Lemonade Fixes AMD APU Model Streaming, Drops OpenMOSS ROCm As ~40x Slower Than Vulkan
Lemonade 2026.40 RC fixes APU streaming by using GTT pool and removes OpenMOSS ROCm due to ~40x performance gap versus Vulkan.Summary
The AMD-aligned Lemonade open-source local AI serving project released version 2026.39.1 and a release candidate for 2026.40. The 2026.40 RC addresses streaming model failures on AMD integrated GPUs (APUs) by sizing against the APU GTT pool instead of the fixed vRAM carve-out, fixing issues seen with models like DeepSeek-V4-Flash-IQ2XXS-DS4 on AMD Ryzen AI Max (Strix Halo). The RC also removes the OpenMOSS ROCm back-end on Windows and Linux because it runs approximately 40 times slower than the Vulkan back-end on the same hardware, likely due to CPU fall-backs in the AMD ROCm code paths for OpenMOSS. Additionally, the release adds a new launch agent for JetBrains' Junie. The stable 2026.39.1 release introduced configurable vRAM auto-eviction and adopted a new year-and-week versioning scheme.
(Source：Phoronix)