# This file is designed for fusion C++ farward operators, which manages the
# generated code for dynamic mode and static mode.
# The operators in the file have extra configuration item "support_dygraph_mode".
# If one operator have "support_dygraph_mode : True", it supports dygraph mode.

- op : embedding_with_eltwise_add_xpu
  args : (Tensor[] ids, Tensor[] tables, int64_t padding_idx)
  output: Tensor
  infer_meta :
    func: EmbeddingWithEltwiseAddXPUInferMeta
  kernel:
    func: embedding_with_eltwise_add_xpu
    data_type: tables

- op : fc_xpu
  args : (Tensor x, Tensor x_max, Tensor w, Tensor w_max, Tensor bias, int in_num_col_dims, bool transpose_x, float alpha, float beta, int act_type, float act_alpha)
  output : Tensor(out), Tensor(out_max)
  infer_meta :
    func : FcXPUInferMeta
  kernel :
    func : fc_xpu
    data_type : x
  optional : bias, x_max

- op : fused_dropout_add
  args : (Tensor x, Tensor y, Scalar p, bool is_test, str mode, int seed, bool fix_seed)
  output : Tensor(out), Tensor(seed_offset)
  infer_meta :
    func : FusedDropoutAddInferMeta
  kernel :
    func : fused_dropout_add
    data_type : x
  backward : fused_dropout_add_grad
  support_dygraph_mode : true

- op : fused_linear_param_grad_add
  args : (Tensor x, Tensor dout, Tensor dweight, Tensor dbias, bool multi_precision = true)
  output : Tensor(dweight_out), Tensor(dbias_out)
  infer_meta:
    func : FusedLinearParamGradAddInferMeta
  optional : dweight, dbias
  kernel:
    func : fused_linear_param_grad_add
    data_type : dout
  support_dygraph_mode : true

- op : fused_multi_transformer_xpu
  args : (Tensor x, Tensor[] ln_scale, Tensor[] ln_bias, Tensor[] qkvw, Tensor[] qkvw_max, Tensor[] qkv_bias, Tensor[] out_linear_w, Tensor[] out_linear_wmax, Tensor[] out_linear_bias, Tensor[] ffn_ln_scale, Tensor[] ffn_ln_bias, Tensor[] ffn1_weight, Tensor[] ffn1_weight_max, Tensor[] ffn1_bias, Tensor[] ffn2_weight, Tensor[] ffn2_weight_max, Tensor[] ffn2_bias, Tensor[] cache_kv, Tensor[] pre_caches, Tensor rotary_pos_emb, Tensor time_step, Tensor seq_lengths, Tensor src_mask, bool pre_layer_norm, int rotary_emb_dims, float epsilon, float dropout_rate, bool is_test, str dropout_implementation, str act_method, bool trans_qkvw, int ring_id)
  output : Tensor(out), Tensor[](cache_kv_out){out_linear_w.size()}
  infer_meta :
    func : FusedMultiTransformerXpuInferMeta
  kernel :
    func : fused_multi_transformer_xpu
    data_type : x
  optional : cache_kv, pre_caches, rotary_pos_emb, time_step, seq_lengths, src_mask

- op : generate_sequence_xpu
  args : (Tensor x, DataType dtype)
  output : Tensor
  infer_meta :
    func : GenerateSequenceXPUInferMeta
  kernel :
    func : generate_sequence_xpu
    data_type : dtype

- op : multi_encoder_xpu
  args : (Tensor x, Tensor[] fc_weight, Tensor[] fc_weight_max, Tensor[] fc_bias, Tensor[] ln_scale, Tensor[] ln_bias, Tensor mask, int layer_num, bool norm_before, int hidden_dim, int head_num, int size_per_head, int ffn_hidden_dim_scale, int act_type, int relative_type, int slice_idx)
  output : Tensor(out), Tensor(x_fp16), Tensor(out_fp16)
  infer_meta :
    func : MultiEncoderXPUInferMeta
  kernel :
    func : multi_encoder_xpu
    data_type : x
  optional : mask, x_fp16, out_fp16