r"""Query Layer at the final layer.
| 107 | |
| 108 | |
| 109 | class QueryLayer(TransformerEncoderLayer): |
| 110 | r"""Query Layer at the final layer.""" |
| 111 | |
| 112 | def __init__( |
| 113 | self, |
| 114 | batch_size, |
| 115 | hidden_size, |
| 116 | ffn_hidden_size, |
| 117 | num_heads, |
| 118 | seq_length, |
| 119 | attention_dropout_rate=0.1, |
| 120 | hidden_dropout_rate=0.1, |
| 121 | post_layernorm_residual=False, |
| 122 | param_init_type=mstype.float32, |
| 123 | hidden_act="fast_gelu", |
| 124 | use_past=False, |
| 125 | parallel_config=None, |
| 126 | softmax_compute_type=mstype.float32, |
| 127 | ): |
| 128 | super(QueryLayer, self).__init__( |
| 129 | batch_size=batch_size, |
| 130 | hidden_size=hidden_size, |
| 131 | ffn_hidden_size=ffn_hidden_size, |
| 132 | num_heads=num_heads, |
| 133 | seq_length=seq_length, |
| 134 | attention_dropout_rate=attention_dropout_rate, |
| 135 | hidden_dropout_rate=hidden_dropout_rate, |
| 136 | post_layernorm_residual=post_layernorm_residual, |
| 137 | param_init_type=param_init_type, |
| 138 | hidden_act=hidden_act, |
| 139 | use_past=use_past, |
| 140 | parallel_config=parallel_config.dp_mp_config, |
| 141 | softmax_compute_type=softmax_compute_type, |
| 142 | ) |
| 143 | |
| 144 | def construct( |
| 145 | self, |
| 146 | x, |
| 147 | query_vector, |
| 148 | input_mask, |
| 149 | init_reset=True, |
| 150 | batch_valid_length=None, |
| 151 | ): |
| 152 | r""" |
| 153 | The forward process of the block. |
| 154 | """ |
| 155 | # [bs * seq_length, embedding_size] |
| 156 | input_x = self.layernorm1(x) |
| 157 | input_x = F.cast(input_x, self.dtype) |
| 158 | |
| 159 | # indicate whether reset saved states |
| 160 | key_reset = None |
| 161 | value_reset = None |
| 162 | |
| 163 | if self.use_past: |
| 164 | # reset states, init_reset True for reuse and False for reset |
| 165 | key_reset = self.assign( |
| 166 | self.key_past, |