Capture every activation and gradient of any PyTorch model — forward and backward — with automatic graph visualization, rich metadata, and live interventions. Works on any architecture, including dynamic and recurrent ones.
QR, symmetric eigendecomposition and SVD for batches of matrices on Apple Silicon GPUs, for MLX and PyTorch. Metal kernels per regime, routed to the GPU or the CPU by policies measured per Mac. C++, Python (pip: metal-linalg, metal-linalg-torch), Swift and C.
imbalanced-losses is a PyTorch library of training losses for class-imbalanced classification — including Focal Loss, Smooth-AP, and Recall-at-Quantile — with built-in DDP all-gather support for globally-correct rank estimation and normalization across multi-GPU training.