torch.lerp: discrepancy between CUDA and CPU (with extremal inputs)

[kshitij12345]

🐛 Describe the bug

import torch

st = torch.tensor(0.2345)
en = torch.tensor(float("inf"))
weight = torch.tensor(-0.9890)

print(torch.lerp(st, en, weight))   # -inf
print(torch.lerp(st.cuda(), en.cuda(), weight.cuda()))  # nan

Versions

master

[khushi-411]

Hi, @kshitij12345; I think this is fixed in the master branch and can be closed. Please see:

In [1]: import torch
   ...: 
   ...: st = torch.tensor(0.2345)
   ...: en = torch.tensor(float("inf"))
   ...: weight = torch.tensor(-0.9890)
   ...: 
   ...: print(torch.lerp(st, en, weight))   # -inf
   ...: print(torch.lerp(st.cuda(), en.cuda(), weight.cuda()))  # nan

tensor(nan)
tensor(nan, device='cuda:0')

In [2]: torch.__version__
Out[2]: '2.0.0a0+gitb7c2a65'

Thanks!

[kshitij12345]

This looks like regression on CPU, as I'd expect -inf as the output.

>>> start = 0.2345
>>> end = -0.9890
>>> weight = float('inf')
>>> start + weight * (end - start)
-inf

[ZailiWang]

Yes this should be a regression, the original CPU result (-inf) should be expected.
I found the PR introduced the bug is #84844, in which weight < 0.5 is updated to abs(weight) < 0.5 so the calc is executed by the formula at the other branch. This judgement is for numerical stability when weight closer to 0, but disregarded this edge case.

import torch

st = torch.tensor(0.2345)
en = torch.tensor(float("inf"))
weight = torch.tensor(-0.9890)

print(st + weight * (en - st)) # -inf
print(en - (en-st)*(torch.tensor(1)-weight)) # nan