Local Softmax parallelize the softmax computation by splitting the tensor into smaller sub-tensors and applying the softmax function on each of these smaller tensors independently. In other words, we want to compute a "local" softmax on each chunk of the tensor, instead of on the entire tensor.

• Lucidrains
• Agorians

pip install local-sfmx

import torch
from local_sfmx import local_softmax

tensor = torch.rand(10, 5)
result = local_softmax(tensor, 2)
print(result)

function LocalSoftmax(tensor, num_chunks): split tensors into num_chunks smaller tensors for each smaller tensor: apply standard softmax concatenate the results return concatenated tensor

MIT