fused_ops.yaml 4.2 KB
Newer Older
1
# This file is designed for fusion C++ farward operators, which manages the
Z
zyfncg 已提交
2
# generated code for static mode and dynamic mode (when `support_dygraph_mode` is true).
Z
zyfncg 已提交
3
# "support_dygraph_mode" is an extra configuration item in this file,
Z
zyfncg 已提交
4 5
# if one operator have "support_dygraph_mode : true", it supports dygraph mode,
# otherwise the operator only could be used in static mode.
6

W
wz1qqx 已提交
7 8 9 10 11 12 13 14 15 16
- 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

17
- op : conv2d_xpu
W
wz1qqx 已提交
18
  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)
19
  output : Tensor(out), Tensor(out_max)
20 21 22 23
  infer_meta :
    func : Conv2dXPUInferMeta
  kernel :
    func : conv2d_xpu
24
    data_type : x
W
wz1qqx 已提交
25
  optional : bias, branch, branch_max ,x_max
26

27
- op : embedding_with_eltwise_add_xpu
28 29
  args : (Tensor[] ids, Tensor[] tables, Tensor mask, int64_t padding_idx)
  output: Tensor(out), Tensor(seq_lod), Tensor(max_seq_len)
30 31 32 33 34
  infer_meta :
    func: EmbeddingWithEltwiseAddXPUInferMeta
  kernel:
    func: embedding_with_eltwise_add_xpu
    data_type: tables
35
  optional : mask, seq_lod, max_seq_len
36 37 38 39 40 41 42 43 44 45 46 47

- 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
48 49
  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
50 51 52
  output : Tensor(out), Tensor(seed_offset)
  infer_meta :
    func : FusedDropoutAddInferMeta
53
    param : [x, y]
54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71
  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
72
  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)
73 74 75 76 77 78
  output : Tensor(out), Tensor[](cache_kv_out){out_linear_w.size()}
  infer_meta :
    func : FusedMultiTransformerXpuInferMeta
  kernel :
    func : fused_multi_transformer_xpu
    data_type : x
79
  optional : cache_kv, pre_caches, rotary_pos_emb, time_step, seq_lengths, src_mask, gather_index
80 81 82 83 84 85 86 87 88 89 90

- 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
91
  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)
92 93 94 95 96 97
  output : Tensor(out), Tensor(x_fp16), Tensor(out_fp16)
  infer_meta :
    func : MultiEncoderXPUInferMeta
  kernel :
    func : multi_encoder_xpu
    data_type : x
98
  optional : mask, seq_lod, max_seq_len, x_fp16, out_fp16