Fixed device binding for active learning loop

[mehdiataei]

PhysicsNeMo Pull Request

Fixed device binding. DistributedManager.device binds the descriptor, not the torch.device.

Description

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.
• The CHANGELOG.md is up to date with these changes.
• An issue is linked to this pull request.
• If I am implementing a new model or modifying any existing model, I have followed the Models Implementation Coding Standards.

Dependencies

Review Process

All PRs are reviewed by the PhysicsNeMo team before merging.

Depending on which files are changed, GitHub may automatically assign a maintainer for review.

We are also testing AI-based code review tools (e.g., Greptile), which may add automated comments with a confidence score.
This score reflects the AI’s assessment of merge readiness and is not a qualitative judgment of your work, nor is
it an indication that the PR will be accepted / rejected.

AI-generated feedback should be reviewed critically for usefulness.
You are not required to respond to every AI comment, but they are intended to help both authors and reviewers.
Please react to Greptile comments with 👍 or 👎 to provide feedback on their accuracy.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[mehdiataei]

/blossom-ci

[greptile-apps]

Greptile Summary

This PR fixes a bug where DefaultTrainingLoop.__init__ read DistributedManager.device — the raw property descriptor object on the class — instead of DistributedManager().device, the actual torch.device returned by the instance property. A new test confirms the resolved value is a torch.device and matches the singleton's device.

• loop.py: One-line change from DistributedManager.device to DistributedManager().device; safe because the call is already guarded by DistributedManager.is_initialized().
• test_loop.py: New test initializes the singleton when needed and asserts both the type and value of loop.device; teardown uses _shared_state.clear() rather than DistributedManager.cleanup() (see inline comment).

Important Files Changed

Filename	Overview
physicsnemo/active_learning/loop.py	Single-line fix: DistributedManager.device (class-level property descriptor) replaced with DistributedManager().device (instance property returning a torch.device); already guarded by is_initialized().
test/active_learning/test_loop.py	New test verifies the fix; teardown uses _shared_state.clear() rather than the official DistributedManager.cleanup() API.
CHANGELOG.md	CHANGELOG entry added correctly under the unreleased "Fixed" section.

_{Reviews (1): Last reviewed commit: "Fixed device binding" | Re-trigger Greptile}

[greptile-apps]

The teardown accesses the internal Borg-pattern dict directly with .clear() instead of using DistributedManager.cleanup(), the official API. cleanup() also handles process-group teardown when _distributed is True, so the manual .clear() could skip that step in an environment where distributed training is actually active (e.g. when this test runs as part of a multi-process suite).

Suggested change

	finally:
	if not was_initialized:
	DistributedManager._shared_state.clear()
	finally:
	if not was_initialized:
	DistributedManager.cleanup()