pytorch

Breaking changes

• SVE compile checks in FindARM.cmake have been strengthened, so source builds with incorrect SVE configurations that previously passed may now fail at the CMake configuration stage.

Key features

• Added torch.accelerator.Graph API: unifies graph capture/replay across CUDA, XPU, and out-of-tree backends.
• torch.export.save now supports Microscaling (MX) quantization formats, enabling export of more heavily compressed models.
• Adagrad now supports fused=True, allowing it to use the single-kernel optimizer implementation alongside Adam, AdamW, and SGD.
• torch.cond control flow can now be captured/replayed inside CUDA Graphs.
• ROCm now supports expandable memory segments, rocSHMEM symmetric memory collectives, and FlexAttention pipelining.

Performance improvements

• Batched linalg.eigh on CUDA is up to 100x faster thanks to updated cuSolver backend selection.