[Test] Cover tensors excluded from WeightChecker reset

[yuchenwang3]

Motivation

#40777 already made reset, comparison, and checksumming share _is_skip_weight_check, covering the fix originally proposed here. The existing reset tests do not cover a parameter marked _skip_weight_check.

Modifications

Keep a regression in TestResetTensors that marks a parameter as excluded, calls _reset_tensors, and checks that its value is preserved. The redundant production guard has been removed.

Testing

Rebased onto main 10edb05e72afb79ecfdf7dbe7603f90ba05ad3ee on October 3.

• All applicable changed-file pre-commit hooks pass.
• The existing regression body passes against the actual reset implementation using a manual CPU fixture. This bypasses the CUDA fixture and the constructor; it is a focused reset check.
• The registered CUDA suite has not been rerun on this revision. The earlier 63-test GPU result belongs to cad0e99dc, before this rebase.

Accuracy Tests

Validation checks preservation of the excluded parameter during reset.

Speed Tests and Profiling

No performance benchmark; this revision adds test coverage only.

Checklist

• Changed-file pre-commit hooks pass.
• Regression added to the existing registered test suite.
• Documentation applicability reviewed; this revision adds no API or configuration.
• Validation scope stated above.
• Followed the existing test style.

AI assistance: OpenAI Codex rebased the branch, removed the redundant guard, and ran local validation.

---

CI States

Latest PR Test (Base): ❌ Run #37160496178
Latest PR Test (Extra): ❌ Run #37160495962
Latest PR Test (AMD ROCm 10): ❌ Run #37160496146

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[yuchenwang3]

@fzyzcjy pinging you since weight_checker.py is mostly your code. This fixes _reset_tensors resetting tensors that _compare/_compute_checksum already exclude via _skip_weight_check (e.g. fp8 KV-cache scales) — the self-test corrupts state the weight update never restores. One-file change, open for ~3 weeks without review. Could you take a look?

[yuchenwang3]

Rebased this onto current main and added a regression case in the existing TestResetTensors suite. The test covers a tensor marked _skip_weight_check and verifies reset leaves it untouched. @fzyzcjy, could you take a look and add run-ci if this direction looks right?

[yuchenwang3]

@b8zhong @ch-wan, could one of you review/triage this? You both recently reviewed or merged #27692 in the same WeightChecker path.

I have now rebased the PR onto current main; the current head is 3cd3ae989d83a097b2ded652d9538341dfa36154. The targeted reset symmetry harness passes, current-head lint is green, and the red checks are PR-gate failures with Missing required label 'run-ci'; the actual CPU/GPU suites were skipped.

The change makes _reset_tensors honor the same _skip_weight_check marker already honored by _compare and _compute_checksum, so unchecked derived tensors are not randomized and then left unverified. If the direction looks right, could you review it and run /tag-and-rerun-ci extra?

[yuchenwang3]

@fzyzcjy, following up on this small WeightChecker fix: reset should honor the same _skip_weight_check marker as comparison and checksumming. The regression test is included, but there is still no review. Could you review it or point me to the current owner, and authorize CI if appropriate? Thanks.

[b8zhong]

/tag-and-rerun-ci

[yuchenwang3]

CI follow-up for current head cad0e99dc560b33289406a114809b4a02ff8158e:

• The GPU WeightChecker job ran test_weight_checker.py: 63 tests passed, including test_preserves_tensors_excluded_from_weight_check.
• The MLX shutdown failure also reproduces on this revision's base (a98d921658b2cb78ca1257a4d72a6a3969620456). On Apple Silicon, both TestOverlapLoopGracefulExit tests raise the same _StopLoop on base and head; applying the existing [MLX] Fix the overlap-loop shutdown tests after unified request intake #39917 test-fixture patch to an isolated base checkout makes both pass. The scheduler and fixture blobs are identical between base and head.
• The B200 failure is TestB200NumaTopology.test_gpu_numa_mapping, which assumes the visible four GPUs share one NUMA node; the runner reports {0, 2}. That test is also unchanged by this PR, and [CI] Check B200 NUMA mapping against sysfs numa_node #40055 has since merged the sysfs-based correction. I have not rerun B200 locally.

These findings cover those two failures, not every red check. The branch is unchanged. Could you review the WeightChecker change and advise whether to refresh CI once the pending MLX fixture fix lands?

Assisted by OpenAI Codex for log inspection, the local base/head comparison, and this write-up.

[yuchenwang3]

@b8zhong Rebased onto 10edb05e72 at a3f146f34c. #40777 already implemented the reset exclusion through _is_skip_weight_check, so I removed the redundant guard and kept the missing regression. This PR now changes only the test. Changed-file pre-commit checks and the regression body with a manual CPU fixture pass; the CUDA suite has not been rerun on this head. Could you review the remaining test coverage?

The current Intel image-build job failed before tests with no space left on device. The extra workflows also stop at the optional run-ci-extra label gate.

The AMD EAGLE3 job fails during RoPE graph capture with expected scalar type Half but found Float. The scheduled main job at ab9d8f029c12 has the same failure in TestBasicSanityEagle3.setUpClass. The HIP fallback retains the FP32 cache introduced by f6fcda8827, while rotary_embedding dispatches on the FP16 query and reads the cache as that type. This PR only adds nine lines to the WeightChecker test; it does not change either path. These logs do not establish a current-head pass for the WeightChecker CUDA suite.

[yuchenwang3]

@yueming-yuan Since #40777 covers the reset logic, this now only adds the missing regression test. Could you take a look?