Support PyTorch 2.8 #2630

M-Quadra · 2025-12-18T12:26:14Z

Adapt torch 2.8, torchao 0.12.0, executorch 0.7.0.

Environment

executorch==0.7.0
├── torch [required: >=2.8.0,<2.9.0, installed: 2.8.0]
└── torchao [required: ==0.12.0, installed: 0.12.0]

Features and Bug Fixes

Fix test_batchnorm_dynamic_stress

pytest coremltools/converters/mil/frontend/torch/test/test_torch_ops.py::TestBatchNorm::test_batchnorm_dynamic_stress

For torch.nn.functional.batch_norm:

E       ValueError: running_mean and running_var must either both be None or neither be None

Adapt ones_like dtype

Adapt ones_like dtype for torch 2.8.0

pytest coremltools/converters/mil/frontend/torch/test/test_torch_ops.py::TestIndexPut::test_index_put_updates_bool

ValueError: In op, of type scatter_nd, named index_put, the named input `updates` must have the same data type as the named input `data`. However, updates has dtype fp32 whereas data has dtype int32.

import torch
import numpy as np
import coremltools as ct


class Model(torch.nn.Module):
    def forward(self, x):
        x = torch.ones(x.shape, dtype=torch.bool)
        y = torch.ones_like(x).bool()
        mask = torch.tensor([True, False, False, False, True, True]).view(3, 2)
        x[mask] = y[mask]
        return x


x = torch.randn(3, 2)
model = Model().eval()

exported_model = torch.export.export(model, (x,)).run_decompositions({})
mlmodel = ct.convert(
    exported_model,
    minimum_deployment_target=ct.target.iOS16,
)

y0 = model(x).numpy()
y1 = mlmodel.predict({"x": x.numpy()})["index_put"]
assert np.equal(y0, y1).all()

exported_model.graph in torch 2.7.0

%c_lifted_tensor_0 : [num_users=1] = placeholder[target=c_lifted_tensor_0]
%x : [num_users=0] = placeholder[target=x]
%ones : [num_users=2] = call_function[target=torch.ops.aten.ones.default](args = ([3, 2],), kwargs = {dtype: torch.bool, device: cpu, pin_memory: False})
%ones_like : [num_users=1] = call_function[target=torch.ops.aten.ones_like.default](args = (%ones,), kwargs = {pin_memory: False})
%_to_copy : [num_users=1] = call_function[target=torch.ops.aten._to_copy.default](args = (%ones_like,), kwargs = {dtype: torch.bool})
%clone : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%c_lifted_tensor_0,), kwargs = {})
%view : [num_users=2] = call_function[target=torch.ops.aten.view.default](args = (%clone, [3, 2]), kwargs = {})
%index : [num_users=2] = call_function[target=torch.ops.aten.index.Tensor](args = (%_to_copy, [%view]), kwargs = {})
%sym_size_int_1 : [num_users=3] = call_function[target=torch.ops.aten.sym_size.int](args = (%index, 0), kwargs = {})
%sym_constrain_range_for_size_default : [num_users=0] = call_function[target=torch.ops.aten.sym_constrain_range_for_size.default](args = (%sym_size_int_1,), kwargs = {})
%ge_2 : [num_users=1] = call_function[target=operator.ge](args = (%sym_size_int_1, 0), kwargs = {})
%_assert_scalar_default : [num_users=0] = call_function[target=torch.ops.aten._assert_scalar.default](args = (%ge_2, Runtime assertion failed for expression u0 >= 0 on node 'ge_2'), kwargs = {})
%le_1 : [num_users=1] = call_function[target=operator.le](args = (%sym_size_int_1, 6), kwargs = {})
%_assert_scalar_default_1 : [num_users=0] = call_function[target=torch.ops.aten._assert_scalar.default](args = (%le_1, Runtime assertion failed for expression u0 <= 6 on node 'le_1'), kwargs = {})
%index_put : [num_users=1] = call_function[target=torch.ops.aten.index_put.default](args = (%ones, [%view], %index), kwargs = {})
return (index_put,)

