Function paged_decode_attn_reduce_fp8 

pub fn paged_decode_attn_reduce_fp8(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    workspace: DevicePtr,
    output: DevicePtr,
    seq_lens: DevicePtr,
    num_q_heads: u32,
    head_dim: u32,
    num_splits: u32,
    num_seqs: u32,
    stream: u64,
) -> Result<()>

Reduce split-K partials into final BF16 output (FP8 variant).

DTYPE-AGNOSTIC in everything but its name: the workspace is F32 [o[head_dim], m, l] whatever the KV cache held, so the BF16 twin’s reduce handle launches through this same wrapper (#928).

Grid: (num_q_heads, num_seqs, 1) Block: (32, 1, 1)