[Erik-Lundell]

TensorSpec.from_tensor(..., const=True) makes test tensors contiguous, discarding channels-last strides and reordering their bytes. Bundled inputs can then have incorrect values, while output verification can fail its stride check.

Preserve the layout of bundled test tensors. Construct ATen tensors with strides derived from the serialized dimension order before copying their bytes.

Cover channels-last inputs and outputs, offset views, and non-dense views.

Additionally, a small type hint fix in tensor.py

cc @Gasoonjia @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/23482

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

⏳ No Failures, 1 Pending

As of commit ded0066 with merge base aee0a33 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[Copilot]

🟡 Changes recommended

The updated dim_order_from_stride signature retains an incorrect return type contract.

1 open finding

• Correct dim_order_from_stride return annotation to Tuple[int, ...] · New

What changed in this PR

Preserves bundled tensor layouts so channels-last inputs and outputs serialize and load correctly.

Changes:

• Adds layout-preserving tensor serialization.
• Reconstructs ATen tensors using serialized dimension-order strides.
• Adds coverage for channels-last tensors and views.

File	Description
extension/​pybindings/​test/​test_pybindings.py	Adds end-to-end channels-last bundle verification.
exir/​tests/​test_tensor.py	Tests layout-preserving TensorSpec creation.
exir/​tensor.py	Adds optional layout preservation.
devtools/​bundled_program/​test/​test_bundle_data.py	Tests serialized layouts and view data.
devtools/​bundled_program/​core.py	Preserves bundled input/output layouts.
devtools/​bundled_program/​bundled_program.cpp	Reconstructs tensors with serialized strides.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[zingo]

Hi @rascani and @digantdesai this is in the general code and need a non Arm review :)

[rascani]

@Gasoonjia - do you think you could take a look?

[rascani]

@claude - Review this PR.

[claude]

Claude finished @rascani's task in 1m 22s —— View job

---

PR Review: #23482

Summary: Bundled test tensors now keep their eager strides, so channels-last inputs and outputs load correctly. However, any dense non-contiguous expected output (for example the result of permute or transpose) now aborts verification. Before this PR those outputs worked.

Detailed review

Backward Compatibility

• devtools/bundled_program/core.py:114,129: the layout comes from the eager tensor, not from the program's I/O dim order. clone(memory_format=torch.preserve_format) keeps the exact strides of any dense, non-overlapping tensor, not just channels-last ones. Most callers build expected outputs by running the eager model (model(*inputs)). Take a model ending in x.permute(0, 2, 3, 1) or x.transpose(1, 2):

  • The eager output is a dense permuted view, so it gets serialized with dim order (0, 2, 3, 1) (or similar).
  • The ExecuTorch program's output comes from permute_copy, so it has the default dim order.
  • tensors_are_close then hits ET_CHECK_MSG(... same strides) and aborts (bundled_program.cpp:182 in ATen mode, :200 in lean mode).

  Before this PR, .contiguous() made those cases pass. Inputs have the same problem: a permuted test input used to be normalized to the program's contiguous input and now arrives with a non-default dim order.

  The cleaner design is to normalize each test tensor to the dim order the program declares for that input/output. That is the layout the runtime compares against, and the method's I/O TensorSpecs carry it. At minimum, limit preservation to tensors whose dim order is default or channels-last, and make everything else contiguous as before.

  Please add a test with a permute/transpose eager output, which would have caught this. Fix this →

API Design

• exir/tensor.py:214: preserve_layout is a boolean switch on a widely used constructor, and it only matters when const=True. If you take the normalization approach above, the conversion can happen in BundledProgram before calling from_tensor, and the flag can go. Separately, the "type hint fix" leaves dim_order_from_stride's return type as Tuple[bytes] (Copilot already raised this).

Testing

• extension/pybindings/test/test_pybindings.py:346: the end-to-end test skips non-ATen kernel modes. The lean impl_like / tensors_are_close path is the one used on device, so it should also get an end-to-end channels-last check.

Recommendation

Request Changes. The arbitrary-permutation regression should be fixed before merge. The channels-last fix itself is correct.

[Erik-Lundell]

Sorry, this is not quite ready

[zingo]

Yeah I was overrideing @Erik-Lundell s origina draft as I needed part of this fix for something else and undrafted sorry for the extra work.

[zingo]

Replaced with this #23927