[Multi-GPU] Refactors multi-gpu logic into module, fix ghost process on keyboard interrup

[StafaH]

Description

Refactors multi-gpu logic from train_multigpu into isaaclab_rl

Checklist

• I have read and understood the contribution guidelines
• I have run the pre-commit checks with ./isaaclab.sh --format
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• I have added a changelog fragment under source/<pkg>/changelog.d/ for every touched package (do not edit CHANGELOG.rst or bump extension.toml — CI handles that)
• I have added my name to the CONTRIBUTORS.md or my name already exists there

[greptile-apps]

Greptile Summary

The PR moves multi-GPU launcher construction into isaaclab_rl, introduces a shared per-rank training worker, and adds process-group supervision intended to prevent workers surviving an interrupt.

• Replaces the script-local launcher implementation with a package API and thin compatibility shim.
• Resolves the packaged worker entrypoint for torchrun and skrl JAX launches.
• Adds POSIX process-group signal forwarding, escalation, cleanup, and shutdown tests.
• Leaves the non-process-group fallback unable to clean up worker descendants.

Confidence Score: 4/5

The fallback shutdown behavior should be fixed before merging because interrupted distributed runs can still leave worker processes and GPU allocations behind on platforms without process groups.

POSIX shutdown is substantially improved, but the newly retained fallback ends only the direct launcher process and provides no mechanism to terminate or reap its distributed worker descendants.

Files Needing Attention: source/isaaclab_rl/isaaclab_rl/entrypoints/multigpu.py

Important Files Changed

Filename	Overview
source/isaaclab_rl/isaaclab_rl/entrypoints/multigpu.py	Centralizes command construction and adds robust POSIX group supervision, but its fallback terminates only the immediate launcher and can leave workers alive.
source/isaaclab_rl/isaaclab_rl/entrypoints/train_worker.py	Provides a packaged worker entrypoint while preserving Warp configuration ordering and torch elastic traceback recording.
scripts/reinforcement_learning/train_multigpu.py	Becomes a thin compatibility shim over the new package-level multi-GPU API.
source/isaaclab_rl/test/test_multigpu_shutdown.py	Covers POSIX process groups, interrupt cleanup, forced escalation, and exit codes, but skips the fallback path entirely.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    CLI[train_multigpu CLI] --> Check{os.killpg available?}
    Check -->|Yes| Group[Start launcher in new session]
    Group --> Forward[Forward signal to process group]
    Forward --> Escalate[Escalate and reap descendants]
    Check -->|No| Fallback[Start direct child]
    Fallback --> Terminate[Terminate direct launcher only]
    Terminate --> Orphans[Worker descendants may survive]

Loading

_{Reviews (1): Last reviewed commit: "Stop multi-GPU Ctrl-C from leaving worke..." | Re-trigger Greptile}

[greptile-apps]

Fallback leaves workers running

When multi-GPU training is interrupted on a platform without os.killpg, this handler terminates only the direct torchrun or skrl launcher and performs no descendant cleanup or escalation, causing worker processes to continue running and retain GPU resources after the CLI exits.

[isaaclab-review-bot]

Isaac Lab Review Bot

The launcher refactor centralizes multi-GPU command construction and process supervision in isaaclab_rl.entrypoints, but the single-process training shim no longer sets Warp’s enable_backward flag before importing the entrypoints package, weakening the import-order guarantee the previous implementation explicitly preserved.

• Design and architecture: Moving launcher logic into the package and retaining the script as a compatibility shim provides a cleaner shared entrypoint. The new worker module is a reasonable per-rank target, but importing it through isaaclab_rl.entrypoints executes the package initializer before the worker can configure Warp.
• API: The existing multi-GPU CLI behavior is retained through run_train_multigpu_cli, which is exported from both isaaclab_rl.entrypoints and the package-level lazy API. No incompatible CLI change is evident in the supplied patch.
• Implementation: The process-group supervision, signal escalation, command building, and shutdown regression tests directly address orphaned workers. However, train.py now imports isaaclab_rl.entrypoints.train_worker before the worker sets wp.config.enable_backward = False; Python executes isaaclab_rl.entrypoints.__init__ first, so the flag should remain configured in train.py before that import to preserve the stated ordering guarantee.

Minor fixes needed. Posted 1 actionable finding inline.

Automated review; human maintainers own approval decisions.

[isaaclab-review-bot]

🟡 Warning · Implementation — Warp flag set after entrypoints package import

from isaaclab_rl.entrypoints.train_worker import main first executes isaaclab_rl/entrypoints/__init__.py (importing api, dispatch, multigpu); only afterwards does train_worker's body run wp.config.enable_backward = False. The removed code deliberately set the flag before that exact import, so the ordering guarantee the new comment claims no longer holds on the single-process path. Set the flag in train.py before importing the worker.