exported_model.graph in torch 2.8.0

%c_lifted_tensor_0 : [num_users=1] = placeholder[target=c_lifted_tensor_0]
%x : [num_users=0] = placeholder[target=x]
%ones : [num_users=2] = call_function[target=torch.ops.aten.ones.default](args = ([3, 2],), kwargs = {dtype: torch.bool, device: cpu, pin_memory: False})
%ones_like : [num_users=2] = call_function[target=torch.ops.aten.ones_like.default](args = (%ones,), kwargs = {pin_memory: False})
%_assert_tensor_metadata : [num_users=0] = call_function[target=torch.ops.aten._assert_tensor_metadata.default](args = (%ones_like, None, None, torch.bool), kwargs = {device: cpu, layout: torch.strided})
%clone : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%c_lifted_tensor_0,), kwargs = {})
%view : [num_users=2] = call_function[target=torch.ops.aten.view.default](args = (%clone, [3, 2]), kwargs = {})
%index : [num_users=2] = call_function[target=torch.ops.aten.index.Tensor](args = (%ones_like, [%view]), kwargs = {})
%sym_size_int_1 : [num_users=3] = call_function[target=torch.ops.aten.sym_size.int](args = (%index, 0), kwargs = {})
%sym_constrain_range_for_size_default : [num_users=0] = call_function[target=torch.ops.aten.sym_constrain_range_for_size.default](args = (%sym_size_int_1,), kwargs = {})
%ge_2 : [num_users=1] = call_function[target=operator.ge](args = (%sym_size_int_1, 0), kwargs = {})
%_assert_scalar_default : [num_users=0] = call_function[target=torch.ops.aten._assert_scalar.default](args = (%ge_2, Runtime assertion failed for expression u0 >= 0 on node 'ge_2'), kwargs = {})
%le_1 : [num_users=1] = call_function[target=operator.le](args = (%sym_size_int_1, 6), kwargs = {})
%_assert_scalar_default_1 : [num_users=0] = call_function[target=torch.ops.aten._assert_scalar.default](args = (%le_1, Runtime assertion failed for expression u0 <= 6 on node 'le_1'), kwargs = {})
%index_put : [num_users=1] = call_function[target=torch.ops.aten.index_put.default](args = (%ones, [%view], %index), kwargs = {})
return (index_put,)

In torch 2.8.0, the dtype of ones_like may be moved to _assert_tensor_metadata.

Fix test_unfold

pytest coremltools/converters/mil/frontend/torch/test/test_torch_ops.py::TestUnfold::test_unfold

E                   torch.fx.experimental.symbolic_shapes.ConstraintViolationError: Constraints violated (w)! For more information, run with TORCH_LOGS="+dynamic".
E                     - Not all values of w = L['args'][0].size()[3] in the specified range 3 <= w <= 128 satisfy the generated guard max(1, (1 + L['args'][0].size()[3]) // 2) == ((1 + L['args'][0].size()[3]) // 2).

To fix the unit test, replace the following code:

h_torch, w_torch = torch.export.Dim("h", min=min_h, max=128), torch.export.Dim("w", min=min_w, max=128)

with:

h_torch, w_torch = torch.export.Dim.AUTO, torch.export.Dim.AUTO

Adapt torch.nn.SELU for executorch 0.7.0

pytest coremltools/converters/mil/frontend/torch/test/test_torch_ops.py::TestActivation::test_selu

>               np.testing.assert_allclose(coreml_result, torch_result, atol=atol, rtol=rtol)
E               AssertionError: 
E               Not equal to tolerance rtol=1e-05, atol=0.0001
E               
E               Mismatched elements: 6 / 7 (85.7%)
E               Max absolute difference among violations: 0.3042059
E               Max relative difference among violations: 0.04825448
E                ACTUAL: array([-1.669116, -1.642616, -1.446812,  0.      ,  2.      ,  4.      ,
E                       6.      ], dtype=float32)
E                DESIRED: array([-1.753741, -1.725899, -1.520167,  0.      ,  2.101402,  4.202804,
E                       6.304206], dtype=float32)

coremltools/converters/mil/frontend/torch/test/testing_utils.py:315: AssertionError

import torch
import executorch.exir


x = torch.tensor([-6.0, -4.0, -2.0, 0.0, 2.0, 4.0, 6.0])
model = torch.nn.SELU(inplace=False).eval()
model_spec = torch.export.export(
    model, (x,)
).run_decompositions({})
model_spec = executorch.exir.to_edge(model_spec).exported_program()
print(model_spec.graph)

graph in executorch 0.7.0

%input_1 : [num_users=1] = placeholder[target=input]
%aten_elu_default : [num_users=1] = call_function[target=executorch.exir.dialects.edge._ops.aten.elu.default](args = (%input_1, 1.6732632423543772, 1.0507009873554805), kwargs = {})
return (aten_elu_default,)

@register_torch_op
def elu(context, node):

In executorch 0.7.0, torch.nn.SELU is decomposed to aten.elu. But the elu operator implementation is missing scale support.

Fix source code assert

pytest -x coremltools/converters/mil/frontend/torch/test/test_torch_export_conversion_api.py::TestExecuTorchExamples::test_add
pytest -x coremltools/converters/mil/frontend/torch/test/test_torch_export_conversion_api.py::TestExecuTorchExamples::test_add_mul
pytest -x coremltools/converters/mil/frontend/torch/test/test_torch_export_conversion_api.py::TestExecuTorchExamples::test_linear
pytest -x coremltools/converters/mil/frontend/torch/test/test_torch_export_conversion_api.py::TestExecuTorchExamples::test_mul
pytest -x coremltools/converters/mil/frontend/torch/test/test_torch_export_conversion_api.py::TestExecuTorchExamples::test_softmax

