[RL] Fix lazy import in isaaclab_rl subpackage

[StafaH]

Description

Make the isaaclab_rl root package and unified entrypoints use the standard stub-driven lazy export mechanism. This prevents single-GPU play imports from loading the multi-GPU Torch Elastic launcher during Torch initialization, which caused a circular-import failure on Windows with Torch 2.11.

The change also adds static export stubs for both public namespaces and subprocess regression coverage for every isaaclab_rl package initializer, ensuring namespace imports do not eagerly load Torch or unrelated RL frameworks.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Release backport

• Backport this pull request to the active release branch after it merges into develop

Testing

• 54 focused entrypoint, installed-workflow, and multi-GPU command tests passed
• Full Isaac Lab formatting and lint checks passed
• isaaclab_rl changelog validation passed

Checklist

• I have read and understood the contribution guidelines
• I have run the pre-commit checks with uv run isaaclab -f
• Documentation changes are not required for this internal import fix
• My changes generate no new warnings
• I have added tests that prove the fix is effective
• I have added a changelog fragment for the touched package
• My name already exists in CONTRIBUTORS.md

[greptile-apps]

Greptile Summary

This PR replaces eager and manually maintained isaaclab_rl exports with the repository’s stub-driven lazy-export mechanism, preventing ordinary entrypoint imports from initializing Torch Elastic and unrelated RL frameworks.

• Adds runtime-adjacent .pyi export maps for the root and entrypoints namespaces.
• Adds isolated subprocess regression coverage for package initialization and selected public exports.
• Documents the lazy-import fix in the package changelog.

Confidence Score: 5/5

The PR appears safe to merge, with the public export mappings preserved and package metadata ensuring the runtime-required stubs are installed.

The new lazy exporter is supplied with complete sibling stubs, representative public imports remain resolvable, and the added subprocess tests verify that ordinary namespace and play-entrypoint imports avoid the unwanted framework initialization.

Important Files Changed

Filename	Overview
source/isaaclab_rl/isaaclab_rl/init.py	Replaces the custom root-level lazy forwarding implementation with the standard stub-driven lazy exporter while preserving version initialization.
source/isaaclab_rl/isaaclab_rl/init.pyi	Defines the root package’s existing public entrypoint exports for runtime lazy resolution and static analysis.
source/isaaclab_rl/isaaclab_rl/entrypoints/init.py	Removes eager imports of API, dispatch, and multi-GPU modules in favor of stub-driven symbol loading.
source/isaaclab_rl/isaaclab_rl/entrypoints/init.pyi	Maps each entrypoint export to its implementation module, allowing the multi-GPU launcher to remain unloaded until requested.
source/isaaclab_rl/test/test_entrypoints.py	Adds isolated import checks demonstrating that package initialization and play entrypoint resolution do not eagerly load Torch or unrelated RL frameworks.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
    A[Import isaaclab_rl] --> B[lazy_export reads __init__.pyi]
    B --> C[Public names registered lazily]
    C --> D{Symbol requested}
    D -->|run_play_cli| E[Load entrypoints dispatch]
    D -->|run_train_multigpu_cli| F[Load Torch Elastic multigpu module]
    A -. no eager load .-> F

Loading

_{Reviews (1): Last reviewed commit: "Fix lazy imports in isaaclab_rl" | Re-trigger Greptile}

[isaaclab-review-bot]

Isaac Lab Review Bot

The PR replaces the root and entrypoints package initializers with the shared stub-driven lazy export mechanism, preserving their public symbol sets while deferring Torch Elastic and unrelated framework imports.

• Design and architecture: The revised import path consistently uses isaaclab.utils.module.lazy_export() with adjacent stubs. The root namespace delegates symbols to entrypoints, while the entrypoints stub maps each symbol to api, dispatch, or multigpu, allowing the multi-GPU launcher to remain deferred.
• API: The new stubs preserve the previous exports: twelve root symbols and thirteen entrypoints symbols, including BackendName only under isaaclab_rl.entrypoints. __version__ remains eagerly available and is represented in the root stub. No public API removal or rename is evident.
• Implementation: The subprocess tests directly exercise clean-interpreter namespace imports and selected symbol imports, checking that Torch, Torch Elastic, and unrelated RL frameworks are not loaded eagerly. The changelog fragment matches the affected source package. Installed-wheel inclusion of the runtime-consumed .pyi files remains the main packaging surface to validate, but the supplied patch does not establish a concrete packaging failure.

No blocking issues. No inline issue met the actionable-evidence threshold; the assessment above records the review feedback.

Automated review; human maintainers own approval decisions.

[StafaH]

run-ci

[kellyguo11]

run-ci