segfault with large tensor for pytorch

[Peach-He]

Hi,
I am facing segmentation fault issue with large input tensor.
Test code:

import torch
from warprnnt_pytorch import RNNTLoss
rnnt_loss = RNNTLoss()

acts = torch.ones(128,256,129,1024, requires_grad=True)
labels = torch.ones(128,128)
act_length = torch.ones(128) * 256
label_length = torch.ones(128) * 128

if acts.dtype != torch.float:
    acts = acts.float()
if labels.dtype != torch.int32:
    labels = labels.int()
if act_length.dtype != torch.int32:
    act_length = act_length.int()
if label_length.dtype != torch.int32:
    label_length = label_length.int()

loss = rnnt_loss(acts, labels, act_length, label_length)
print(loss)
loss.backward()

How can I fix this issue?
Thanks!

[csukuangfj]

Can you try https://github.com/csukuangfj/optimized_transducer ?
It uses less memory and may fix your issue.