pub fn dequant_4bit_block_to_bf16(
raw: &[u8],
bs: usize,
nkv: usize,
hd: usize,
lut: &[f32; 16],
out: &mut [bf16],
)Expand description
Dequant a 4-bit packed (NVFP4 or Turbo4) KV block to BF16.
Layout per block:
data: bs * nkv * (hd / 2) bytes (2 nibbles/byte).
scales: bs * nkv * (hd / NVFP4_GROUP_SIZE) bytes (1 FP8 scale per group).