torch.multinomial raises no error when sampling from zero weight and replacement=False.

[FlyDogDaDa]

🐛 Describe the bug

As in the example, torch.multinomial should raise an error.

If errors are not raised to block developers, it will happen that the wrong index samples are used to train the model without knowing it, like me.

Here is the code I executed:

import torch
print("git_version =", torch.version.git_version)
print("torch_version =", torch.version.__version__)
weights = torch.tensor([0, 10, 3, 0], dtype=torch.float)
a = torch.multinomial(weights, 2, replacement=False)
b = torch.multinomial(weights, 4, replacement=False)
print(a)
print(b)

I get the output:

git_version = 39901f229520a5256505ec24782f716ee7ddc843
torch_version = 2.2.2+cpu
tensor([1, 2])
tensor([2, 1, 0, 3])

Versions

Here is the version I'm using:torch_version=2.2.2+cpu

[drisspg]

I think this is an issue with the docs but not with the actual code since there are 4 elements in the weights matrix and there are 4 things to draw. If you said this to 5 you should see the error