Will Haste work without CUDA? about haste HOT 2 CLOSED

At the moment it's CUDA-only but we'll definitely get to a CPU implementation.

from haste.

You can now use these layers without CUDA. When the layer and input tensor are on CPU, the layer will use the CPU implementation. When the layer and input tensor are on GPU, it will use the fast CUDA implementation.

Here's an example of an LSTM on CPU:

import torch
import haste_pytorch as haste

batch_size = 128
seq_len = 256
input_size = 128
hidden_size = 256

x = torch.rand(seq_len, batch_size, input_size).cpu()
lstm = haste.LSTM(input_size, hidden_size).cpu()
y, state = lstm(x)

Note that all Haste RNN features (e.g. Zoneout, DropConnect) are supported by the CPU implementation as well.

from haste.