(self,
input_ids: Tensor,
input_lengths=None,
position_ids=None,
token_type_ids=None,
hidden_states=None,
max_input_length=None,
prompt_embedding_table=None,
prompt_tasks=None,
prompt_vocab_size=None,
attention_mask=None,
lora_params: LoraParams = None,
language_adapter_routings: Optional[Tensor] = None)
| 714 | config.set_if_not_exist('residual_scaling', 1.0) |
| 715 | |
| 716 | def forward(self, |
| 717 | input_ids: Tensor, |
| 718 | input_lengths=None, |
| 719 | position_ids=None, |
| 720 | token_type_ids=None, |
| 721 | hidden_states=None, |
| 722 | max_input_length=None, |
| 723 | prompt_embedding_table=None, |
| 724 | prompt_tasks=None, |
| 725 | prompt_vocab_size=None, |
| 726 | attention_mask=None, |
| 727 | lora_params: LoraParams = None, |
| 728 | language_adapter_routings: Optional[Tensor] = None): |
| 729 | # In PP, layer 0 has ids as inputs, all other layers have hidden_states as inputs |
| 730 | if self.mapping.is_first_pp_rank(): |
| 731 | ptuning_args = [ |
| 732 | prompt_embedding_table, prompt_tasks, prompt_vocab_size |
| 733 | ] if prompt_embedding_table is not None else [] |
| 734 | |
| 735 | hidden_states = self.transformer.embedding(input_ids, position_ids, |
| 736 | token_type_ids, |
| 737 | *ptuning_args) |
| 738 | self.register_network_output('embedding_layer_output', |
| 739 | hidden_states) |
| 740 | else: |
| 741 | hidden_states = recv(hidden_states, self.mapping.prev_pp_rank()) |
| 742 | |
| 743 | for layer_idx, encoder_layer in enumerate(self.transformer.layers): |
| 744 | lora_layer_params = None |
| 745 | if lora_params is not None and lora_params.lora_ranks is not None: |
| 746 | lora_layer_params = lora_params.get_layer_params(layer_idx) |
| 747 | hidden_states = encoder_layer( |
| 748 | hidden_states=hidden_states, |
| 749 | attention_mask=attention_mask, |
| 750 | input_lengths=input_lengths, |
| 751 | max_input_length=max_input_length, |
| 752 | lora_layer_params=lora_layer_params, |
| 753 | language_adapter_routings=language_adapter_routings) |
| 754 | |
| 755 | if self.mapping.is_last_pp_rank(): |
| 756 | if self.has_model_final_layernorm: |
| 757 | hidden_states = self.transformer.ln_f(hidden_states) |
| 758 | hidden_states.mark_output('encoder_output', self._dtype) |
| 759 | else: |
| 760 | hidden_states = send(hidden_states, self.mapping.next_pp_rank()) |
| 761 | hidden_states.mark_output('hidden_states_output', self._dtype) |
| 762 | |
| 763 | return hidden_states |
| 764 | |
| 765 | def prepare_inputs(self, |
| 766 | max_batch_size, |
nothing calls this directly
no test coverage detected