-
1 · Cited, not measured on the local desktop
Linux install on Ubuntu 24.04+
Lemonade Ubuntu docs (checked 2026-10-14): tested on Ubuntu 24.04 LTS. Official PPA: sudo add-apt-repository ppa:lemonade-team/stable then sudo apt install lemonade-server. Optional desktop package: sudo apt install lemonade-desktop. Snap alternative: sudo snap install lemonade-server. Linux GUI is http://localhost:13305 once the service is up. Prove with lemonade --version and lemonade status. CLI defaults: --host 127.0.0.1 --port 13305 (env LEMONADE_HOST / LEMONADE_PORT). Ubuntu 24.04.5 is the measured AMDai llama.cpp OS; that is not a Lemonade test claim. Fedora and Arch have cited Lemonade packages on the v2026.41.1 notes and the AMD playbook (yay -S lemonade-server) — those paths are not verified on our rig. Keep the bind on 127.0.0.1 unless you mean the LAN. Not measured on the local desktop.
-
2 · Cited, not measured on the local desktop
Windows note — short, then back to Linux
Lemonade Windows docs (checked 2026-10-14): download the MSI from lemonade-server.ai and run it. Silent: msiexec /i lemonade-server-minimal.msi /qn. v2026.41.1 ships lemonade.msi and lemonade-server-minimal.msi. AMD playbook: the lemonade CLI lands on PATH; the server is expected to run in the background. Minimal MSI has no app — open http://127.0.0.1:13305. Official 35B cards on this site are Ubuntu llama.cpp, not Windows Lemonade. Vulkan llama.cpp remains the honest Windows first-token path on /setup/windows. Not measured on the local desktop.
-
3 · Cited, not measured on the local desktop
Pick Vulkan vs ROCm on Radeon — cite their gfx, not ours
Lemonade uses llama.cpp as the primary LLM backend. Install: lemonade backends install llamacpp:vulkan or lemonade backends install llamacpp:rocm. Per-run: lemonade run MODEL --llamacpp vulkan|rocm, or lemonade load MODEL --llamacpp vulkan. Persist: lemonade config set llamacpp.backend=rocm (or vulkan). Env LEMONADE_LLAMACPP=vulkan|rocm|cpu. Lemonade’s own llamacpp page (checked 2026-10-14): Vulkan is “recommended for most GPU users” on Windows/Linux, AMD iGPU and dGPU. ROCm is Windows/Linux for “AMD Radeon RX 6000/7000 series (RDNA2/RDNA3/RDNA4)” and Ryzen AI iGPUs (Strix Point/Halo). Their decision tree says try ROCm first on that RX 6000/7000 / Ryzen AI set, Vulkan if it fails; RX 5000 and earlier are Vulkan only (ROCm not supported). This site still picks Vulkan first on consumer Radeon — /compare/vulkan-vs-rocm and /rocm/vulkan-hip — and we do not invent a Lemonade speed delta. Nightly ROCm channel (lemonade config set rocm_channel=nightly then lemonade backends install llamacpp:rocm) publishes architecture-specific bins: gfx1150, gfx1151, gfx103X, gfx110X, gfx120X. Stable channel is “common builds for supported architectures” — we do not invent that the stable blob lists gfx1201 / R9700. gfx120X is the RDNA4 family on the nightly README (9070 / 9060 names); R9700 is gfx1201 in that family, not a Lemonade SKU we invent. Link /rocm/gfx. v2026.41.1 moves ROCm backends to ROCm 10; first ROCm load after upgrade downloads about 2.6 GB. Stable can reuse ROCM_PATH / /opt/rocm when the HIP runtime matches. CUDA is the wrong family on a Radeon. NPU / Hybrid / FLM recipes are a Ryzen AI XDNA 2 path on the AMD playbook — the R9700 has no NPU; do not treat those as a house-rig how-to. Not measured on the local desktop.
-
4 · Cited, not measured on the local desktop
Pull a GGUF — Q6, Q5, or Q4; Q3 is a labeled estimate
House family is Qwen3.6-35B-A3B. Lemonade CLI pull form (docs checked 2026-10-14) is owner/repo[:variant] from Hugging Face or ModelScope, or lemonade pull user.qwen36-house --checkpoint main owner/repo:variant --recipe llamacpp. House shape: lemonade pull unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q5_K_XL then lemonade load unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q5_K_XL --ctx-size 262144 --llamacpp vulkan. Confirm the variant stem on the Unsloth listing and lemonade list — we do not invent a Lemonade library tag. Prefer Qwen3.6-35B-A3B-UD-Q5_K_XL.gguf when the calculator says it fits. Start at Q6, Q5, or Q4. A cited Q3 file (Hugging Face listing for Qwen3.6-35B-A3B-UD-Q3_K_XL.gguf (16.8 GB)) is a labeled estimate. Q1/Q2 stay off. AMD’s playbook chat examples use other library names; stay on the house Qwen3.6 family already on /models. Decode numbers on this site are Ubuntu llama.cpp llama-bench, not a speed measured inside Lemonade.
