AMDGPU.jl versions 2.6 and 2.7 bring significant expansions to GPU-accelerated linear algebra on AMD hardware. New features include GPU Cholesky, LU, and SVD factorizations via rocSOLVER (with both QR iteration and Jacobi algorithms), mixed-precision matrix multiplication, Hermitian support, and a sparse array interface backed by rocSPARSE supporting CSR, CSC, and COO formats. RDNA4 GPUs gain WMMA matrix-core support alongside existing RDNA3 support. The rocFFT path was redesigned for more predictable behavior. Load time improvements come from a new precompilation workload and moving SpecialFunctions into a package extension. Toolchain support now covers Julia 1.10–1.13 and LLVM up to 21.1. Documentation has been substantially expanded with usage guides, library references, and an FAQ.
