# 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
import torch
[docs]def create_fixed_sparse_attention_mask(
max_sequence_length,
n_heads,
dtype=None,
local_attn_ctx=16,
num_verts=64,
vert_size=16,
different_layout_per_head=False,
):
"""
Create GPT-3 Fixed Sparse mask.
Adapted from https://github.com/openai/sparse_attention/blob/master/attention.py#L135
:param int max_sequence_length: Max sequence length.
:param dtype: Dtype of the resulting mask.
Returns:
The autoregressive fixed sparse mask of shape
[n_heads, max_sequence_length, max_sequence_length].
"""
n_ctx = max_sequence_length
assert n_heads % num_verts == 0
stride = local_attn_ctx
assert vert_size <= stride
assert stride % vert_size == 0
indices = [i for i in range(stride - 1, -1, -1)]
indices = np.array(indices).reshape([-1, vert_size])
if num_verts == 1:
layout = np.zeros([n_ctx, n_ctx])
for idx in indices[0]:
layout[:, idx::stride] = 1
for q_idx in range(n_ctx):
# Each thing can attend to its local block
row = q_idx // stride
layout[q_idx, row * stride : (row + 1) * stride] = 1
# Any query cannot attend to keys above it
layout[q_idx, q_idx + 1 :] = 0
else:
layouts = []
indices = indices[:num_verts]
for h in range(n_heads):
layout = np.zeros([n_ctx, n_ctx])
subindices = indices[h % num_verts]
for idx in subindices:
layout[:, idx::stride] = 1
for q_idx in range(n_ctx):
# Each position can attend to its local block
row = q_idx // stride
layout[q_idx, row * stride : (row + 1) * stride] = 1
# Any query cannot attend to keys above it
layout[q_idx, q_idx + 1 :] = 0
layouts.append(layout)
layout = np.array(layouts)
if not different_layout_per_head:
layout = layout[0, :, :]
# Swap 0s and 1s since we use 1 to indicate masked positions
mask = 1 - layout
fixed_sparse_attn_mask = torch.Tensor(mask).to(dtype)
return fixed_sparse_attn_mask