Support item() and related shape assertion operators

[dayo09]

What?

Let's support below torch graph.

class SimpleItem(TestModuleBase):
    def __init__(self):
        super().__init__()

    def forward(self, x, y):
        return x.item() + y

    def get_example_inputs(self):
        return (torch.tensor(1), 2), {}

opcode         name                    target                       args                                                                  kwargs
-------------  ----------------------  ---------------------------  --------------------------------------------------------------------  --------
placeholder    x                       x                            ()                                                                    {}
placeholder    y                       y                            ()                                                                    {}
call_function  item                    aten.item.default            (x,)                                                                  {}
call_function  ge                      <built-in function ge>       (item, 2)                                                             {}
call_function  _assert_scalar_default  aten._assert_scalar.default  (ge, "Runtime assertion failed for expression u0 >= 2 on node 'ge'")  {}
call_function  add                     <built-in function add>      (item, 2)                                                             {}
output         output                  output                       ((add,),)                                                             {}
ERROR:tico.utils.convert:NOT SUPPORTED OPERATOR
        (op) item.default
        (trace)   File "/home/dayoung.lee/git/TICO/test/modules/op/item.py", line 25, in forward
    return x.item() + y

ERROR:tico.utils.convert:NOT SUPPORTED OPERATOR
        (op) ge
        (trace)   File "/home/dayoung.lee/git/TICO/test/modules/op/item.py", line 25, in forward
    return x.item() + y

ERROR:tico.utils.convert:NOT SUPPORTED OPERATOR
        (op) _assert_scalar.default
        (trace)   File "/home/dayoung.lee/git/TICO/test/modules/op/item.py", line 25, in forward
    return x.item() + y

ERROR:tico.utils.convert:NOT SUPPORTED OPERATOR
        (op) add
        (trace)   File "/home/dayoung.lee/git/TICO/test/modules/op/item.py", line 25, in forward
    return x.item() + y


TEST FAILED at 'convert_pt2_to_circle ...'
 NOT SUPPORTED OPERATOR IN GRAPH MODULE
ERROR

My simple plan is to remove assertion operators from remove_nop pass and remove dangling builtin ops.

[dayo09]

• torch.ops.aten.item.default maps to aten._local_scalar_dense.default. Both extracts scalar value from tensor.
  • Scalar value has SymInt fake tensor value.
  • It's possible to add SymInt as integer tensor, but I am not sure whether it's just hack or proper implementation.