Function prefill_attention_paged 

pub fn prefill_attention_paged(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    q: DevicePtr,
    k_cache: DevicePtr,
    v_cache: DevicePtr,
    output: DevicePtr,
    block_table: DevicePtr,
    q_len: u32,
    kv_len: u32,
    q_offset: u32,
    num_q_heads: u32,
    num_kv_heads: u32,
    head_dim: u32,
    cache_block_size: u32,
    sliding_window: u32,
    inv_sqrt_d: f32,
    stream: u64,
) -> Result<()>

Paged prefill Flash Attention — reads K/V from paged KV cache via block_table.

For chunked prefill chunk 1+: Q comes from GEMM (contiguous), K/V reside in the paged cache from prior chunks. Replaces per-token paged decode loop with a single Flash Attention pass (O(N) per chunk instead of O(N^2) total).

Kernel: inferspark_prefill_paged(Q, K_cache, V_cache, O, block_table, q_len, kv_len, q_offset, nq, nkv, hd, block_size, inv_sqrt_d) Grid: (num_q_heads, ceil(q_len/32), 1) Block: (128, 1, 1)