python.builtin

dynamic_shape_round

Note

Tags: python.builtin, torch.dynamic-shape

Support Level: NOT_SUPPORTED_YET

Original source code:

import torch

from torch.export import Dim

x = torch.ones(3, 2)
dim0_x = Dim("dim0_x")

def dynamic_shape_round(x):
    """
    Calling round on dynamic shapes is not supported.
    """
    return x[: round(x.shape[0] / 2)]

Result:

Unsupported: Calling round() on symbolic value is not supported. You can use floor() to implement this functionality

tensor_setattr

Note

Tags: python.builtin

Support Level: SUPPORTED

Original source code:

import torch



def tensor_setattr(x, attr):
    """
    setattr() call onto tensors is not supported.
    """
    setattr(x, attr, torch.randn(3, 2))
    return x + 4

Result:

ExportedProgram:
    class GraphModule(torch.nn.Module):
        def forward(self, l_x_: "f32[3, 2]", arg1):
                add: "f32[3, 2]" = torch.ops.aten.add.Tensor(l_x_, 4);  l_x_ = None
            return (add,)

Graph signature: ExportGraphSignature(input_specs=[InputSpec(kind=<InputKind.USER_INPUT: 1>, arg=TensorArgument(name='l_x_'), target=None), InputSpec(kind=<InputKind.USER_INPUT: 1>, arg=ConstantArgument(value='attr'), target=None)], output_specs=[OutputSpec(kind=<OutputKind.USER_OUTPUT: 1>, arg=TensorArgument(name='add'), target=None)])
Range constraints: {}
Equality constraints: []

type_reflection_method

Note

Tags: python.builtin

Support Level: SUPPORTED

Original source code:

import torch



class A:
    @classmethod
    def func(cls, x):
        return 1 + x


def type_reflection_method(x):
    """
    type() calls on custom objects followed by method calls are not allowed
    due to its overly dynamic nature.
    """
    a = A()
    return type(a).func(x)

Result:

ExportedProgram:
    class GraphModule(torch.nn.Module):
        def forward(self, l_x_: "f32[3, 4]"):
                add: "f32[3, 4]" = torch.ops.aten.add.Tensor(l_x_, 1);  l_x_ = None
            return (add,)

Graph signature: ExportGraphSignature(input_specs=[InputSpec(kind=<InputKind.USER_INPUT: 1>, arg=TensorArgument(name='l_x_'), target=None)], output_specs=[OutputSpec(kind=<OutputKind.USER_OUTPUT: 1>, arg=TensorArgument(name='add'), target=None)])
Range constraints: {}
Equality constraints: []

You can rewrite the example above to something like the following:

def type_reflection_method_rewrite(x):
    """
    Custom object class methods will be inlined.
    """
    return A.func(x)