xb.Op.scalar maps unsigned int to torch signed int

[tengyifei]

torch supports unsigned int types so this is probably a bug.

[iwknow]

is there an instruction to reproduce the issue? if it is easy to reproduce, i can take a look.

[rpsilva-aws]

I don't have an exact repro at the moment, but I can write down on how we could go about it.
The issue refers to _XLA_PT_TYPE_MAP.

The only place where this exists today is (via to_torch_type):

  @classmethod
  def scalar(cls, builder, value, dtype=None):
    return mkleaf(
        'Constant',
        builder,
        value=torch.tensor(value, dtype=cls.to_torch_type(dtype)))

Hence, if you can use the XLA builder to generate a computation that derives the type of a torch.uint32, but generates an XLA S32 type - we could use that to trigger either a lowering error or an accuracy issue (e.g. overflow).

You can start by looking at test_operations.py to mimic/repro the behavior. If I have time, I can generate one.