llama.cpp fixes GDN, CUDA and Vulkan correctness across 12 builds

Twelve signed llama.cpp prereleases landed during 7 September IST. Beyond the four early builds, later tags fix a divergent CUDA barrier in f16 flash attention and a Vulkan GET_ROWS crash on misaligned views, add Vulkan TQ1_0 and gfx90c HIP support, and tune Vulkan RMS normalization and CUDA Q4_K/Q5_K paths.

What changed

Why it matters

The completed date set adds two correctness fixes that can prevent unsafe CUDA synchronization and hard Vulkan crashes, while the wider backend changes affect performance and compatibility without changing model weights.

What to do

If an affected CUDA, Vulkan, OpenCL, HIP, GDN or conversion path is in service, pin b10842—the last release inside the IST date—in a disposable canary. Compare golden outputs, crashes, memory, throughput, quantized accuracy and rollback with the current production build.

Caveat

All twelve tags are prereleases and results are project-reported. The approximately 4% Gemma 4 Vulkan gain covers one contributor's system, and later b10844 and b10850 tags fall on 8 September in India rather than this edition.