Fix: a2a3 finalize calls rtDeviceReset on the rt-only path (cascade fix)

[hw-native-sys-bot]

Summary

Fixes the chronic `st-onboard-a2a3` cascade we've been seeing across many recent PRs.

Symptom

CI shows the same shape repeatedly: one test errors out (varies — `test_dedup_shared_so_independent_unregister`, `spmd_paged_attention`, etc.), then every subsequent test on the same device fails its `simpler_init` with `rtStreamCreate failed: 507899`. Reproduced locally with the full ST suite (`pytest examples tests/st --platform a2a3 --device 0-7 -n 2 --pto-isa-commit $PTO_ISA_COMMIT`).

Root cause

a2a3's `DeviceRunner::finalize()` gated `aclrtResetDevice` on `acl_ready_`, which is only set by `ensure_acl_ready()` — i.e. only the HCCL/comm test path. Non-HCCL tests (most of the suite) skipped any form of device reset on teardown. When a non-HCCL test left the device in error state (AICPU exception, stuck kernel, anything), the next `ChipWorker` on the same device tripped `rtStreamCreate`.

a5 had been doing the right thing all along: its `finalize` calls `rtDeviceReset(device_id_)` unconditionally. This change mirrors that on a2a3 for the non-ACL path. The asymmetry is exactly why a5 onboard has been ~100% reliable in CI while a2a3 onboard has chronically cascaded.

The fix

`finalize()` now always resets the device:

• `acl_ready_=true` (HCCL/comm path) — unchanged: `aclrtResetDevice` + `aclFinalize` (tears down both device + the ACL bring-up).
• `acl_ready_=false` (rt-only, the common case) — NEW: `rtDeviceReset(device_id_)`. Previously skipped entirely.

Diff: 1 file, +34/-14.

Local validation (a2a3 onboard, device 1)

• `vector_example` with pinned PTO-ISA: PASS.

(A full ST-suite repro of the cascade requires xdist parallelism + device pool exhaustion timing that CI provides naturally. CI is the real validation gate for this fix.)

Expected effect in CI

• The chronic cascade on `st-onboard-a2a3` should stop. A test that genuinely fails (e.g. PTO-ISA HEAD AICPU exception on spmd_paged_attention) still fails that specific test, but the next test recovers cleanly — no 507899 chain.
• If the cascade does stop, this is also the answer to the long-standing "what's the state-teardown bug?" investigation that's been open all session.

Test plan

• CI green on st-onboard-a2a3 (the real signal).
• No regression on st-onboard-a5 (this PR doesn't touch a5).

[coderabbitai]

Important

Review skipped

Auto incremental reviews are disabled on this repository.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 720d1132-015f-4cc4-a418-3bdbca2a333a

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

📝 Walkthrough

Walkthrough

DeviceRunner::finalize() is modified to reset devices via two conditional paths after freeing device memory: ACL-ready paths call aclrtResetDevice() and aclFinalize() with error logging, while non-HCCL rt-only paths call rtDeviceReset() instead, ensuring both runtime types have equivalent cleanup.

Changes

Device Finalization Reset

Layer / File(s)	Summary
Dual-path device reset in finalization src/a2a3/platform/onboard/host/device_runner.cpp	DeviceRunner::finalize() conditionally resets devices based on acl_ready_ state: ACL devices use aclrtResetDevice/aclFinalize with error logging, rt-only devices use rtDeviceReset. Both paths now log failures and propagate errors via rc.

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~20 minutes

Poem

🐰 Devices cleaned up in two fine ways,
ACL and rt-only both get their days,
Two reset paths, one finalization call,
Runtime cleanup handles it all! 🎉

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately describes the main change: adding rtDeviceReset call on the rt-only path in a2a3's finalize, which directly addresses the chronic cascade issue detailed in the PR.
Description check	✅ Passed	The description is comprehensively related to the changeset, explaining the symptom, root cause, the fix, and validation—all directly relevant to the code modification.
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.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[gemini-code-assist]

Code Review

This pull request updates the DeviceRunner::finalize() method in device_runner.cpp to ensure the device is reset even when acl_ready_ is false (the pure runtime path). When acl_ready_ is true, it continues to perform aclrtResetDevice and aclFinalize. When acl_ready_ is false, it now calls rtDeviceReset to clear any remaining per-device runtime state, preventing subsequent device initialization failures. There are no review comments to address.