Download Latest Version v4.9.0 source code.zip (4.0 MB)
Home / v4.9.0
Name Modified Size InfoDownloads / Week
Parent folder
rocm-python-wheels-Linux.zip 2026-11-10 431.2 MB
rocm-python-wheels-Windows.zip 2026-11-10 194.0 MB
README.md 2026-11-10 1.0 kB
v4.9.0 source code.tar.gz 2026-11-10 3.7 MB
v4.9.0 source code.zip 2026-11-10 4.0 MB
Totals: 5 Items   632.9 MB 2

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
Source: README.md, updated 2026-11-10