Lemonade Project Removed ROCm Support in Latest Release
The software update improves streaming performance on AMD hardware by shifting how the project manages model memory.
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 2026.39.1 and 2026.40-RC, bringing significant changes to model execution on AMD hardware. The update removes OpenMOSS ROCm support due to substantial performance lags compared to the Vulkan back-end.
Why it matters
The removal of the OpenMOSS ROCm back-end ensures users can access more stable performance when running complex streaming models on AMD Ryzen AI Max devices. By defaulting to the Vulkan back-end, the project resolves failures that previously prevented models like DeepSeek-V4-Flash from running.
The 2026.40-RC candidate introduces model streaming support on AMD APUs by sizing against the GTT pool rather than fixed vRAM carve-outs. Additionally, version 2026.39.1 introduced configurable vRAM auto-eviction.
The players
Lemonade
Lemonade is an open-source software project focused on optimizing the local execution of large language models.
AMD
AMD is a global semiconductor company that designs high-performance processors and graphics hardware for computing.
The details
Lemonade now optimizes model execution by dynamically sizing against the APU GTT pool, allowing for better hardware utilization on AMD systems. Developers made the decision to abandon the ROCm code path after tests confirmed it performed 40x slower than the Vulkan-based alternative.
Timeline
September 23, 2026: Lemonade versions 2026.39.1 and 2026.40-RC were officially released.
The Tech Race
The shift away from ROCm toward the Vulkan API highlights the ongoing effort to standardize AI model performance across heterogeneous hardware environments. By prioritizing cross-platform APIs, the project aims to replace niche hardware-specific implementations with more reliable software abstractions.
Users running streaming models on AMD Ryzen AI Max hardware will experience improved model stability and faster load times. Those who previously relied on the OpenMOSS ROCm path must transition to the Vulkan back-end to maintain software functionality.
The takeaway
Developers should prioritize mature, cross-platform APIs like Vulkan to ensure consistent performance across diverse hardware configurations. Users of local AI software should monitor release notes for backend deprecations that may affect hardware compatibility.
Further reading
For more insights on the latest developments in AI model optimization, visit our Artificial Intelligence section.
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?
