Lemonade and ROCm integration in Ubuntu 26.04

Mario Limonciello and Ken VanDine explain the Ubuntu 26.04 packaging work behind Lemonade, then present applications built around local inference. The talk covers both the system changes needed to run models locally and the software that uses them.

Limonciello describes AMD Strix Halo systems configurable with 128 GB of shared memory. He explains how variable graphics memory reserves a region for the GPU, avoiding IOMMU translation overhead. On supported hardware, fwupd clients can change that BIOS setting from Linux for the next boot.

The speakers say Canonical and AMD packaged the ROCm stack in Ubuntu's archives. Lemonade's system backend uses the installed llama.cpp server and ROCm stack instead of downloading executable artifacts from GitHub. Its browser interface replaces the need for Electron in this setup, while a confined systemd service limits the server's access. The chat interface supports text, speech and images, and Lemonade also exposes an Ollama-compatible API.

VanDine's demos include Stable Diffusion image generation and an Omni model collection that connects text, vision and speech models. He describes agents running on his Strix Halo machine that sort community messages and maintain snaps through pull requests and visual tests. The snap workflow is incomplete and still tailored to his needs. A role-playing game using generated images is also a work in progress; VanDine says he has not personally tried its audio features.