Function _prepare_inputs

examples/summarize.py:242–292

(batch_input_texts,
                        eval_task='summarize',
                        add_special_tokens=True,
                        min_input_length=0)

    ppls_hf = [[] for _ in range(num_sequences)]

    def _prepare_inputs(batch_input_texts,
                        eval_task='summarize',
                        add_special_tokens=True,
                        min_input_length=0):
        batch_size = len(batch_input_texts)
        append_str = ' TL;DR: ' if eval_task == 'summarize' else ''
        batch_input_ids = []
        for i in range(batch_size):
            curr_text = batch_input_texts[i] + append_str
            curr_text = curr_text.strip().replace(" n't", "n't")

            # TODO: The below lines are used to be compatible with the original code; may need fix
            if 'GLM' in model_name and model_version in ('chatglm2',
                                                         'chatglm3'):
                input_ids = tokenizer.encode(curr_text,
                                             return_tensors='pt').squeeze(0)
                input_ids = input_ids[:test_token_num]
            elif 'qwen' in model_name.lower() and model_version == 'qwen':
                # use make_content to generate prompt
                system_prompt = "You are a useful assistant, please directly output the corresponding summary according to the article entered by the user."
                _, input_id_list = make_context(
                    tokenizer=tokenizer,
                    query=curr_text,
                    history=[],
                    system=system_prompt,
                    max_input_length=test_token_num,
                )
                input_ids = torch.tensor(input_id_list)
            else:
                if 'qwen' in model_name.lower() and 'qwen2' in model_version:
                    messages = [{
                        "role":
                        "system",
                        "content":
                        "You are a helpful assistant, please summarize the article entered by the user with one or two sentences."
                    }, {
                        "role": "user",
                        "content": curr_text
                    }]
                    curr_text = tokenizer.apply_chat_template(
                        messages, tokenize=False, add_generation_prompt=True)
                input_ids = tokenizer.encode(
                    curr_text,
                    return_tensors='pt',
                    add_special_tokens=add_special_tokens,
                    truncation=True,
                    max_length=test_token_num).squeeze(0)

            if input_ids.numel() > min_input_length:
                batch_input_ids.append(input_ids)
        return batch_input_ids

    def eval_trt_llm(datapoint,
                     eval_task='summarize',