[dynamo] torch._dynamo.exc.Unsupported: comparison SymNodeVariable() <built-in function is_> ListVariable()

[ezyang]

🐛 Describe the bug

Self explanatory.

Occurred when implementing self-referentiality check in torch.tensor ref

            if cur_item is obj:
                raise TypeError("new(): self-referential lists are incompatible")

Versions

main

cc @msaroufim @wconstab @bdhirsh @anijain2305 @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @Xia-Weiwen @wenzhe-nrv @jiayisunx @chenyang78 @aakhundov @kadeng

[aqbewtra]

This is my first issue. Do you mind providing more detail on how I can recreate the bug?

[ezyang]

import torch

@torch.compile(backend='eager', fullgraph=True, dynamic=True)
def f(x, xs):
    if x.size(0) is xs:
        return x + 1
    else:
        return x * 2

f(torch.randn(2), [1, 2])

[vdesai2014]

@ezyang

Currently when the 'is' operator is used with a SymNodeVariable, an unimplemented exception is throw because the op isn't currently part of supported_tensor_comparison_ops (line 1442, dynamo/variables/built_in.py). There is an existing if statement to handle 'is' operator called on different types (lines 1458-1461) but the fact that its lower down in the function causes the unimplemented exception to get thrown first.

I think moving the catch for 'is' op on different types to the top of the function should fix this issue without any second-order effects. Can you confirm this makes sense? My first Pytorch issue so want to make sure.

[ezyang]

Give it a try

[jon-chuang]

How about this test case?

import torch

@torch.compile(backend='eager', fullgraph=True, dynamic=True)
def f(x, xs):
    if x.size(0) is xs:
        return x + 1
    else:
        return x * 2

f(torch.randn(2), 2)

We also want to handle the case where it is indeed self-referential etc. For that we need something closer to #111550

[laithsakka]

@vdesai2014 @jon-chuang is this still an open issue? is anyone working on this?

[ezyang]

Feel free to grab this

[ezyang]

still repros as of 0a1b3be

[joshue031]

If this is still free to work on, I would propose adding the following logic to _comparison_with_symnode in torch/_dynamo/variables/builtin.py:

if op in [operator.is_, operator.is_not]:
    is_result = (
        isinstance(left, SymNodeVariable)
        and isinstance(right, SymNodeVariable)
        and left.sym_num is right.sym_num
    )
    if op is operator.is_:
        return ConstantVariable.create(is_result)
    else:
        return ConstantVariable.create(not is_result)

This is similar to what is done in _comparison_with_tensor and seems to work for the two example test cases above.