Expand description
Affine Q4 groups: eight low-first nibbles per word, BF16 scale and bias.
Structs§
- Group 🔒
Functions§
- quantize 🔒
- Use MLX’s affine endpoint selection and ties-to-even rounding. Quantize with float parameters, then store them as BF16, as MLX does for BF16 input. Reference: mlx/backend/cpu/quantized.cpp, quantize<T, U>.