Function prefill_attention_fp8kv 

pub fn prefill_attention_fp8kv(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    q: DevicePtr,
    k_fp8: DevicePtr,
    v_fp8: DevicePtr,
    output: DevicePtr,
    seq_len: u32,
    batch: u32,
    num_q_heads: u32,
    num_kv_heads: u32,
    head_dim: u32,
    inv_sqrt_d: f32,
    causal: bool,
    stream: u64,
) -> Result<()>

Contiguous prefill Flash Attention — FP8 E4M3 K/V variant (BR=64).

Q is BF16, K/V are FP8 E4M3 (dequantized to BF16 in shared memory). Halves K/V memory reads compared to the BF16 kernel.

Grid: (num_q_heads, ceil(seq_len/64), batch) Block: (256, 1, 1)