April 2026 monthly summary for the pytorch/pytorch repository focusing on stability and robustness improvements in ROCm AMDSMI integration. Delivered a targeted fix to prevent crashes when amdsmi is not installed but pynvml is present, improving reliability across ROCm environments.

March 2026 monthly summary for pytorch/pytorch focused on aligning CUDA behavior with the CPU path for the binomial distribution by adding dtype validation on CUDA and expanding tests. This ensures only floating-point tensors are accepted for both count and probability, with clear, user-friendly error messages. The work improves reliability, reduces support overhead, and provides a consistent developer experience across CPU and CUDA paths.

February 2026 performance summary for ROCm/PyTorch focusing on reliability, performance, and correctness on ROCm-enabled systems. Delivered the AOTriton integration for ROCm memory-efficient attention with environment-based configurability to pin AOTriton commits at build time. Added CI-friendly AOTRITON commit override to support testing specific versions. Implemented critical ROCm build and runtime hardening by fixing the ROCm preprocessor path for P2P connectivity detection and guarding the ROCm environment file, reducing flaky builds. Strengthened numerical correctness and stability with targeted fixes to NLLLoss backward for non-contiguous 4D inputs and to isclose broadcasting with equal_nan, addressing real-world data edge cases. These changes improve performance opportunities, build stability, and numerical reliability across ROCm-enabled PyTorch deployments.