ggml-metal

Introduced chunked SSD MMA for Mamba-2 prefill optimization on Metal, improving performance by dropping the scalar SSD path and utilizing MMA with a sequential tail. Several constants and internal states have been refactored for clarity and efficiency.

This release includes improvements to ggml, specifically shortening virtual device naming in CUDA and Metal. It also updates the device description build process for ggml-metal and refines naming conventions.

Introduced a new FWHT kernel for the metal backend in ggml-metal and fixed a narrowing issue. This release also provides pre-compiled binaries for various platforms and configurations.

This release introduces core support for the NVFP4 quantization type, including fixes for subnormal value handling and significant performance optimizations for ARM NEON dot products. However, NVFP4 support has been temporarily removed from GPU backends and architecture-specific optimized dot products pending further review.

This release includes a fix for ggml-metal header copying during embedded builds and provides updated binary distributions across macOS, Linux, Windows, and openEuler platforms.

This release optimizes the Metal backend by adjusting the Flash Attention (FA) buffer size to prevent unnecessary memory reallocations.

This release introduces Metal kernel support for TRI, FILL, EXPM1, and SOFTPLUS operations while announcing a transition in Linux package formats from .zip to .tar.gz.