[AxelNoun]

Summary

_tensor_spec_to_evalue reads a constant's bytes by casting spec.storage.data_ptr() to a host pointer. When the storage is on CUDA, that pointer is a device address, so to_executorch() segfaults (exit 139) with no Python exception. This PR checks the storage device first and raises an InternalError, as @shoumikhin suggested in #21545, that names the device and says how to fix the export. As discussed in the issue, lowering a CUDA module onto the portable path is not supported; only the way it fails changes. The check applies only to non-empty constants that the emitter serializes itself and whose storage is not on CPU, which are exactly the cases that reach the host read of a device pointer.

Fixes #21545

Test plan

New test test_emit_non_cpu_constant_raises in exir/emit/test/test_emit.py. It uses meta storage, which is not host-readable either and needs no accelerator. Without the guard, that case already fails with a wrapped "NULL pointer access" InternalError, so the test matches on the device name in the message.

Built and installed from this branch after merging main at d6abe23 (pip install . --no-build-isolation on Ubuntu 24.04, Python 3.12, torch 2.14.0+cpu):

pytest exir/emit/test/test_emit.py

70 passed, 2 skipped, as before the merge. With _emitter.py replaced by main's, the new test fails on the "NULL pointer access" message. lintrunner --skip MYPY reports no issues on the two files.

I also ran the reproducer from #21545 on CUDA (RTX 3060 under WSL), with the same package and torch replaced by the cu130 build of the same version (2.14.0+cu130). With main's _emitter.py it segfaults in _tensor_spec_to_evalue (exit 139). With this branch it raises InternalError: ... Constant tensor storage is on device 'cuda:0', ... (exit 1).

I used Claude Code to write this PR.

cc @JacobSzwejbka @angelayi @nil-is-all @Gasoonjia @digantdesai

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24202

• 📄 Preview Python docs built from this PR

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

❌ 1 New Failure, 1 Cancelled Job, 1 Unrelated Failure

As of commit 5e310ec with merge base ec56cfa ():

NEW FAILURE - The following job has failed:

• Test QNN Windows MSVC build / build-qnn-windows-arm64 (gh)

CANCELLED JOB - The following job was cancelled. Please retry:

• pull / nxp-mcuxpresso-test / linux-job (gh)
  Could not load credentials from any providers

FLAKY - The following job failed but was likely due to flakiness present on trunk:

• pull / unittest / windows / windows-job (gh) (matched win rule in flaky-rules.json)
  ##[error]The operation was canceled.

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

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

[AxelNoun]

The red nxp-mcuxpresso-test job here is unrelated to this PR, which only touches exir/emit/_emitter.py and its test.

That job runs west init against mcuxsdk-manifests without a revision, so it follows that repository's default branch. The branch currently pins the fw_bin project to NXP/wifi_nb_fw@2553ab1d2097f0adb6e0846f054142c86045db3f, a commit the remote no longer serves, and the project is declared clone-depth: 1, so west has no fallback:

fatal: remote error: upload-pack: not our ref 2553ab1d2097f0adb6e0846f054142c86045db3f
ERROR: update failed for project fw_bin
FATAL ERROR: Command failed: Update failed

The same job is failing on main (runs 41546129726, 41545177768, 41542178609), so re-running it here will not help until the manifest is pinned or NXP restores the commit. #24479 pins the manifest to v26.09.00.

[Gasoonjia]

thx lgtm i will stamp it after ci passes