dequant_turbo3_block_to_bf16

Function dequant_turbo3_block_to_bf16 

Source
pub fn dequant_turbo3_block_to_bf16(
    raw: &[u8],
    bs: usize,
    nkv: usize,
    hd: usize,
    out: &mut [bf16],
)
Expand description

2026-09-25: A Turbo3 block to BF16.

Layout per block: data: bs * nkv * (hd * 3 / 8) bytes, 8 values in 3 bytes. scales: bs * nkv * (hd / NVFP4_GROUP_SIZE) bytes. Unpacking of v0..v7 from b0, b1, b2, as in nvfp4_dequant of kernels/gb10/common/paged_decode_attn_turbo3.cu: v0 = b0 & 0x7 v1 = (b0 >> 3) & 0x7 v2 = ((b0 >> 6) | (b1 << 2)) & 0x7 v3 = (b1 >> 1) & 0x7 v4 = (b1 >> 4) & 0x7 v5 = ((b1 >> 7) | (b2 << 1)) & 0x7 v6 = (b2 >> 2) & 0x7 v7 = (b2 >> 5) & 0x7