Click a topic below to open the notes for that subject:

Data Engineering DevOps Email Kafka Kubernetes macOS Mongo DB Productivity Programming Python Powershell REST RDBMS SCM Security Spark Unix Tools

Lemonade

Lemonade Server installation

Lemonade server was installed via a Debian package (.deb).

Install method: dpkg / .deb package

  • Package: lemonade-server version 10.0.0 (amd64)
  • Source .deb file: /home/ojitha/Downloads/lemonade-server_10.0.0_amd64.deb
  • Binaries installed to: /usr/bin/lemonade-server, /usr/bin/lemonade-web-app
  • Also in: /opt/bin/ (lemonade-router, lemonade-server, lemonade-web-app)
  • Config: /etc/lemonade/lemonade.conf and secrets.conf
  • Systemd service: /usr/lib/systemd/system/lemonade-server.service
  • VS Code extension: lemonade-sdk.lemonade-sdk-0.0.7 also installed

To reinstall or upgrade, you can run: sudo dpkg -i
/home/ojitha/Downloads/lemonade-server_10.0.0_amd64.deb

Install FLM models in Lemonde Server

The model location for the Lemonade server is export HF_HOME=/opt/var/lib/lemonade/.cache/huggingface'

Check what FLM itself knows

flm list

Pull via FLM directly, example gemma3:4b:

flm pull gemma3:4b

Then register it with lemonade under the user namespace

lemonade-server pull user.Gemma3-4b-it-FLM \
  --checkpoint gemma3:4b \
  --recipe flm

The --checkpoint value should match whatever flm list shows as the model identifier (likely gemma3:4b based on the upstream registry metadata).

Verify FLM is actually present:

which flm
flm --version

The default location for the FLM models are ~/.config/flm/models/.

Ollama

Here is my current Ollama models:

ollama ls

image-20260523111937157

Ollama ROCm Setup on OJAI Machine

Hardware

Component Detail
Machine MINISFORUM AI X1 Pro Mini PC
CPU AMD Ryzen AI 9 HX 470 (Zen5, 12-core)
iGPU AMD Radeon 890M (gfx1150, RDNA 3.5)
NPU 86 TOPS (aie2p / RyzenAI-npu4)
OS Ubuntu 24.04.4
Kernel 6.17.0-1012-oem
ROCm 7.2.0
MIGraphX 2.15.0.dev+20250912

Problem 1 — Ollama Server Already Running

Attempting to start Ollama failed because port 11434 was already bound:

Error: listen tcp 127.0.0.1:11434: bind: address already in use

Note: ollama stop stops a model, not the server daemon — it requires a model name argument.

Fix

sudo fuser -k 11434/tcp

Problem 2 — Ollama Not Detecting ROCm / GPU

Running the debug check returned no output:

OLLAMA_DEBUG=1 ollama serve 2>&1 | grep -i "gpu\|rocm\|gfx"

This indicated Ollama was not recognising the gfx1150 GPU — it is too new for Ollama’s bundled ROCm to detect by default.

Fix — Override GFX Version

HSA_OVERRIDE_GFX_VERSION=11.5.0 ollama serve &

To make permanent via systemd

sudo systemctl edit ollama

Add:

[Service]
Environment="HSA_OVERRIDE_GFX_VERSION=11.5.0"

Then reload:

sudo systemctl daemon-reload && sudo systemctl restart ollama

Verification

GPU utilisation confirmed at 98% via rocm-smi --showuse:

GPU[0] : GPU use (%): 98

Ollama is fully offloading inference to the Radeon 890M via ROCm. ✅

Models Available

Model Size
gemma4:26b 17 GB
gemma4:e4b 9.6 GB

gemma4:31b (19 GB) was removed with ollama rm gemma4:31b.