What it is
AMD’s own local AI server. The primary LLM recipe wraps llama.cpp (Vulkan or ROCm backends, plus a Linux system-backend). It also ships other recipes (whisper.cpp, sd-cpp, Ryzen AI NPU/FLM). OpenAI-compatible HTTP on 127.0.0.1:13305/api/v1.
Architecture in plain English
lemonade / lemond service. You install a recipe backend (llamacpp:vulkan or llamacpp:rocm), pull owner/repo:variant, and serve /api/v1. It is not this site’s Vulkan llama.cpp tree. NPU / Hybrid / FLM paths are Ryzen AI, not a R9700 how-to.
What it does well
An AMD-promoted wrapper with an explicit Vulkan vs ROCm switch, if a GPU meter moves.
What it does badly
Matching this site’s compile flags or benches; silent CPU; treating /v1 like llama-server.
ROCm / Radeon
Official llamacpp page (checked 2026-10-14): ROCm for AMD Radeon RX 6000/7000 series (RDNA2/RDNA3/RDNA4) and Ryzen AI iGPUs (Strix Point/Halo). Nightly bins: gfx1150, gfx1151, gfx103X, gfx110X, gfx120X. v2026.41.1 moves ROCm backends to ROCm 10; first ROCm load after upgrade downloads about 2.6 GB. We do not invent that the stable blob lists gfx1201 / R9700.
Vulkan. Yes — “recommended for most GPU users” on their llamacpp page. First pick on this site.
Supported quants
GGUF via llama.cpp. Q3 is listed as estimate/community unless measured. Q1/Q2 stay off.
Q3 is listed as estimate/community unless measured; Q1/Q2 stay off. Prefer Q5. See quants and the calculator.
Install on an R9700
# Ubuntu 24.04+ (Lemonade Ubuntu docs, checked 2026-10-14)
# sudo add-apt-repository ppa:lemonade-team/stable
# sudo apt install lemonade-server
# lemonade backends install llamacpp:vulkan
# Prefer the Vulkan llama.cpp you already built if this sits on CPU.
Run
lemonade status
# lemonade pull unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q5_K_XL
# lemonade load unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q5_K_XL --ctx-size 262144 --llamacpp vulkan
curl -s http://127.0.0.1:13305/api/v1/models
Tuning flags
lemonade config set llamacpp.backend=vulkan (or rocm). Env LEMONADE_LLAMACPP. --llamacpp-args reaches llama-server. Do not spray HSA_OVERRIDE_GFX_VERSION. If the meter stays idle, come back to llama.cpp.
When to choose it
Only if the meter moves. Otherwise llama.cpp.
You want AMD’s server and you will read the backend log.