CoreML export → .pte fails with Metadata is invalid or missing

🐛 Describe the bug

I’m trying to export an EfficientViT segmentation model trained on custom data to ExecuTorch .pte with CoreML delegation, following the docs

Problem arises only when using CoreML partitioner not XNNPACK.

-macOS: 15.6 (Apple Silicon, arm64)
-Python: 3.10.18

• torch 2.8.0
• executorch 0.7.0
• coremltools 8.3.0

Script (simplified)

import torch, coremltools as ct
from torch.export import export as torch_export
from executorch.exir import to_edge_transform_and_lower
from executorch.backends.apple.coreml.partition import CoreMLPartitioner
from executorch.backends.apple.coreml.compiler import CoreMLBackend
from executorch.backends.xnnpack.partition.xnnpack_partitioner import XnnpackPartitioner

model = ...  # EfficientViT segmentation, checkpoint loaded, eval mode
example_input = torch.randn(1, 3, 640, 640)

with torch.no_grad():
    ep = torch_export(model, (example_input,), strict=True)

compile_specs = CoreMLBackend.generate_compile_specs(
    minimum_deployment_target=ct.target.iOS17,
    compute_precision=ct.precision.FLOAT16,
    compute_unit=ct.ComputeUnit.ALL,
    model_type=CoreMLBackend.MODEL_TYPE.MODEL
)

edge_mgr = to_edge_transform_and_lower(
    ep,
    partitioner=[
        CoreMLPartitioner(compile_specs=compile_specs),
        XnnpackPartitioner(per_op_mode=True),
    ],
)

pte = edge_mgr.to_executorch()
with open("segmentation_model_ios.pte", "wb") as f:
    f.write(pte.buffer)

Output:

