aten.alias.default operation is not supported by Coremltools converter.

[himalayjor]

🌱 Describe your Feature Request

• Please provide a description of your feature request, clear and consise.
• If this is a feature request for the Core ML Framework or Xcode, please instead submit your feature request using the Feedback Assistant for Developers: https://developer.apple.com/bug-reporting/

aten.alias.default operation is not supported by Coremltools converter.

How can this feature be used?

Please provide some examples where this feature can be used.

Trying to convert complex model from Torch to CoreML model.

Describe alternatives you've considered

NO alternative.

Additional context

Stack trace:

  File "/opt/anaconda3/envs/uat/lib/python3.11/site-packages/coremltools/converters/mil/frontend/torch/converter.py", line 633, in __init__
    self.graph = InternalTorchIRGraph.from_exir(
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/anaconda3/envs/uat/lib/python3.11/site-packages/coremltools/converters/mil/frontend/torch/internal_graph.py", line 590, in from_exir
    nodes.append(InternalTorchIRNode.from_exir_node(node=node))
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/anaconda3/envs/uat/lib/python3.11/site-packages/coremltools/converters/mil/frontend/torch/internal_graph.py", line 289, in from_exir_node
    raise ValueError(f"Unsupported fx node {str(node)}, kind {kind}")
ValueError: Unsupported fx node alias, kind alias

[TobyRoseman]

@himalayjor - can you please give us a model stub which use this op?

[himalayjor]

Dummy Model Code:

    def forward(self, x):
        y = torch.ops.aten.alias(x)
        y = y + 5
        return y

[himalayjor]

Is there an ETA for this ? Work is blocked because of this.

[TobyRoseman]

Complete code to reproduce:

import torch
import torch.nn as nn

import coremltools as ct

class Model(torch.nn.Module):
    def forward(self, x):
        y = torch.ops.aten.alias(x)
        y = y + 5
        return y

exported_model = torch.export.export(Model(), (torch.rand(2, 32),))
ct.convert(exported_model)

Any ideas @YifanShenSZ?

[TobyRoseman]

@himalayjor - you could edit your model to remove the alias usage. Seems like that shouldn't be difficult to do.

[Lxrd-AJ]

@himalayjor - you could edit your model to remove the alias usage. Seems like that shouldn't be difficult to do.

Some models exported from torch.export.export include this op and sometimes are not easily editable.

@himalayjor Does this composite op work for you?

if 'alias' not in _TORCH_OPS_REGISTRY:
    @register_torch_op
    def alias(context, node):
        x = _get_inputs(context, node, expected=1)[0]
        x_identity = mb.identity(x=x, name=node.name)
        context.add(x_identity)