llama.cpp b9934: WebGPU flash attention tuning plus CUDA 13.3 prebuilts

llama.cpp build b9934 ships a single WebGPU change — tuning the subgroup split (d_split) in flash_attn_vec — alongside prebuilt binaries for 30+ platform configurations.

What’s in the build matrix:

• macOS: Apple Silicon (arm64), Intel (x64), iOS XCFramework. KleidiAI variant on Apple Silicon is disabled this release.
• Linux: Ubuntu CPU builds for x64, arm64, and s390x. GPU builds via Vulkan (x64, arm64), ROCm 7.2, OpenVINO, and SYCL (FP32 + FP16).
• Windows: CPU builds for x64 and arm64. GPU builds for CUDA 12 (12.4 DLLs), CUDA 13 (13.3 DLLs), Vulkan, OpenVINO, SYCL, and HIP. Plus OpenCL Adreno for arm64.
• Android: arm64 CPU only.
• openEuler: Ascend 310p and 910b (ACL Graph) for both x86 and aarch64. Base openEuler build is disabled.

The actual code delta is one PR — #25418, which tunes how WebGPU subgroups split during flash attention vector operations. That’s a performance optimization for browser-based inference paths, not a feature drop.

CUDA 13.3 DLLs in the Windows build matrix is the standout. Most local-LLM toolchains still target CUDA 12.x. If you’re running recent NVIDIA drivers, b9934 gives you a prebuilt binary that matches.

Full release artifacts and checksums are on the b9934 GitHub release page. Pick the binary that matches your exact backend — the build matrix is specific down to DLL versions.