Lemonade GPU deployment (Ubuntu 24.04 + llama.cpp)

Deploy Gemma 4 E4B-it GGUF with Lemonade Server + llamacpp on the Ryzen AI iGPU. Not FastFlowLM, not vLLM.

Prerequisite: Lemonade environment. NPU: FastFlowLM · Lemonade NPU.

---

Item	Value
Lemonade id	Gemma-4-E4B-it-GGUF (confirm with lemonade list)
Weights	unsloth/gemma-4-E4B-it-GGUF (often Q4_K_M)
Backend	--llamacpp vulkan (try rocm if the kernel qualifies)
API	http://127.0.0.1:13305/api/v1
Device	gfx1151 iGPU, does not occupy /dev/accel

lemonade backends install llamacpp:vulkan
lemonade pull Gemma-4-E4B-it-GGUF --source modelscope
lemonade load Gemma-4-E4B-it-GGUF --llamacpp vulkan
lemonade status

If llamacpp:rocm reports Linux kernel missing support, stay on Vulkan. See gfx1151 Linux.

Smoke:

BASE=http://127.0.0.1:13305/api/v1 MODEL=Gemma-4-E4B-it-GGUF \
  python3 src/npu/examples/chat_gemma4.py

Frontend: Base URL http://127.0.0.1:13305/api/v1, API key lemonade. Stop: lemonade unload and, if needed, sudo systemctl stop lemond.