fix(snapshot): exit vLLM source without engine teardown

[galletas1712]

Summary

• terminate the source vLLM process directly after a successful initial
  Snapshot capture;
• avoid unwinding and destroying the checkpointed CUDA engine during source
  shutdown; and
• leave the restored path unchanged.

Motivation

This is a generic Dynamo Snapshot/vLLM lifecycle fix, independent of GMS V1.

The initial source path raises SystemExit after successful capture. That
unwinds vLLM and may enter CUDA and EngineCore teardown after the process image
has already been checkpointed. In the observed failure, teardown produced a
source-process SIGSEGV even though capture itself had succeeded.

The successful initial-capture boundary does not need Python cleanup: the
source container is deliberately being replaced by a restore-standby
container. This change therefore exits directly with os._exit(0), matching
the existing TensorRT-LLM Snapshot lifecycle. The restored path still returns
normally.

This fix is split from #12011 because it applies to Dynamo Snapshot with or
without GMS.

Validation

Focused test

The focused unit test covers both lifecycle branches:

• successful initial capture exits directly with status 0; and
• restored execution returns without exiting.

nscale Snapshot E2E

The fix passed in the final graph-enabled Qwen3-0.6B TP1/DP1 Snapshot + GMS V1
proof on nscale-dev, node s2877.

The source engine logged:

Initial vLLM snapshot captured successfully; exiting without destroying the engine

Its terminal state was:

exitCode=0
reason=Completed

The replacement entered restore standby, Snapshot reported
nvidia.com/snapshot-restore-status.engine=completed, restore completed in
2.674 seconds, and the restored engine served five coherent requests with
cached_tokens=0.

The previous destructor-path SIGSEGV did not recur. No CUDA fault, illegal
memory access, late free, engine crash, or restore retry was observed.

Qualification build:
https://github.com/ai-dynamo/dynamo/actions/runs/30227329178.

Summary by CodeRabbit

• Bug Fixes
  • Improved the snapshot capture completion flow to exit cleanly after a successful restore.
  • Added a confirmation message when the vLLM snapshot is captured successfully.

[datadog-official]

🎯 Code Coverage (details)
• Patch Coverage: 100.00%
• Overall Coverage: 40.64% (-10.28%)

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: 58272b1 | Docs | Datadog PR Page | Give us feedback!}

[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.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 1dc563b1-3d4f-4730-b993-1417135c0e13

📥 Commits

Reviewing files that changed from the base of the PR and between 329aba5 and 6207d3b.

📒 Files selected for processing (1)

• components/src/dynamo/vllm/snapshot.py

---

Walkthrough

The vLLM snapshot success path now logs completion and terminates the process with os._exit(0) after snapshot restoration.

Changes

Snapshot restore termination

Layer / File(s)	Summary
Restore completion and process exit components/src/dynamo/vllm/snapshot.py	Imports os, logs successful snapshot capture, and replaces SystemExit(0) with os._exit(0) after restore completion.

Estimated code review effort: 2 (Simple) | ~5 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Description check	⚠️ Warning	The description covers summary, motivation, and validation, but it misses the required reviewer-start section and the required issue-linking section.	Add the required template sections: Overview, Details, Where should reviewer start?, and Related Issues with either a Closes/Relates reference or the no-issue checkbox.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly matches the main change: vLLM source now exits directly after snapshot capture to avoid engine teardown.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

---

_{Comment @coderabbitai help to get the list of available commands.}

[galletas1712]

/ok to test 6207d3b

[galletas1712]

/ok to test f02b4cb

[galletas1712]

/ok to test 7b55ac9

[galletas1712]

/ok to test d8a06c3

[galletas1712]

/ok to test b1f88b2

[galletas1712]

/ok to test afc98d1

[galletas1712]

/ok to test 58272b1