Function rpadding

generation_utils.py:329–345

(input_ids, channels, tokenizer)

def rpadding(input_ids, channels, tokenizer):
    attention_masks = [np.ones(inputs.shape[0]) for inputs in input_ids]
    max_length = max(ids.shape[0] for ids in input_ids)
    padded_input_ids, padded_attns = [], []

    for ids, attn in zip(input_ids, attention_masks):
        pad_len = max_length - ids.shape[0]
        input_pad = np.full((pad_len, channels), 1024)
        input_pad[:, 0] = tokenizer.pad_token_id
        padded_input_ids.append(np.concatenate([input_pad, ids]))
        attn_pad = np.zeros(pad_len)
        padded_attns.append(np.concatenate([attn_pad, attn]))

    input_ids = torch.tensor(np.stack(padded_input_ids))
    attention_mask = torch.tensor(np.stack(padded_attns))

    return input_ids, attention_mask


def find_max_valid_positions(C: torch.Tensor, invalid_value=1024) -> torch.Tensor: