fix: Fix device state restoration in NamedEntityExtractor and add release …

[ritikraj2425]

Related Issues

• fixes [Bug] NamedEntityExtractor (spaCy) fails to restore device state after execution #11242

Proposed Changes:

Previously, the NamedEntityExtractor (spaCy backend) would unconditionally call spacy.require_cpu() after execution. Since spaCy and Thinc use a global state for device configuration, this would override any pre-existing user configuration (e.g., if the user was using a specific GPU for other parts of their application).

This PR:

• Captures the current Thinc Ops state at the start of the _select_device context manager.
• Restores the original Ops state in the finally block instead of forcing a reset to CPU.
• Removes the outdated TODO regarding device restoration.

How did you test it?

• Regression Test: Added TestNamedEntityExtractorDeviceRestoration to test/components/extractors/test_named_entity_extractor.py. This test sets a custom attribute on the global Ops object and verifies that it is preserved after the component's internal device switching logic.
• Unit Tests: Ran the existing NamedEntityExtractor test suite (11 tests passed).
• Manual Verification: Verified that custom NumpyOps objects are not replaced by fresh instances after component execution.

Checklist

• I have read the contributors guidelines and the code of conduct.
• I have updated the related issue with new insights and changes.
• I have added unit tests and updated the docstrings.
• I've used one of the conventional commit types for my PR title: fix:.
• I have documented my code.
• I have added a release note file.
• I have run pre-commit hooks and fixed any issue.

[vercel]

@ritikraj2425 is attempting to deploy a commit to the deepset Team on Vercel.

A member of the Team first needs to authorize it.

[sjrl]

Hey @ritikraj2425 is this an issue you were running into with scripts you were running yourself? Also please look at the failing CI and fix the issues there.

[ritikraj2425]

Hi @sjrl!
Yes, I was running some custom test scripts while exploring how Haystack components manage hardware resources. I noticed that NamedEntityExtractor calls a global spacy.require_cpu() in its finally block, which could unintentionally wipe out device settings for other components sharing the same GPU. I implemented the Thinc Ops state restoration to prevent this global side-effect!

I also pushed a fix for the CI. I had mistakenly placed my new device restoration test in the standard unit tests (where spaCy isn't installed in CI), causing an ImportError during collection. I have now moved it to the e2e pipeline tests where the spaCy backend is properly tested.

Let me know if there's anything else I should adjust!

[sjrl]

Hey @ritikraj2425 thanks for the changes! Could you fix the mypy issues? https://github.com/deepset-ai/haystack/actions/runs/25561973899/job/75037156133?pr=11259

[ritikraj2425]

Hey @sjrl!
I synced with main and noticed that some recent upstream changes caused mypy failures in the HuggingFace and OpenAI response components. I've re-added the necessary # type: ignore comments in this PR to bypass the errors and get the CI green again.

Everything should be passing now! Let me know if there's anything else you need.

[ritikraj2425]

Done! I had just added those because they were failing the mypy CI on my branch and I was trying to get the build green, but I've reverted them now to keep this PR strictly scoped to the NER fix.

[sjrl]

Thanks!