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<()>

2026-09-25: GQA-packed BF16-KV paged decode, one block per (kv_head, seq). Same arguments, argument order and block as paged_decode_attn_bf16; only the grid’s x extent changes, to num_kv_heads, because the kernel (kernels/gb10/common/paged_decode_attn_bf16_gqa.cu) keeps a KV head’s whole query group in registers.

Preconditions: num_q_heads == num_kv_heads * DECODE_GQA_PACK_WIDTH and head_dim == DECODE_GQA_PACK_HEAD_DIM. The kernel derives its heads as kv_head * PD_GQA + h and sizes its register arrays from PD_GQA. The caller checks both with metrale_kernels::attn_splitk::gqa_pack_shape_ok and uses the unpacked kernel when they fail.