RuntimeError: Failed running call_function aten.uniform(...

[silvasean]

🐛 Describe the bug

import torch
import torch._dynamo as dynamo

@dynamo.optimize()
def forward(x):
    return torch.ops.aten.uniform(x)

forward(torch.rand(1000))

Error

RuntimeError: Failed running call_function aten.uniform(*(FakeTensor(FakeTensor(..., device='meta', size=(1000,)), cpu),), **{}):
uniform() missing 2 required keyword-only arguments: 'dtype' and 'device'

Error logs

Ful log: https://gist.github.com/silvasean/c502b2a4df4feb97403cdbd8b56eb1b5

Minified repro

See above

[yanboliang]

I think this is a bug about the uniform ref implementation, working on a PR to fix it.

[kvathupo]

I'm not sure this is a bug: torch.ops.aten.uniform() just fills a tensor of the input shape with values from some uniform distribution. torch.ops.aten.uniform(x.size(), dtype=x.dtype, device=my_device) should work to generate a tensor of the input dimensions.

That said, I suppose it's reasonable to expect to be able to call it like torch.ops.aten.randn().

[yanboliang]

The fix has been merged, @silvasean can we close this?

[silvasean]

Feel free to close when the test case in the OP passes!