-
5 · Cited, not measured on the local desktop
Stage models on fast NVMe, not a slow bulk mount
Lemonade config docs (checked 2026-10-14): on the systemd package, models cache under /var/lib/lemonade/.cache/huggingface (Debian/Ubuntu migrate from the older /opt/var/lib/lemonade path). Point that store — or the GGUF you already keep — at a fast NVMe directory. A USB or spinning bulk mount will stall pulls and swaps; we do not invent a GB/s. Size the file on the calculator before the pull. Not measured on the local desktop.
-
6 · Cited, not measured on the local desktop
Context and both RAM lines
Lemonade load/run take --ctx-size SIZE (docs default auto). House window is 262144: lemonade load unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q5_K_XL --ctx-size 262144 --llamacpp vulkan. Extra llama-server flags go through --llamacpp-args (example in their CLI: --flash-attn on --load-mode none). v2026.41.1 changelog: they renamed the deprecated --no-mmap option to --load-mode none for llama.cpp. Do not assume Lemonade defaults to f16 KV or -ngl 99 — those are house llama.cpp flags. The calculator still prints Minimum system RAM needed (rounded kit estimate with a breakdown — 192 or 256 GB when a 35B file spills) and RAM needed at full context when we have a measured llama.cpp peak (37.32 GB at 4K, 42.9 GB at 256K on UD-Q5). Those peaks are llama.cpp on the local desktop, not Lemonade. Label estimate unless measured.
-
7 · Cited, not measured on the local desktop
Point a harness at /api/v1
lemonade status should say the server is on port 13305. AMD playbook OpenAI-compatible base is http://127.0.0.1:13305/api/v1. Confirm with curl -s http://127.0.0.1:13305/api/v1/models. API key placeholder on that playbook is "lemonade" (env LEMONADE_API_KEY / LEMONADE_ADMIN_API_KEY). Point Hermes, Open WebUI, Aider, Cline, or Continue at that base — /harnesses/connect and /setup/coding-assistant. Note the /api/v1 suffix: llama-server and Ollama speak /v1. House Hermes on the local desktop talks to llama-server :8080, not a measured Lemonade attach. Keep 127.0.0.1 unless you mean the LAN. Do not present llama.cpp tok/s as a Lemonade result.
-
8 · Cited, not measured on the local desktop
Where Lemonade sits vs llama-server, Ollama, and LM Studio
Honest llama.cpp-first: plain llama-server you compiled is the measured daily driver (:8080 /v1, house flags, local desktop cards). Lemonade is AMD’s wrapper over that same engine — easier PPA/MSI, explicit lemonade backends install llamacpp:vulkan|rocm, a tray/web UI on :13305, and an OpenAI-shaped /api/v1. You give up matching compile flags and this site’s llama-bench cards. Versus Ollama: both are packaged llama.cpp. Ollama is :11434/v1 with library tags; Lemonade is :13305/api/v1 with owner/repo:variant pulls and a documented Vulkan vs ROCm switch. Versus LM Studio: both are desktop-ish wrappers. LM Studio is GUI-first on :1234/v1; Lemonade is CLI-first plus a web UI. None of the three wrappers is a measured local desktop tok/s path. We do not invent a speed ranking. Leave Lemonade when you need this site’s flags, recipes, or Stack Lab doctor.
-
9 · Cited, not measured on the local desktop
When it sits on CPU
lemonade status and lemonade logs first. GPU meter idle: Vulkan never installed (lemonade backends install llamacpp:vulkan), ROCm kernels missed this gfx, or you are on CPU. Do not spray HSA_OVERRIDE_GFX_VERSION — unofficial here, and the wrong gfx1201 fix. Missing nightly gfx → stay Vulkan or move to a llama.cpp you compiled. First ROCm load after ${SETUP_LEMONADE_RELEASE} pulling ~2.6 GB is expected, not a hang we invent. Harness 404 on /v1: you wanted /api/v1. Start from /troubleshoot. The wrapper is not the measured daily driver.