Function paged_decode_attn_fp8k_turbo3v 

pub fn paged_decode_attn_fp8k_turbo3v(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    q: DevicePtr,
    k_cache: DevicePtr,
    v_cache: DevicePtr,
    output: DevicePtr,
    block_tables: DevicePtr,
    seq_lens: DevicePtr,
    max_blocks_per_seq: u32,
    num_seqs: u32,
    num_q_heads: u32,
    num_kv_heads: u32,
    head_dim: u32,
    block_size: u32,
    inv_sqrt_d: f32,
    k_scale: f32,
    q_stride: u32,
    v_block_stride_bytes: u64,
    v_data_section_bytes: u64,
    sliding_window: u32,
    stream: u64,
) -> Result<()>

Paged decode attention for Fp8K + Turbo3V asymmetric KV cache.

K is read as FP8 NHD with per-tensor k_scale dequant, V as 3-bit Lloyd-Max packed bytes + FP8 per-group scale (sparse-V threshold on batched + remainder paths).

Kernel: paged_decode_attn_fp8k_turbo3v(Q, K_cache, V_cache, O, block_tables, seq_lens, max_blocks_per_seq, num_q_heads, num_kv_heads, head_dim, block_size, inv_sqrt_d, k_scale, q_stride, v_block_stride_bytes, v_data_section_bytes, sliding_window) Grid: (num_q_heads, num_seqs, 1) Block: (256, 1, 1)