* [optim] 32x32 f32 AMX kernel (for Apple Silicon M family) * [optim] bunch of AMX512F kernels (square, skinny, vector) * [ONNX] introduce Trilu, TopK * [NNEF/OPL] submodel loader * [ONNX] support alternative layout for LSTM (layout=1, batch becomes first axis) * [ONNX] If operators with dynamic condition (very basic optimisations, no nnef support yet).
0.19.9
* HardSwiwh ONNX, tract_core_hard_swish in NNEF/OPL * introducing tract_core_submodel in NNEF/OPL * JSON resource loader in NNEF/OPL * Profiling API tweaks * `--folded` view for model command line dump (hide Scan loops)
0.19.8
* Various bug fixes
0.19.7
* more bug fixes * wip on python doc auto-deploy
0.19.5
0.19.3
* lots of bugfixes following 0.19 big changes * introducing the JSON NNEF resource