{
  "multiquery_attention_c8": {
    "name": "multiquery_attention_c8",
    "function_name": "MultiQueryAppendC8Attention",
    "impl_file": "multiquery_attention_c8_impl.cuh",
    "template_params": [
      "T",
      "GROUP_SIZE",
      "HEAD_DIM",
      "BLOCK_SIZE",
      "CAUSAL",
      "BLOCK_SHAPE_Q",
      "NUM_WARP_Q",
      "OutT",
      "ENABLE_PREFILL",
      "IsFP8",
      "IsDynamicC8"
    ],
    "dispatch_params": {
      "GROUP_SIZE": [1, 2, 4, 5, 6, 7, 8, 12, 14, 16],
      "HEAD_DIM": [128],
      "BLOCK_SIZE": [64],
      "CAUSAL": [0, 1],
      "BLOCK_SHAPE_Q": [16, 32, 64, 128],
      "ENABLE_PREFILL": [0, 1],
      "IsFP8": [0, 1],
      "IsDynamicC8": [0, 1]
    },
    "data_types": [
      ["paddle::float16", "paddle::float16", "float16_float16"],
      ["paddle::float16", "paddle::float8_e4m3fn", "float16_fp8"],
      ["paddle::float16", "int8_t", "float16_int8"],
      ["paddle::bfloat16", "paddle::bfloat16", "bfloat16_bfloat16"],
      ["paddle::bfloat16", "paddle::float8_e4m3fn", "bfloat16_fp8"],
      ["paddle::bfloat16", "int8_t", "bfloat16_int8"]
    ],
    "max_instances_per_file": 80,
    "file_prefix": "multiquery_attention_c8_",
    "function_signature": "template void {function_name}{template_args}(\n    const AppendAttnMetaData &meta_data,\n    const paddle::Tensor &qkv,\n    const paddle::Tensor &cache_k,\n    const paddle::Tensor &cache_v,\n    const paddle::optional<paddle::Tensor> &attn_mask,\n    const paddle::Tensor &cache_k_scale,\n    const paddle::Tensor &cache_v_scale,\n    const paddle::optional<paddle::Tensor> &shift_bias,\n    const paddle::optional<paddle::Tensor> &smooth_weight,\n    const paddle::Tensor &seq_lens_q,\n    const paddle::Tensor &seq_lens_kv,\n    const paddle::Tensor &seq_lens_encoder,\n    const paddle::Tensor &batch_id_per_token,\n    const paddle::Tensor &cu_seqlens_q,\n    const paddle::Tensor &block_table,\n    const paddle::Tensor &batch_ids,\n    const paddle::Tensor &tile_ids_per_batch,\n    const int num_blocks_x_cpu,\n    const int max_seq_len,\n    const int max_dec_len,\n    const float quant_max_bound,\n    const float quant_min_bound,\n    const float in_scale,\n    const int max_partition_size,\n    const int encoder_max_partition_size,\n    const int speculate_max_draft_token_num,\n    const bool is_decoder,\n    cudaStream_t &stream,\n    paddle::Tensor *out);\n\n"
  },
  "multiquery_attention_c4": {
    "name": "multiquery_attention_c4",
    "function_name": "MultiQueryAppendC4Attention",
    "impl_file": "multiquery_attention_c4_impl.cuh",
    "template_params": [
      "T",
      "GROUP_SIZE",
      "HEAD_DIM",
      "BLOCK_SIZE",
      "CAUSAL",
      "BLOCK_SHAPE_Q",
      "NUM_WARP_Q",
      "OutT",
      "ENABLE_PREFILL"
    ],
    "dispatch_params": {
      "GROUP_SIZE": [1, 2, 4, 5, 6, 7, 8, 12, 14, 16],
      "HEAD_DIM": [128],
      "BLOCK_SIZE": [64],
      "CAUSAL": [0, 1],
      "BLOCK_SHAPE_Q": [16, 32, 64, 128],
      "ENABLE_PREFILL": [0, 1]
    },
    "data_types": [
      ["paddle::float16", "paddle::float16", "float16_float16"],
      ["paddle::float16", "paddle::float8_e4m3fn", "float16_fp8"],
      ["paddle::float16", "int8_t", "float16_int8"],
      ["paddle::bfloat16", "paddle::bfloat16", "bfloat16_bfloat16"],
      ["paddle::bfloat16", "paddle::float8_e4m3fn", "bfloat16_fp8"],
      ["paddle::bfloat16", "int8_t", "bfloat16_int8"]
    ],
    "max_instances_per_file": 160,
    "file_prefix": "multiquery_attention_c4_",
    "function_signature": "template void {function_name}{template_args}(\n    const AppendAttnMetaData &meta_data,\n    const paddle::Tensor &qkv,\n    const paddle::Tensor &cache_k,\n    const paddle::Tensor &cache_v,\n    const paddle::optional<paddle::Tensor> &attn_mask,\n    const paddle::Tensor &cache_k_scale,\n    const paddle::Tensor &cache_v_scale,\n    const paddle::optional<paddle::Tensor> &cache_k_zp,\n    const paddle::optional<paddle::Tensor> &cache_v_zp,\n    const paddle::optional<paddle::Tensor> &shift_bias,\n    const paddle::optional<paddle::Tensor> &smooth_weight,\n    const paddle::Tensor &seq_lens_q,\n    const paddle::Tensor &seq_lens_kv,\n    const paddle::Tensor &seq_lens_encoder,\n    const paddle::Tensor &batch_id_per_token,\n    const paddle::Tensor &cu_seqlens_q,\n    const paddle::Tensor &block_table,\n    const paddle::Tensor &batch_ids,\n    const paddle::Tensor &tile_ids_per_batch,\n    const int num_blocks_x_cpu,\n    const int max_seq_len,\n    const int max_dec_len,\n    const float quant_max_bound,\n    const float quant_min_bound,\n    const float in_scale,\n    const int max_partition_size,\n    const int encoder_max_partition_size,\n    const int speculate_max_draft_token_num,\n    const bool is_decoder,\n    cudaStream_t &stream,\n    paddle::Tensor *out);\n\n"
