llama.cpp fixes MiniCPM3 Metal and adds sparse Vulkan attention

Ten signed llama.cpp prereleases through b10985 fixed Android packaging, added CUDA 13.4 Windows builds, supplied a missing MiniCPM3 Metal flash-attention kernel, expanded OpenCL and OpenVINO paths, added sparse Vulkan flash attention for DeepSeek V4 and GLM while avoiding nondeterministic atomic addition, and limited RPC hash caching to weights.

What changed

Why it matters

The run combines concrete startup and correctness fixes with new accelerator paths. The MiniCPM3 kernel and Vulkan determinism change can affect whether a model runs and whether repeated inference is reproducible on the affected backends.

What to do

Canary b10985 beside llama.cpp 0.4.1 using the exact MiniCPM3, DeepSeek V4 or GLM model and backend you operate. Compare golden outputs across repeated runs, Android and Windows packaging, memory, latency and RPC behavior; keep the stable binary and configuration ready for rollback.

Caveat

Every build is a fast-moving prerelease and most changes target narrow hardware or model paths. The project artifacts establish code and signatures, not cross-device correctness or performance. b10988 and later were released after the India date boundary and are deferred to 16 September.