[Test] Check GLM-5.3 end-to-end logits alongside loss

[xs1997zju]

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}

Stack #2112 (bottom to top):

1. [Feature] Add GLM-5.3-Flash F0: 25B cropped reference checkpoint builder #2105 — F0: cropped reference checkpoint
2. [Feature] Add GLM-5.3-Flash F3: Kimi Delta Attention (KDA) #2106 — F3: KDA
3. [Feature] Add GLM-5.3-Flash F4: mHC four-stream residual #2107 — F4: mHC
4. [Feature] Add GLM-5.3-Flash F5: NoPE DSA + KPool indexer + clamped SwiGLU #2108 — F5: NoPE DSA
5. [Feature] Add GLM-5.3-Flash F1: VL data preprocessing pipeline #2109 — F1: VL data preprocessing
6. [Feature] Add GLM-5.3-Flash F2: vision tower + projector (eager) #2110 — F2: vision tower and projector
7. [Feature] Add GLM-5.3-Flash F6 core: text model + MTP + compose model #2111 — F6: text and compose model
8. [Test] Check GLM-5.3 end-to-end logits alongside loss #2139 — F7: end-to-end loss and logits parity ← you are here

Base is #2111's branch (feat/glm53flash-f6-text-moe). Review this PR's own diff; the seven lower layers are dependencies.

---

The existing TestGlm53TextMoEAccuracy.test_fsdp_accuracy compares native HF and XTuner forward loss on the reduced GLM-5.3 checkpoint. Retain that test and its loss assertion, and additionally compare full-vocabulary logits at the final eight next-token positions for each existing text sample.

The test still uses the same checkpoint, sentences and EP1/4/8 parameterization. Capture HF logits during its existing forward; request XT logits with loss_ctx=None in an additional no-grad forward so the public training-loss path remains covered. Both models run in eval mode. There is no backward or parameter update.

The small metric helper checks shape/finiteness and computes relative L2, cosine and maximum absolute error with CPU FP64 reductions. Each sample must satisfy relative L2 <5% and cosine >0.998, independently of the unchanged loss-curve thresholds. Six CPU checks cover identity, shape/nonfinite failures, position permutations, and a common logits offset that leaves CE unchanged but must fail the logits check.

Validation:

• Existing real-checkpoint accuracy test with the added logits assertions: 3 passed, EP1/4/8 on 8×H200. These retained files are unchanged from that GPU-validated version.
• Metric regressions rerun after scope reduction: 6 passed.
• Diff whitespace check passed.

This PR contains only the existing text test change, the metric helper and its regression tests. The separate multimodal replay/data-preparation scripts and documentation are excluded. No model implementation, preprocessing or training recipe is changed. Base is #2111; the vision RoPE fix is already in the underlying stack.

[xs1997zju]

Superseded by #2141, an ordinary PR targeting the existing F6 layer (#2111), combining the logits and image parity tests without adding a stack layer.