise --q-len |
| OOM with a GPU model | unified memory; stop vLLM / Ollama / large GGUF first |
flm validate fails in a sandbox | no /dev/accel is expected; run on a real terminal |