LLM/Ryzen AI 7 350/2026-07-09

Install lemonade from PPA

sudo add-apt-repository ppa:lemonade-team/stable
sudo apt update
sudo apt install lemonade-server

Run lemonade server

sudo systemctl start lemond
sudo systemctl enable lemond

Install Backends

Open http://localhost:13305, then ...

Install llama.cpp cuda

• Click backends at left sidebar.
• Find Llama.cpp GPU
• Find child item cuda
• Click download

sudo vim /etc/gdm3/custom.conf

[daemon]
# 強制 GDM 的 Wayland 渲染走 AMD 的基底
Environment=VK_LOADER_DRIVERS_SELECT=*radeon*
Environment=MESA_VK_DEVICE_SELECT=pci-0000_66_00_0

sudo vim /etc/default/grub

GRUB_CMDLINE_LINUX_DEFAULT="quiet splash nvidia-drm.modeset=0"

sudo update-grub

Install Fastflow

• Click backends at left sidebar.
• Find FastFlowLM NPU
• Find child item npu
• Click download
• Install XDNA driver

sudo apt install amdxdna-dkms
sudo reboot

• Get FastFlowLM release from https://github.com/FastFlowLM/FastFlowLM/releases/
• Download the latest db for Ubuntu 26.04

cd 下載
sudo apt install ./fastflowlm_0.9.44_ubuntu26.04_amd64.deb
flm --version
flm validate

Memlock limit is too low (8MB). Please ...

Fix memory limit problem

sudo nano /etc/security/limits.conf

Add before `# End of file`

* soft memlock unlimited
* hard memlock unlimited

sudo reboot

ulimit -l
flm validate

Enable NPU and set default models

sudo systemctl edit lemond

[Service]
# Enable NPU
Environment="LEMONADE_ENABLE_NPU=true"
# Prevent from model unloaded
Environment="LEMONADE_MODEL_TTL=0"
# Default models
ExecStartPost=/usr/bin/lemonade load model --ctx-size 65535 Gemma-4-12B-it-GGUF --backend cuda
ExecStartPost=/usr/bin/lemonade load model gpt-oss-sg-20b-FLM --backend npu

sudo systemctl daemon-reload
sudo systemctl restart lemond

Let iGPU can use large memory size

Default iGPU memory limit is 536.87MB in Ubuntu 26.04

sudo nano /etc/modprobe.d/amdgpu.conf

options amdgpu ttm_pages_limit=25600000
options amdgpu page_pool_size=25600000

sudo update-initramfs -u
sudo reboot
sudo dmesg | grep -iE "amdgpu.*memory"

[    5.304706] amdgpu 0000:66:00.0: Trusted Memory Zone (TMZ) feature disabled as experimental (default)
[    5.304936] amdgpu 0000:66:00.0:  512M of VRAM memory ready
[    5.304939] amdgpu 0000:66:00.0:  46677M of GTT memory ready.

sudo apt install vulkan-tools sudo vim /var/lib/lemonade/.cache/lemonade/config.json

Optimize Lemonade config

Find Cache dir

sudo journalctl -u lemond -f -n 100

 7月 09 17:37:22 agentlab-zen5 lemond[411649]: 2026-07-09 17:37:22.869 [Info] (main) Starting Lemonade Server...
 7月 09 17:37:22 agentlab-zen5 lemond[411649]: 2026-07-09 17:37:22.869 [Info] (main)   Version: 10.9.0
 7月 09 17:37:22 agentlab-zen5 lemond[411649]: 2026-07-09 17:37:22.869 [Info] (main)   Cache dir: /var/lib/lemonade/.cache/lemonade
 7月 09 17:37:22 agentlab-zen5 lemond[411649]: 2026-07-09 17:37:22.869 [Info] (main)   Port: 13305
 7月 09 17:37:22 agentlab-zen5 lemond[411649]: 2026-07-09 17:37:22.869 [Info] (main)   Host: localhost
 7月 09 17:37:22 agentlab-zen5 lemond[411649]: 2026-07-09 17:37:22.869 [Info] (main)   Log level: info
 7月 09 17:37:22 agentlab-zen5 lemond[411649]: 2026-07-09 17:37:22.869 [Info] (main)   Telemetry: disabled

Edit config.json

sudo vim /var/lib/lemonade/.cache/lemonade/config.json

• host: 0.0.0.0
• max_loaded_models: -1

sudo systemctl restart lemond

Now Lemonade can load multiple models.

Debug

sudo journalctl -u lemond.service -n 50 --no-pager
sudo journalctl -u lemond -f -n 100