frink-quant 0.46.0

Quantized weight formats and fused dequant kernels for Frink
Documentation
1
2
3
4
5
6
# frink-quant

Quantized weight formats and fused dequant kernels for Frink.

Block layouts (Q4_K, Q8_0, …), repacking, and CPU fused dequant+dot used by
`frink-core` and the Metal/CUDA backends.