Skip to main content

dequantize_q5_0

Function dequantize_q5_0 

Source
pub fn dequantize_q5_0(
    data: &[u8],
    start: usize,
    num_elements: usize,
) -> Result<Vec<f32>, AprenderError>
Expand description

Dequantize Q5_0 format to f32

Q5_0: blocks of 32 elements, each block has:

  • 2-byte f16 scale
  • 4 bytes of high bits (32 bits = 1 per element)
  • 16 bytes of low 4-bit values (32 values packed 2 per byte)

Total: 22 bytes per block