functions:
  - cpp_name: "at_npu::native::atb::_npu_paged_attention_quant"
    impl_name: "atb::_npu_paged_attention_quant"
  - cpp_name: "at_npu::native::atb::npu_fused_add_topk_div"
    impl_name: "atb::npu_fused_add_topk_div"
  - cpp_name: "at_npu::native::atb::npu_fused_add_topk_div_out"
    impl_name: "atb::npu_fused_add_topk_div_out"
  - cpp_name: "at_npu::native::atb::_npu_group_topk"
    impl_name: "atb::_npu_group_topk"
  - cpp_name: "at_npu::native::atb::_npu_matmul_add_fp32"
    impl_name: "atb::_npu_matmul_add_fp32"
  - cpp_name: "at_npu::native::atb::npu_mla_preprocess"
    impl_name: "atb::npu_mla_preprocess"
  - cpp_name: "at_npu::native::atb::npu_mla_preprocess_out"
    impl_name: "atb::npu_mla_preprocess_out"
  - cpp_name: "at_npu::native::atb::npu_multi_head_latent_attention"
    impl_name: "atb::npu_multi_head_latent_attention"
  - cpp_name: "at_npu::native::atb::npu_multi_head_latent_attention_out"
    impl_name: "atb::npu_multi_head_latent_attention_out"
  - cpp_name: "at_npu::native::atb::npu_multi_head_latent_attention_lse"
    impl_name: "atb::npu_multi_head_latent_attention_lse"
  - cpp_name: "at_npu::native::atb::_npu_multi_head_latent_attention_lse_get_workspace"
    impl_name: "atb::_npu_multi_head_latent_attention_lse_get_workspace"
  - cpp_name: "at_npu::native::atb::npu_multi_head_latent_attention_lse_out"
    impl_name: "atb::npu_multi_head_latent_attention_lse_out"
  - cpp_name: "at_npu::native::atb::_npu_paged_attention"
    impl_name: "atb::_npu_paged_attention"
  - cpp_name: "at_npu::native::atb::_npu_paged_attention_get_workspace"
    impl_name: "atb::_npu_paged_attention_get_workspace"
  - cpp_name: "at_npu::native::atb::_npu_paged_attention_v2"
    impl_name: "atb::_npu_paged_attention_v2"
  - cpp_name: "at_npu::native::atb::_npu_paged_attention_v2_get_workspace"
    impl_name: "atb::_npu_paged_attention_v2_get_workspace"
  - cpp_name: "at_npu::native::atb::_npu_paged_attention_mla"
    impl_name: "atb::_npu_paged_attention_mla"
  - cpp_name: "at_npu::native::atb::_npu_paged_attention_splitfuse"
    impl_name: "atb::_npu_paged_attention_splitfuse"
  - cpp_name: "at_npu::native::atb::npu_paged_cache_load_out"
    impl_name: "atb::npu_paged_cache_load_out"
  - cpp_name: "at_npu::native::atb::npu_paged_cache_load"
    impl_name: "atb::npu_paged_cache_load"
  - cpp_name: "at_npu::native::atb::_npu_quantize_per_tensor"
    impl_name: "atb::_npu_quantize_per_tensor"
  - cpp_name: "at_npu::native::atb::_npu_quant_rms_norm"
    impl_name: "atb::_npu_quant_rms_norm"
  - cpp_name: "at_npu::native::atb::_npu_reshape_and_cache"
    impl_name: "atb::_npu_reshape_and_cache"
  - cpp_name: "at_npu::native::atb::_npu_reshape_and_cache_siso"
    impl_name: "atb::_npu_reshape_and_cache_siso"
  - cpp_name: "at_npu::native::atb::npu_ring_mla"
    impl_name: "atb::npu_ring_mla"
  - cpp_name: "at_npu::native::atb::npu_ring_mla_out"
    impl_name: "atb::npu_ring_mla_out"
  - cpp_name: "at_npu::native::atb::_npu_rotary_embedding"
    impl_name: "atb::_npu_rotary_embedding"
  - cpp_name: "at_npu::native::atb::_npu_flash_attention"
    impl_name: "atb::_npu_flash_attention"
  - cpp_name: "at_npu::native::atb::_npu_flash_attention_v2"
    impl_name: "atb::_npu_flash_attention_v2"
  - cpp_name: "at_npu::native::atb::_npu_flash_attention_prefix_v2"
    impl_name: "atb::_npu_flash_attention_prefix_v2"
  - cpp_name: "at_npu::native::atb::npu_self_attention_prefix_encoder"
    impl_name: "atb::npu_self_attention_prefix_encoder"
  - cpp_name: "at_npu::native::atb::npu_self_attention_prefix_encoder_out"
    impl_name: "atb::npu_self_attention_prefix_encoder_out"
  - cpp_name: "at_npu::native::atb::_npu_flash_attention_qlens"
    impl_name: "atb::_npu_flash_attention_qlens"
  - cpp_name: "at_npu::native::atb::_npu_flash_attention_unpad"
    impl_name: "atb::_npu_flash_attention_unpad"