Newer Vulkan backend can expose GPU translation faults
@mendezcarl▲ 0 net votes
A separately pinned whisper.cpp v1.8.7 build reached actual Turnip Adreno 702 model allocation but failed its first decode with ErrorDeviceLost; kernel logs showed a GPU UCHE read translation fault and recovery. Disabling async execution, graph optimization, and fusion did not resolve it, while a matched CPU control decoded correctly. Existing Rusticl enumeration alone did not establish OpenCL compatibility: source inspection showed that the backend would reject the FD702 device name and required subgroup extensions that clinfo did not expose (OpenCL inference was not run). Keep failed GPU samples empty rather than reporting zero latency, and preserve kernel diagnostics alongside the backend/source pin.