Lemonade Fixes AMD APU Model Streaming, Drops OpenMOSS ROCm As ~40x Slower Than Vulkan

The AMD-aligned Lemonade open-source project for serving as a local AI server released 2026.39.1 today as well as issuing a release candidate of 2026.40 as their next release. Besides adapting to a new versioning scheme, the Lemonade SDK updates today bring a few notable changes. The Lemonade 2026.40 release candidate now supports streaming models on AMD integrated GPUs (APUs) by sizing against the APU GTT pool rather than the fixed vRAM carve-out. This came about as some streaming models like DeepSeek-V4-Flash-IQ2XXS-DS4 were failing to run on AMD Ryzen AI Max (Strix Halo). While there is sufficient RAM for doing so, Lemonade had been checking against the smaller carve-out window rather than the addressable GTT. Thus that's being fixed to allow for better streaming model support on AMD iGPUs/APUs. Another interesting change is the OpenMOSS back-end for ROCm being removed on Windows and Linux. OpenMOSS is an AI research initiative and model ecosystem out of China. The OpenMOSS ROCm back-ends are disabled since they are running around 40x slower than using the Vulkan back-end on the same hardware. It appears there are at least some CPU fall-backs happening with the AMD ROCm code paths for OpenMOSS and what's leading to the big disparity. This merge request took the step of removing OpenMOSS ROCm support until it can be properly fixed and shared the atrocious performance numbers for OpenMOSS ROCm. Lemonade 2026.40 RC is also adding a new launch agent for JetBrains' Junie. Plus other enhancements as outlined on GitHub. Lemonade 2026.39.1 released today brought configurable vRAM auto-eviction and other smaller changes. It's the first stable release under the new versioning system in relying on the year and week number.