Function paged_decode_attn_bf16_gqa 

pub fn paged_decode_attn_bf16_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,
    q_stride: u32,
    sliding_window: u32,
    stream: u64,
) -> Result<()>

GQA-PACKED BF16-KV paged decode: one CTA per (kv_head, seq).

Same arguments, same kernel ABI order and same block as paged_decode_attn_bf16 — only the grid’s x extent changes, from num_q_heads to num_kv_heads, because the kernel (kernels/gb10/common/paged_decode_attn_bf16_gqa.cu) carries the whole query group of a KV head in registers and reads each K and V row once for all of them.

⚠️ num_q_heads == num_kv_heads * DECODE_GQA_PACK_WIDTH and head_dim == DECODE_GQA_PACK_HEAD_DIM are hard preconditions — the kernel recovers its heads as kv_head * PD_GQA + h and sizes its register arrays from PD_GQA. The caller checks both with avarok_kernels::attn_splitk::gqa_pack_shape_ok and keeps the unpacked kernel when they do not hold.