pub fn reshape_and_cache_bf16k_turbo4v(
gpu: &dyn GpuBackend,
kernel: KernelHandle,
key: DevicePtr,
value: DevicePtr,
k_cache: DevicePtr,
v_cache: DevicePtr,
slot_mapping: DevicePtr,
num_tokens: u32,
num_kv_heads: u32,
head_dim: u32,
block_size: u32,
key_stride: u32,
value_stride: u32,
k_block_stride_bytes: u64,
v_block_stride_bytes: u64,
v_data_section_bytes: u64,
stream: u64,
) -> Result<()>Expand description
Write K/V to paged Bf16K + Turbo4V (TurboQuant+ safer-asym) cache.
K written as raw BF16 NHD, V as 4-bit Lloyd-Max + FP8 per-group scale with matched-norm correction. K and V pools have separate strides.