- Ampere, RTX 30 series GPUs now compatible with the library.
v0.0.22:
- Fixed an error where a `reset_parameters()` call on the `StableEmbedding` would lead to an error in older PyTorch versions (from 1.7.0).
v0.0.23:
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.
v0.0.24:
- Fixed a bug where a float/half conversion led to a compilation error for CUDA 11.1 on Turning GPUs.
