Function paged_decode_attn_fp8_gqa 

pub fn paged_decode_attn_fp8_gqa(
    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,
    v_scale: f32,
    q_stride: u32,
    cache_stride: u64,
    sliding_window: u32,
    stream: u64,
) -> Result<()>

GQA-PACKED FP8-KV paged decode: one CTA per (kv_head, seq).

The FP8 twin of paged_decode_attn_bf16_gqa — same argument list and same ABI order as paged_decode_attn_fp8, grid x extent num_kv_heads. The same two preconditions apply, checked by the same predicate.