fix(kda): preserve gate alias contract during GLM decode

[voipmonitor]

Behavior

When a B12X KDA decode output overlaps the live gate tensor, vLLM binds a disjoint caller-owned output, runs recurrence, then copies the result into the requested output. Disjoint outputs retain the existing allocation-free path. Other B12X binding failures still propagate.

Reason and compatibility

B12X correctly rejects a mutable output that overlaps its read-only raw_g input. A reported GLM-5.3-Flash TP2/DCP2/MTP3 service crashed under concurrent long-context traffic with exactly that error. This change preserves the B12X alias contract and existing output semantics; the exceptional alias case adds one temporary output and copy.

Validation

• A forced-alias unit reproducer checks that B12X sees a disjoint output, reads the unchanged gate, and returns the requested result. Four targeted model tests passed in a GPU-capable environment; Ruff passed.
• A patched-source GLM-5.3-Flash-NVFP4-Spark TP2/DCP2/MTP3 server on two RTX PRO 6000 GPUs completed full and piecewise CUDA graph capture, a health check, and overlapping decode with a roughly 30k-token cold prefill.
• The reported 850k-token, three-concurrent-request workload was not reproduced. This is bounded correctness validation, not a claim that every long-context failure is fixed.

Report: https://discord.com/channels/1466898002793857221/1471527895439638528/1551693471558017085

Summary by CodeRabbit

• Bug Fixes
  • Improved gated attention decoding when output buffers overlap with read-only gate data.
  • Preserved gate data during decoding while continuing to propagate unrelated binding errors.
  • Added coverage to verify correct results and buffer handling in this scenario.

[coderabbitai]

Understand this PR’s impact

Explore downstream dependencies and potential security impact with Blast Radius.

View blast radius →

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Advanced

Run ID: 9c2265f5-49d9-4d62-8585-ad8905f9f737

📥 Commits

Reviewing files that changed from the base of the PR and between 9e5d179 and 0aa1e08.

📒 Files selected for processing (2)

• tests/models/test_glm5next_model.py
• vllm/model_executor/layers/mamba/gdn/kimi_gdn_linear_attn.py

Included review availability: Your plan provides up to 1 included review per hour; 0 remain after this review.

---

📝 Walkthrough

Walkthrough

The B12X KDA decode path now handles output buffers that overlap the read-only raw_g tensor. A regression test verifies temporary output binding, gate preservation, and final result copying.

Changes

B12X KDA alias-safe decode

Layer / File(s)	Summary
Alias-safe decode and regression coverage vllm/model_executor/layers/mamba/gdn/kimi_gdn_linear_attn.py, tests/models/test_glm5next_model.py	The decode binding uses a temporary destination when the output overlaps raw_g. Other ValueError exceptions remain unchanged. The test verifies that raw_g stays unchanged and that the result is copied back.

Priority: ➖ Normal

Estimated code review effort: 2 (Simple) | ~12 minutes

Change: Bug fix

Suggested reviewers: lukealonso

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 12.50% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 8 functions across 2 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
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.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely describes the main change: preserving the gate alias contract during GLM decode in KDA.

• Fix all pre-merge checks with AI

✨ Finishing Touches 💡 1

📝 Generate docstrings 💡

• Commit to this branch
• Create a new PR

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

---

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.}

[voipmonitor]

Superseded by #821, which now contains this PR’s history, gate/output alias correction, tests and immutable release fragment. The consolidation preserves the exact alias-safe implementation already qualified in beta and resolves its overlap with recovery-record binding. Merge #798, then #821; do not reapply this separate branch. Composed QSA/GDN/warmup/alias checks: 112 passed, nine declared skips.