def _deduplicate_inputs()

in optimum/intel/openvino/modeling_decoder.py [0:0]

• 37 lines of code
• 13 McCabe index (conditional complexity)

    def _deduplicate_inputs(self, model_inputs: Dict):
        input_ids = model_inputs["input_ids"]
        upd_model_inputs = {}
        unique_input_ids, indicies, reverse_indicies = np.unique(
            input_ids, axis=0, return_index=True, return_inverse=True
        )
        export_transformers_version = get_export_transformers_version(self.model, self.config)
        for input_name, input_tensor in model_inputs.items():
            if input_name not in ["input_ids", "beam_idx"]:
                if input_name not in self.key_value_input_names:
                    upd_model_inputs[input_name] = input_tensor[indicies]
                else:
                    shape = input_tensor.shape if isinstance(input_tensor, Tensor) else list(input_tensor.shape)
                    dtype = input_tensor.element_type if isinstance(input_tensor, Tensor) else Type(input_tensor.dtype)
                    upd_batch_size = indicies.shape[0]
                    if self.config.model_type == "bloom" and compare_versions(
                        export_transformers_version, "<", "4.44"
                    ):
                        upd_batch_size *= self.config.num_attention_heads
                    shape[
                        (
                            0
                            if not (self.config.model_type == "chatglm" and not hasattr(self.config, "rope_ratio"))
                            else 1
                        )
                    ] = upd_batch_size
                    upd_model_inputs[input_name] = Tensor(dtype, shape)
        upd_model_inputs["input_ids"] = unique_input_ids
        if "beam_idx" in model_inputs:
            beam_range = (
                unique_input_ids.shape[0] * self.config.num_attention_heads
                if (self.config.model_type == "bloom" and compare_versions(export_transformers_version, "<", "4.44"))
                else unique_input_ids.shape[0]
            )
            beam_idx = np.arange(beam_range, dtype=int)
            upd_model_inputs["beam_idx"] = beam_idx
        return upd_model_inputs, reverse_indicies