March 2026 – PyTorch core: Delivered explicit BLAS backend selection via environment variables, enabling users to choose between cublas and rocblas for CUDA operations by treating TORCH_BLAS_PREFER_CUBLASLT and TORCH_BLAS_PREFER_HIPBLASLT as binary toggles. Updated and extended the testing framework to validate the new behavior, including test_preferred_blas_library_settings. PR 174377 merged with commit 5b1d1004262fa2a119c7815c702589305c5ce2dd. This work improves reproducibility and control across CUDA/HIP backends and lays groundwork for ROCm hipBLASLt defaulting under relevant architectures.

May 2025 monthly summary for graphcore/pytorch-fork focusing on ROCm performance optimization. Implemented hipblaslt support for gfx1200/gfx1201 by updating BlasBackend preferences, enabling improved GPU performance in ROCm environments. No major bug fixes recorded for this repository in May. Impact: faster PyTorch backend performance on ROCm-enabled AMD GPUs, improved platform alignment, and smoother adoption of ROCm optimizations in production pipelines. Technologies/skills demonstrated: ROCm, hipblaslt, gfx1200/gfx1201, BlasBackend configuration, commit-based development, performance-focused optimization.