Adapt torchao 0.12.0

pytest -x coremltools/converters/mil/frontend/torch/test/test_torch_conversion_api.py::TestTorchao::test_weight_only_quantization
pytest -x coremltools/converters/mil/frontend/torch/test/test_torch_quantization_ops.py::TestPytorchQuantizedOps::test_unpack_int4packed_by_mm_with_eye_matrix
pytest -x coremltools/converters/mil/frontend/torch/test/test_torch_quantization_ops.py::TestPytorchQuantizedOps::test_weight_int4pack_mm
pytest -x coremltools/test/optimize/torch/quantization/test_coreml_quantizer.py::test_weight_module_act_fusion

    @functools.wraps(func)
    def decorate_context(*args, **kwargs):
        with ctx_factory():
>           return func(*args, **kwargs)
E           TypeError: quantize_affine() got an unexpected keyword argument 'zero_point_domain'

                and (
                    torch.nn.modules.batchnorm.BatchNorm2d
>                   in [val[1] for val in node.meta["source_fn_stack"]]
                    or torch.nn.modules.batchnorm.BatchNorm1d
                    in [val[1] for val in node.meta["source_fn_stack"]]
                )
            ):
E           KeyError: 'source_fn_stack'

In torchao 0.12.0, torchao_quant.quantize_affine no longer accepts the zero_point_domain argument.

torchao.quantization.pt2e expects node.meta["source_fn_stack"]; ensure node.meta exists by adding a dummy placeholder when missing.

Add dead node elimination

pytest -x coremltools/converters/mil/frontend/torch/test/test_torch_ops.py::TestEinsum::test_ternary_einsum

NotImplementedError: Unsupported fx node or__3, kind or_

Some nodes in the FX graph are not used, but the graph validation does not ignore dead nodes and thus fails.

Related Issue

Closes #2610.

TobyRoseman · 2025-12-18T17:40:40Z

Thanks for the Pull Request @M-Quadra! It would be great to support a newer version of PyTorch.

The code changes look good to me.

I've kicked off a CI run:
https://gitlab.com/coremltools1/coremltools/-/pipelines/2222858268
Even if this passes, I'll likely need to do more internal testing before we can merge this change.

M-Quadra · 2025-12-18T23:24:56Z

PyTorch on the x86_64 platform is older and behaves differently; I'll mark the test with pytest.xfail soon.

TobyRoseman · 2025-12-19T19:21:47Z

That's a whole lot of tests to xfail. I think you should look into why they're failing.

X86_64 machines are running a different version of PyTorch:

coremltools/reqs/pytorch.pip

Line 3 in 0434e85

torch==2.2.0; platform_machine != "arm64"

That's likely the cause.

M-Quadra · 2025-12-20T07:08:21Z

Failing test:

coremltools/converters/mil/frontend/torch/test/test_torch_ops.py::TestUnfold::test_unfold

E           AttributeError: 'function' object has no attribute 'AUTO'

The torch.export.Dim.AUTO feature is only available in higher versions.

Solution

Use platform.machine() == "x86_64" to differentiate older versions.

TobyRoseman · 2026-01-05T16:38:07Z

Updated CI: https://gitlab.com/coremltools1/coremltools/-/pipelines/2245622649

M-Quadra added 11 commits January 4, 2026 20:31

Pin torch 2.8

eb1a2ea

Adapt ones_like dtype for torch 2.8.0

6560c03

Fix test_batchnorm_dynamic_stress

ce9289e

Fix test_unfold

6fd0d2c

Adapt torch.nn.SELU for executorch 0.7.0

86f54bf

Fix source code assert

2e9fc6b

Adapt _weight_int4pack_mm_for_cpu for torchao 0.12.0

717e9a8

Add dead node elimination

994201f

Adapt test_unpack_int4packed_by_mm_with_eye_matrix for torchao 0.12.0

98e0d2a

Adapt test_weight_module_act_fusion for torchao 0.12.0

29ee951

fix test_unfold in x86

3661d9c

M-Quadra force-pushed the torch-28 branch from c914c5b to 3661d9c Compare January 4, 2026 21:40

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Support PyTorch 2.8 #2630

Support PyTorch 2.8 #2630

Uh oh!

M-Quadra commented Dec 18, 2025

Uh oh!

TobyRoseman commented Dec 18, 2025

Uh oh!

M-Quadra commented Dec 18, 2025

Uh oh!

TobyRoseman commented Dec 19, 2025

Uh oh!

M-Quadra commented Dec 20, 2025

Uh oh!

TobyRoseman commented Jan 5, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Support PyTorch 2.8 #2630

Are you sure you want to change the base?

Support PyTorch 2.8 #2630

Uh oh!

Conversation

M-Quadra commented Dec 18, 2025

Environment

Features and Bug Fixes

Fix test_batchnorm_dynamic_stress

Adapt ones_like dtype for torch 2.8.0

Fix test_unfold

Adapt torch.nn.SELU for executorch 0.7.0

Fix source code assert

Adapt torchao 0.12.0

Add dead node elimination

Related Issue

Uh oh!

TobyRoseman commented Dec 18, 2025

Uh oh!

M-Quadra commented Dec 18, 2025

Uh oh!

TobyRoseman commented Dec 19, 2025

Uh oh!

M-Quadra commented Dec 20, 2025

Solution

Uh oh!

TobyRoseman commented Jan 5, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants