Function eval_hf

examples/summarize.py:414–525

(datapoint,
                eval_task='summarize',
                eval_ppl=False,
                add_special_tokens=True,
                min_input_length=0)

        return [], [], [], {}

    def eval_hf(datapoint,
                eval_task='summarize',
                eval_ppl=False,
                add_special_tokens=True,
                min_input_length=0):
        batch_size = len(datapoint[dataset_input_key])
        if batch_size > 1:
            logger.warning(
                f"HF does not support batch_size > 1 to verify correctness due to padding. Current batch size is {batch_size}"
            )
        batch_input_ids = _prepare_inputs(datapoint[dataset_input_key],
                                          eval_task=eval_task,
                                          add_special_tokens=add_special_tokens,
                                          min_input_length=min_input_length)
        batch_size = len(batch_input_ids)
        if batch_size == 0:
            return [], [], [], [[] for _ in range(batch_size)]
        input_lengths = [x.size(0) for x in batch_input_ids]
        # Left padding for HF
        max_length = max(input_lengths)
        paddings = [
            torch.ones(max_length - l, dtype=torch.int32) * pad_id
            for l in input_lengths
        ]
        batch_input_ids = [
            torch.cat([pad, x]) for x, pad in zip(batch_input_ids, paddings)
        ]
        batch_input_ids = torch.stack(batch_input_ids)
        batch_input_ids = batch_input_ids.cuda()

        # specialization for HF
        if early_stopping in [0, 1]:
            local_early_stopping = bool(early_stopping)
        else:
            local_early_stopping = "never"

        with torch.no_grad():
            hf_config = {}
            if num_beams == 1:
                hf_config.update({
                    "top_k": top_k,
                    "top_p": top_p,
                    "do_sample": True,
                })
            else:
                hf_config.update({
                    "num_beams": num_beams,
                    "early_stopping": local_early_stopping,
                })

            outputs = model.generate(batch_input_ids,
                                     max_new_tokens=output_len,
                                     num_return_sequences=num_sequences,
                                     temperature=temperature,
                                     eos_token_id=end_id,
                                     pad_token_id=pad_id,
                                     length_penalty=length_penalty,
                                     output_scores=True,