Ggml v1.0.0

ggml 1.0.0 marks a cleanup and expansion of the tensor-computation library used by machine-learning workloads. It removes deprecated public interfaces and the legacy diagonal-mask operations, so projects relying on them need to update; the RPC protocol also advances to version 9, with faster loading and clearer failure behavior for unreachable servers or incompatible handshakes. New controls let applications configure YaRN truncation, and the release adds FP8 support and a backward pass for triangular inversion. Core scheduling is more selective: it avoids transferring inputs that an operation will not read and prefetches only data produced by earlier execution splits. The mixture-of-experts cache can also send some missed computations to the CPU asynchronously.

Much of the release focuses on faster and more capable hardware backends. CUDA adds low-row-count matrix multiplication paths, FP8 and fused NVFP4 operations, and broader FlashAttention configurations; Metal and SYCL improve mixture-of-experts execution, while Vulkan gains Intel FlashAttention prefill support and additional matrix and attention optimizations. OpenCL adds decomposed FlashAttention for Adreno and more quantized kernels; WebGPU gains batched tensor support and several operations; Hexagon adds quantization formats and convolution capabilities. CPU work includes improved ARM SIMD matrix multiplication and more precise FlashAttention accumulation. Across backends, the update also fixes memory, graph-reuse, compatibility, and determinism issues, making the release relevant to developers maintaining ggml integrations as well as those seeking improved inference performance on varied devices.