[hhhappyshow]

Summary

• rsl_rl wraps the whole PPO rollout in torch.inference_mode(). With deferred finite validation (the default since Follow-up: clean up remaining contract-less defensive code patterns #2591's defer_finite_validation default flip), the NaN guard's device latches (_device_bad_rows, _device_verdict) are typically first allocated inside that context and are therefore inference tensors.
• Poll sites — training cycle boundaries and close() → poll_finite_validation() — run outside inference mode, so reset_verdict()'s in-place zero_() raises RuntimeError: Inplace update to inference tensor outside InferenceMode is not allowed at teardown of every deferred-validation PPO run.
• Fix: run the latch reset under torch.inference_mode() explicitly. The reset touches no autograd bookkeeping, so behavior is unchanged for plain tensors, and inference tensors reset cleanly at every poll site.

Discovered by the downstream perceptive navigation port (Motphys/Perceptive-UniLab staging): the first full PPO train smoke against current main crashed in close() after one successful iteration.

Linked Work

• Issue: none (found through downstream integration testing)
• Base branch: main

Validation

• Focused tests passed on the final local head
• make test-all: in progress separately; see commands below

Commands actually run:

uv run pytest tests/training/test_tensor_diagnostics.py -q            # 11 passed (with fix)
# regression check: stashing the src fix makes the new test fail with the
# exact runtime error above, then passes again after restore
make check                                                            # ruff + mypy + pyright clean

Remote CI route:

• Base main: will run on this PR's head.

Impact

• Backend impact: none (training-loop diagnostics only)
• Platform impact: unknown
• Training effect expected: no (fixes teardown crash; numerics unchanged)

Artifacts

• W&B: none / benchmark result: none / video: none / checkpoint: none

Checklist

• Added or updated tests where needed (test_reset_verdict_allows_inference_tensor_latches fails without the fix)
• Updated docs if behavior or workflow changed (n/a; inline comment documents the contract)
• Linked the driving issue (none exists)
• Noted any follow-up work explicitly (downstream perceptive navigation port continues against this fix)

Made with Cursor