v4.9.0

New features

• Add an experimental Apple Metal (MPS) backend for Apple Silicon GPUs, selected with device="mps" (#2077) by @TBO22. It is included in Apple Silicon Python wheels and can be enabled in source builds with -DWITH_MPS=ON; device="auto" continues to select CPU on Apple Silicon
• Include gfx1031, gfx1032, gfx1035, and gfx1103 targets in the ROCm wheels (#2142) by @jordimas

Fixes and improvements

• Scale rotary embeddings in float32 before the dtype conversion to preserve precision in FP16 and BF16 models (#2146) by @Kyozzz
• Replace the deprecated FindCUDA CMake module with native CMake CUDA support (#2067) by @gaborlukacs
• Bump the CI MKL to 2025.3 and reuse shared build scripts (#2145) by @jordimas
• Add performance and accuracy regression tests to CI (#2147) by @jordimas