Major version formalizes an API break
Ggml has released version 1.0.0 of its open-source tensor library, removing deprecated public interfaces and legacy DIAG_MASK operations while adding new model and hardware capabilities. The release is significant for developers who embed ggml or build inference tools on top of it because the cleanup can require source changes in software that still depends on the old API.
The project's release notes describe a broad engineering update rather than a single headline feature. At the library level, ggml now supports configurable YaRN truncation across several backends and adds FP8-related functionality. The change list includes OCP FP8 support on CUDA and FP8 key-value cache support, alongside tests and precision controls for matrix multiplication.
Removing obsolete interfaces gives the project a cleaner compatibility boundary for the 1.0 line. The accompanying commits show examples being adapted after removal of `ggml_diag_mask_inf`, illustrating the practical migration work downstream projects may face. The release also includes a NumPy requirement fix for Python 3.14 and multiple build, error-handling and memory-safety corrections.
Backend work targets varied hardware
A large share of the release focuses on accelerating inference across different devices. CUDA gains optimized quantized matrix multiplication for small batches, mixture-of-experts fusion work, additional FlashAttention configurations and graph-handling improvements. Metal changes include protections against out-of-bounds access in triangular-solve and FlashAttention kernels.
Vulkan receives a prefill FlashAttention path for Intel hardware, a batched tensor-copy interface and fixes for MoltenVK and FreeBSD builds. SYCL adds a device-side mixture-of-experts scheduler, caching improvements and corrections for large ahead-of-time builds and strided views. OpenCL adds decomposed FlashAttention for Adreno hardware and another matrix-multiplication operation, while WebGPU gains new kernels and reflected one-dimensional padding support.
The CPU and specialist backends also move forward. ARM NEON uses a new SIMD matrix-multiplication tile, Hexagon expands quantized operation coverage and fixes a 64-bit handle issue, and SpacemiT paths receive additional quantized-kernel dispatch. ROCm/HIP workspace handling was changed to avoid repeated allocation by rocBLAS.
Several scheduler and transport changes are aimed at avoiding unnecessary work. The backend can skip uploads for inputs that are not read, prefetch only inputs needed by later execution splits and improve loading over ggml's RPC mechanism. Other corrections make unreachable RPC endpoints fail cleanly and address device-fault reporting.
The range of changes reflects ggml's role as a portable low-level foundation: performance depends on many device-specific paths, while its public API must remain usable by higher-level applications. Developers adopting 1.0 should review removed interfaces first, then benchmark the backend relevant to their deployment rather than assume every listed optimization applies equally across hardware.
