Bugs:
- Unified quantization API: each quantization function now returns `Q, S` where `Q` is the quantized tensor and `S` the quantization state which may hold absolute max values, a quantization map or more. For dequantization all functions now accept the inputs `Q, S` so that `Q` is dequantized with the quantization state `S`.
- Fixed an issue where the CUDA 11.1 binary was not compiled with the right headers
API changes:
- Block-wise quantization for optimizers now enabled by default
Features:
- Block-wise quantization routines now support CPU Tensors.