Method forward

codegeex/paddle/codegeex_model.py:513–560

(
        self,
        hidden_states,
        query_hidden_state,
        attention_mask,
        layer_past=None,
        get_key_value=False,
        prompt_length=None,
        context_length=None,
    )

        self.mlp = MLP(self.hidden_size)

    def forward(
        self,
        hidden_states,
        query_hidden_state,
        attention_mask,
        layer_past=None,
        get_key_value=False,
        prompt_length=None,
        context_length=None,
    ):
        # hidden_states: [b, s, h]
        # assert query_hidden_state != None

        # Use FP32 for Layernorm
        # layernorm_output = self.input_layernorm(hidden_states.cast("float32")).cast("float16")
        layernorm_output = self.input_layernorm(hidden_states)

        # Self attention.
        attention_output = self.attention(layernorm_output,
                                          query_hidden_state,
                                          attention_mask,
                                          layer_past=layer_past,
                                          get_key_value=get_key_value,
                                          prompt_length=prompt_length,
                                          context_length=context_length)

        if get_key_value:
            attention_output, presents = attention_output

        # Residual connection.
        residual = hidden_states
        layernorm_input = attention_output + residual
        
        # Use FP32 for Layernorm
        # layernorm_output = self.post_attention_layernorm(layernorm_input.cast("float32")).cast("float16")
        layernorm_output = self.post_attention_layernorm(layernorm_input)

        # MLP.
        mlp_output = self.mlp(layernorm_output)

        # Second residual connection.
        residual = layernorm_input
        output = mlp_output + residual

        if get_key_value:
            output = [output, presents]

        return output


class Transformer(paddle.nn.Layer):