WARNING:torchao.kernel.intmm:Warning: Detected no triton, on systems without Triton certain kernels will not work
W0930 12:44:31.462000 69374 .venv/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs.
/.venv/lib/python3.10/site-packages/executorch/exir/dialects/edge/_ops.py:9: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
  import pkg_resources
WARNING:coremltools:Torch version 2.8.0 has not been tested with coremltools. You may run into unexpected errors. Torch 2.5.0 is the most recent version that has been tested.
✅ Registrato modello 'efficientvit-seg-b0-cards' nel model zoo.
Missing: []
Unexpected: []
WARNING:root:Torchgen is not able to parse the schema of aten::conv2d(Tensor input, Tensor weight, Tensor? bias=None, SymInt[2] stride=[1, 1], SymInt[2] padding=[0, 0], SymInt[2] dilation=[1, 1], SymInt groups=1) -> Tensor.  This is not fatal.
WARNING:root:Op aten.add.Tensor was requested for preservation by partitioner.  This request is ignored because it is in a blocklist.
WARNING:root:Torchgen is not able to parse the schema of aten::conv2d(Tensor input, Tensor weight, Tensor? bias=None, SymInt[2] stride=[1, 1], SymInt[2] padding=[0, 0], SymInt[2] dilation=[1, 1], SymInt groups=1) -> Tensor.  This is not fatal.
WARNING:root:Op aten._native_batch_norm_legit_no_training.default was requested for preservation by partitioner.  This request is ignored because it aliases output.
WARNING:root:Op aten.add.Tensor was requested for preservation by partitioner.  This request is ignored because it is in a blocklist.
WARNING:root:Op aten.view.default was requested for preservation by partitioner.  This request is ignored because it aliases output.
WARNING:root:Op aten.slice.Tensor was requested for preservation by partitioner.  This request is ignored because it aliases output.
WARNING:root:Op aten.permute.default was requested for preservation by partitioner.  This request is ignored because it aliases output.
WARNING:root:Op aten.expand.default was requested for preservation by partitioner.  This request is ignored because it aliases output.
WARNING:root:Op aten.div.Tensor was requested for preservation by partitioner.  This request is ignored because it is in a blocklist.
WARNING:root:Op aten._to_copy.default was requested for preservation by partitioner.  This request is ignored because it is in a blocklist.
WARNING:root:Op aten.mul.Tensor was requested for preservation by partitioner.  This request is ignored because it is in a blocklist.
WARNING:root:Op aten.sub.Tensor was requested for preservation by partitioner.  This request is ignored because it is in a blocklist.
WARNING:root:Op aten.unsqueeze.default was requested for preservation by partitioner.  This request is ignored because it aliases output.
WARNING:root:Op aten.squeeze.dims was requested for preservation by partitioner.  This request is ignored because it aliases output.
Converting PyTorch Frontend ==> MIL Ops:   0%|                                                                                                         | 0/4 [00:00<?, ? ops/s]WARNING:coremltools:Saving value type of int64 into a builtin type of int32, might lose precision!
WARNING:coremltools:Saving value type of int64 into a builtin type of int32, might lose precision!
WARNING:coremltools:Saving value type of int64 into a builtin type of int32, might lose precision!
WARNING:coremltools:Saving value type of int64 into a builtin type of int32, might lose precision!
Converting PyTorch Frontend ==> MIL Ops:  75%|███████████████████████████████████████████████████████████████████████▎                       | 3/4 [00:00<00:00, 2036.73 ops/s]
Running MIL frontend_pytorch pipeline: 100%|██████████████████████████████████████████████████████████████████████████████████████████████| 5/5 [00:00<00:00, 7679.06 passes/s]
Running MIL default pipeline: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████| 89/89 [00:00<00:00, 7071.82 passes/s]
Running MIL backend_mlprogram pipeline: 100%|██████████████████████████████████████████████████████████████████████████████████████████| 12/12 [00:00<00:00, 29194.69 passes/s]
Converting PyTorch Frontend ==> MIL Ops:  94%|███████████████████████████████████████████████████████████████████████████████████████▊     | 17/18 [00:00<00:00, 9870.32 ops/s]
Running MIL frontend_pytorch pipeline: 100%|██████████████████████████████████████████████████████████████████████████████████████████████| 5/5 [00:00<00:00, 4882.78 passes/s]
Running MIL default pipeline: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████| 89/89 [00:00<00:00, 1777.23 passes/s]
Running MIL backend_mlprogram pipeline: 100%|███████████████████████████████████████████████████████████████████████████████████████████| 12/12 [00:00<00:00, 4067.86 passes/s]
WARNING:coremltools:Core ML does not support scalar input, so dim_order_ops__to_dim_order_copy_default_5 has been wrapped as rank-1 size-1 tensor
WARNING:coremltools:Core ML does not support scalar input, so dim_order_ops__to_dim_order_copy_default_4 has been wrapped as rank-1 size-1 tensor
WARNING:coremltools:Core ML does not support scalar input, so dim_order_ops__to_dim_order_copy_default_11 has been wrapped as rank-1 size-1 tensor
WARNING:coremltools:Core ML does not support scalar input, so dim_order_ops__to_dim_order_copy_default_10 has been wrapped as rank-1 size-1 tensor
Converting PyTorch Frontend ==> MIL Ops: 100%|██████████████████████████████████████████████████████████████████████████████████████████▊| 553/554 [00:00<00:00, 2941.72 ops/s]
Running MIL frontend_pytorch pipeline: 100%|████████████████████████████████████████████████████████████████████████████████████████████████| 5/5 [00:00<00:00, 54.42 passes/s]
Running MIL default pipeline: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████| 89/89 [00:02<00:00, 41.62 passes/s]
Running MIL backend_mlprogram pipeline: 100%|█████████████████████████████████████████████████████████████████████████████████████████████| 12/12 [00:00<00:00, 91.59 passes/s]
Esportato: segmentation_model_ios.pte MB: 2.3654251098632812

Error

When trying to load the .pte:

[Core ML]  Metadata is invalid or missing.
[Core ML] Model init failed Metadata is invalid or missing.
CoreMLBackend: Failed to init the model.
Init failed for backend CoreMLBackend: 0x23
RuntimeError: loading method forward failed with error 0x23

Export with XNNPACK-only → .pte loads and runs fine.

Could you help me?

Versions

-macOS: 15.6 (Apple Silicon, arm64)
-Python: 3.10.18

• torch 2.8.0
• executorch 0.7.0
• coremltools 8.3.0