Lemonade Project Removed OpenMOSS ROCm Support
The release candidates for version 2026.40 drop AMD ROCm support due to significant performance regressions.
Updated on Sept. 23, 2026 in Artificial Intelligence
Live Poll
Do you trust that open-source software updates will reliably improve performance on your hardware?
The Lemonade project has released versions 2026.40-RC and 2026.39.1, officially removing the OpenMOSS ROCm back-end for Windows and Linux. This change addresses severe performance discrepancies where the ROCm path was measured to be 40x slower than the Vulkan alternative.
Why it matters
The removal of the ROCm back-end targets critical stability issues that prevented large streaming models from executing on AMD Ryzen AI Max hardware. By shifting focus toward more performant back-ends, the project aims to ensure reliability for resource-intensive inference tasks.
The OpenMOSS ROCm back-end was removed due to a 40x performance disparity compared to the Vulkan back-end. Version 2026.40-RC now enables streaming models on AMD APUs by dynamically sizing against the Graphics Translation Table pool rather than a fixed vRAM carve-out.
The players
Lemonade
An open-source software project focused on optimizing and running machine learning models.
OpenMOSS
A China-based initiative focused on AI software architecture and hardware abstraction.
JetBrains
A developer tools company that provides integrated development environments and now integrates with the Lemonade launch agent.
The details
The project replaced a fixed vRAM (video random access memory) carve-out—a static allocation of system memory for graphics tasks—with dynamic sizing against the GTT (Graphics Translation Table) pool. The GTT acts as a memory mapping layer that allows the integrated graphics processor to access system memory efficiently. This adjustment was necessary because streaming models such as DeepSeek-V4-Flash-IQ2XXS-DS4 were failing to execute correctly on AMD Ryzen AI Max chips.
Timeline
September 23, 2026: Lemonade released version 2026.39.1 and version 2026.40-RC.
The Tech Race
This development tracks the ongoing optimization lifecycle for the AMD Ryzen AI Max hardware platform. It follows a pattern of consolidating support around stable, high-performance APIs like Vulkan to ensure model compatibility across diverse APU architectures.
Users running streaming models on AMD APUs should update to the 2026.40-RC or 2026.39.1 releases to benefit from the new GTT-based memory management. Those previously reliant on ROCm will need to migrate to the Vulkan back-end to maintain hardware acceleration performance.
The takeaway
The move to remove ROCm support underscores the importance of prioritizing hardware back-ends that deliver consistent throughput for large language models. Watch for future merge requests from the project to see if alternative optimizations for AMD hardware are introduced.
Further reading
For broader technical context on model optimization, visit Artificial Intelligence.
Source note: This article includes information reported by Phoronix.
Live Poll
Do you trust that open-source software updates will reliably improve performance on your hardware?
