# sparse_attention¶

paddle.nn.functional. sparse_attention ( query, key, value, sparse_csr_offset, sparse_csr_columns, name=None ) [source]

This operator sparsify the Attention matrix in Transformer module to achieve the effect of reducing memory consumption and computation. The sparse layout is expressed in CSR format and contains two parameters, offset and columns.

$result=softmax(\frac{ Q * K^T }{\sqrt{d}}) * V$

where : Q, K, and V represent the three input parameters of the attention module. The dimensions of the three parameters are the same. d represents the size of the last dimension of the three parameters.

Parameters
• query (Tensor) – The query tensor in the Attention module. It’s a 4-D tensor with a shape of $$[batch\_size, num\_heads, seq\_len, head\_dim]$$. The dtype can be float32 and float64.

• key (Tensor) – The key tensor in the Attention module. It’s a 4-D tensor with a shape of $$[batch\_size, num\_heads, seq\_len, head\_dim]$$. The dtype can be float32 and float64.

• value (Tensor) – The value tensor in the Attention module. It’s a 4-D tensor with a shape of $$[batch\_size, num\_heads, seq\_len, head\_dim]$$. The dtype can be float32 and float64.

• sparse_csr_offset (Tensor) – The sparsity feature in the Attention module is expressed in the CSR format, and the offset represents the number of non-zero elements in each row of the matrix. It’s a 3-D tensor with a shape of $$[batch\_size, num\_heads, seq\_len + 1]$$. The dtype should be int32.

• sparse_csr_columns (Tensor) – The sparsity feature in the Attention module is expressed in the CSR format, and the columns represent the column index values of non-zero elements in the matrix. It’s a 3-D tensor with a shape of $$[batch\_size, num\_heads, sparse\_nnz]$$. The dtype should be int32.

• name (str, optional) – The default value is None. Normally there is no need for user to set this property. For more information, please refer to Name.

Returns

A Tensor which refers to the result in the Attention module. It’s a 4-D tensor with a shape of $$[batch\_size, num\_heads, seq\_len, head\_dim]$$. The dtype can be float32 and float64.

Examples

# required: skiptest
import numpy as np

query_data = np.array([[[[0, 1,], [2, 3],
[ 0, 1], [2, 3]]]]).astype("float32")
key_data = np.array([[[[0, 1,], [2, 3],
[ 0, 1], [2, 3]]]]).astype("float32")
value_data = np.array([[[[0, 1,], [2, 3],
[ 0, 1], [2, 3]]]]).astype("float32")
sparse_csr_offset_data = np.array([[[0, 2,
4, 6, 8]]]).astype("int32")
sparse_csr_columns_data = np.array([[[0, 1,
0, 1, 2, 3, 2, 3]]]).astype("int32")
print(query_data.shape)
# (1, 1, 4, 2)
print(sparse_csr_offset_data.shape)
# (1, 1, 5)
print(sparse_csr_columns_data.shape)
# (1, 1, 8)