  },
  "multiquery_attention_c16": {
    "name": "multiquery_attention_c16",
    "function_name": "MultiQueryAppendAttention",
    "impl_file": "multiquery_attention_c16_impl.cuh",
    "template_params": [
      "T",
      "GROUP_SIZE",
      "HEAD_DIM",
      "BLOCK_SIZE",
      "CAUSAL",
      "BLOCK_SHAPE_Q",
      "NUM_WARP_Q",
      "OutT",
      "ENABLE_PREFILL"
    ],
    "dispatch_params": {
      "GROUP_SIZE": [1, 2, 4, 5, 6, 7, 8, 12, 14, 16],
      "HEAD_DIM": [128],
      "BLOCK_SIZE": [64],
      "CAUSAL": [0, 1],
      "BLOCK_SHAPE_Q": [16, 32, 64, 128],
      "ENABLE_PREFILL": [0, 1]
    },
    "data_types": [
      ["paddle::float16", "paddle::float16", "float16_float16"],
      ["paddle::float16", "paddle::float8_e4m3fn", "float16_fp8"],
      ["paddle::float16", "int8_t", "float16_int8"],
      ["paddle::bfloat16", "paddle::bfloat16", "bfloat16_bfloat16"],
      ["paddle::bfloat16", "paddle::float8_e4m3fn", "bfloat16_fp8"],
      ["paddle::bfloat16", "int8_t", "bfloat16_int8"]
    ],
    "max_instances_per_file": 160,
    "file_prefix": "multiquery_attention_c16_",
    "function_signature": "template void {function_name}{template_args}(\n    const AppendAttnMetaData &meta_data,\n    const paddle::Tensor &qkv,\n    const paddle::Tensor &cache_k,\n    const paddle::Tensor &cache_v,\n    const paddle::optional<paddle::Tensor> &attn_mask,\n    const paddle::optional<paddle::Tensor> &shift_bias,\n    const paddle::optional<paddle::Tensor> &smooth_weight,\n    const paddle::Tensor &seq_lens_q,\n    const paddle::Tensor &seq_lens_kv,\n    const paddle::Tensor &seq_lens_encoder,\n    const paddle::Tensor &batch_id_per_token,\n    const paddle::Tensor &cu_seqlens_q,\n    const paddle::Tensor &block_table,\n    const paddle::Tensor &batch_ids,\n    const paddle::Tensor &tile_ids_per_batch,\n    const int num_blocks_x_cpu,\n    const int max_seq_len,\n    const int max_dec_len,\n    const float quant_max_bound,\n    const float quant_min_bound,\n    const float in_scale,\n    const int max_partition_size,\n    const int encoder_max_partition_size,\n    const int speculate_max_draft_token_num,\n    const bool is_decoder,\n    cudaStream_t &stream,\n    paddle::Tensor *out);\n\n"
  },
  "multiquery_decoder_attention": {
    "name": "multiquery_decoder_attention",
    "function_name": "MultiQueryDecoderAttention",
    "impl_file": "multiquery_decoder_attention_impl.cuh",
    "template_params": [
      "T",
      "GROUP_SIZE",
      "HEAD_DIM_QK",
      "HEAD_DIM_V",
      "BLOCK_SIZE",
      "CAUSAL",
      "NUM_STAGE",
      "cache_bytes",
      "DEAL_EACH_TIME"
    ],
    "dispatch_params": {
      "GROUP_SIZE": [8, 16, 128],
      "HEAD_DIM_QK": [128, 192, 512, 576],
      "HEAD_DIM_V": [128, 192, 512, 576],
      "BLOCK_SIZE": [64],
      "CAUSAL": [0, 1],
      "NUM_STAGE": [2],
      "cache_bytes": [16],
      "DEAL_EACH_TIME": [32, 64]
    },
    "data_types": [
      ["paddle::float16", "", "float16"],
      ["paddle::bfloat16", "", "bfloat16"]
    ],
    "max_instances_per_file": 60,
    "file_prefix": "multiquery_decoder_attention_",
    "function_signature": "template void {function_name}{template_args}(\n    const AppendAttnMetaData& meta_data,\n    cudaStream_t &stream,\n    const paddle::Tensor &q,\n    const paddle::Tensor &cache_k,\n    const paddle::Tensor &cache_v,\n    const paddle::optional<paddle::Tensor>& attn_mask,\n    const paddle::optional<paddle::Tensor>& shift_bias,\n    const paddle::optional<paddle::Tensor>& smooth_weight,\n    const paddle::Tensor &seq_lens_q,\n    const paddle::Tensor &seq_lens_kv,\n    const paddle::Tensor &batch_id_per_token,\n    const paddle::Tensor &cu_seqlens_q,\n    const paddle::Tensor &block_table,\n    const int max_seq_len,\n    const int max_dec_len,\n    const float rope_scale,\n    const float rope_theta,\n    const float softmax_scale,\n    const float in_scale,\n    paddle::Tensor *out);\n\n"
  }
}