Lemonade shipped stable v2026.40.0 on September 30. On AMD integrated GPUs, streaming-model sizing now uses the GTT pool instead of the fixed VRAM carve-out. GTT is the shared pool an integrated GPU allocates from system memory.
Previously, device-type detection did not match, so integrated GPUs were also judged by the carve-out alone. Models were rejected because of a 1–4 GB carve-out even when a much larger GTT pool was reported.
After the fix, integrated GPUs use the larger of the two pools without adding them. Sizing for discrete GPUs did not change.
Caveats: ROCm backends removed in the same release
The same release temporarily removed openmoss:rocm and sd-cpp:rocm while bugs are being fixed. openmoss:rocm was removed on all OSes.
The sd-cpp:rocm removal applies only to Windows, and Linux keeps ROCm. On Windows, sd-cpp configs with backend rocm no longer resolve, and the PR points to vulkan or auto.
The PR author reported a successful load on a Strix Halo machine, but that is one machine’s report. The sizing fix alone does not confirm load success, speed or usable context on other APUs.
A check order that separates sizing failures from missing backends
Before upgrading, record the engine, the backend and the reported memory pools. After upgrading, see how eligibility changed, then confirm actual loading and inference separately.
A model being rejected and a backend disappearing are different failure paths.
This release did not add memory; it started counting the GTT pool that was already there.