2026-03 monthly summary for pytorch/pytorch: Focused on extending the AOTI eager compilation path with a pluggable compile backend registry and support for custom device backends, enabling extensibility without modifying core code and accelerating hardware backend onboarding.

February 2026: Delivered BMM out variant support in PyTorch meta registrations, enabling out-variant batched matrix multiplication and greater flexibility in tensor operations. The work expands the BMM API surface with additive, backwards-compatible changes, reducing workaround code for downstream users and paving the way for more efficient matmul workflows.

In July 2025, ROCm/pytorch delivered a focused code cleanup to streamline ExportedProgram output handling. The _postprocess_graph_module_outputs method was removed from the ExportedProgram class, eliminating an unnecessary hook that could lead to output-processing issues. This change reduces complexity, lowers risk of regression in export workflows, and improves long-term maintainability of the ExportedProgram module. The work was implemented as a single commit [exported_program] Remove _postprocess_graph_module_outputs (#158059) (dd93883231226576220a8ec0fa266c9e22eae82c). Overall business value: more reliable exports for downstream users and easier maintenance of the ROCm/pytorch codebase.