# Copyright 2022 Cerebras Systems.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
[docs]def construct_attention_mask(datadict, eos_id, pad_id, input_key='input_ids'):
"""
Constructs attention masks based on the provided input_key in the datadict.
The mask is constructed based on the presence of pad_id and eos_id.
"""
input_data = datadict[input_key]
attention_mask = []
for i in range(input_data.shape[0]):
pad_indices = np.where(input_data[i] == pad_id)[0]
if eos_id != pad_id:
# Handle case where eos_id and pad_id are different
non_pad_len = (
pad_indices[0] if len(pad_indices) > 0 else input_data.shape[1]
)
else:
# Handle case where eos_id is the same as pad_id
if len(pad_indices) > 0:
pad_idx = 0
while (
pad_idx + 1 < len(pad_indices)
and pad_indices[pad_idx] + 1 < input_data.shape[1]
and input_data[i][pad_indices[pad_idx] + 1] != pad_id
):
pad_idx += 1
if pad_idx == len(pad_indices) - 1:
# All eos, no pad
non_pad_len = input_data.shape[1]
else:
# Last eos just before pad, input_ids need to be chopped off
non_pad_len = pad_indices[pad_idx]
else:
non_pad_len = input_data.shape[1]
attention_mask.append(
[1] * non_pad_len + [0] * (input_data.shape[1] - non_pad_len)
)
return attention_mask