fused_ops.yaml

# This file is designed for fusion C++ farward operators, which manages the
# generated code for static mode and dynamic mode (when `support_dygraph_mode` is true).
# "support_dygraph_mode" is an extra configuration item in this file,
# if one operator have "support_dygraph_mode : true", it supports dygraph mode,
# otherwise the operator only could be used in static mode.

- op : add_act_xpu
  args : (Tensor x, Tensor x_max, Tensor y, Tensor y_max, int act_type)
  output : Tensor(out), Tensor(out_max)
  infer_meta :
    func : AddActXPUInferMeta
  kernel :
    func : add_act_xpu
    data_type : x
  optional : x_max, y_max

- op : add_layernorm_xpu
  args : (Tensor x, Tensor y, Tensor scale, Tensor bias, int64_t m, int64_t n, float epsilon)
  output : Tensor(out), Tensor(mean), Tensor(variance), Tensor(z_add)
  infer_meta :
    func : AddLayernormXPUInferMeta
  kernel :
    func : add_layernorm_xpu
    data_type : x

- op : conv2d_transpose_xpu
  args : (Tensor x, Tensor x_max, Tensor filter, Tensor filter_max, Tensor bias, int[] strides, int[] paddings, int[] output_padding, IntArray output_size, str padding_algorithm, int groups, int[] dilations, str data_format, bool has_bias, bool with_act, str act_type)
  output : Tensor(out), Tensor(out_max)
  infer_meta :
    func : Conv2dTransposeXPUInferMeta
  kernel :
    func : conv2d_transpose_xpu
    data_type : x
  optional : bias, x_max

- op : conv2d_xpu
  args : (Tensor x, Tensor x_max, Tensor filter, Tensor filter_max, Tensor bias, Tensor branch, Tensor branch_max, int[] paddings, int[] dilations, int[] strides, str padding_algorithm, int groups, bool has_bias, bool has_branch, int act_type, float act_param)
  output : Tensor(out), Tensor(out_max)
  infer_meta :
    func : Conv2dXPUInferMeta
  kernel :
    func : conv2d_xpu
    data_type : x
  optional : bias, branch, branch_max ,x_max

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

- 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, Tensor seed_tensor, Scalar p, bool is_test, str mode, int seed = 0, bool fix_seed = false)
  optional : seed_tensor
  output : Tensor(out), Tensor(seed_offset)
  infer_meta :
    func : FusedDropoutAddInferMeta
    param : [x, y]
  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, Tensor gather_index, 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, int gather_axis)
  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, gather_index

- op : fused_rotary_position_embedding
  args : (Tensor q, Tensor k, Tensor v)
  output : Tensor(out_q), Tensor(out_k), Tensor(out_v)
  infer_meta :
    func : FusedRopeInferMeta
  optional : k,v, out_k, out_v
  kernel :
    func : fused_rotary_position_embedding
    data_type : q
  backward: fused_rotary_position_embedding_grad
  support_dygraph_mode : true

- 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, Tensor seq_lod, Tensor max_seq_len, 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, seq_lod, max_seq_len, x_fp16, out_fp16

- op : yolo_box_xpu
  args : (Tensor x, Tensor x_max, Tensor grid, Tensor stride, Tensor anchor_grid, float offset)
  output : Tensor(out), Tensor(out_max)
  infer_meta :
    func : YoloBoxXPUInferMeta
  kernel :
    func : yolo_box_xpu
    data_type : x
  optional : x_max