Fix TP6 virtual padding for GLM MTP drafts

[voipmonitor]

Summary

• restore virtual-TP padding for independently constructed MTP draft model configs
• apply the draft padding plan before draft parallel-config validation
• add a regression test that verifies padding precedes the divisibility check

Root cause

The GLM target config was padded from 64 to 66 attention heads in VllmConfig.__post_init__, but SpeculativeConfig validates the independently constructed DeepSeekMTPModel first. The v17 stack contained the draft-padding test but omitted the corresponding helper and validator call, so TP6 worked with MTP off and failed with MTP enabled.

Validation

• 19 passed in tests/config/test_virtual_tp.py
• Ruff check and format check passed
• exact config-only reproduction now passes for TP6, DCP6, MTP3 with both target and draft at 66 heads / 2112 MoE intermediate size

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

🗂️ Base branches to auto review (1)

• dev/*

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 Plus

Run ID: 751da9f4-c1ba-4183-9cfe-72f06853d6b8

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

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch codex/fathomless-firmament-v17-tp6-mtp-fix-20260715

---

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]

Validated after rebasing directly onto dev/fathomless-firmament:

• PR head: 522c626de89b629a18d05db21ba02b5acf6e6f30
• focused tests: 19 passed
• real lukealonso/GLM-5.2-NVFP4 config at TP6/DCP6/MTP3: target and draft both resolve to 66 attention heads / 2112 MoE intermediate size
• exact-v17 source image carrying the equivalent fix: voipmonitor/vllm:fathomless-firmament-v17-vllm05f50ae-b12x1377d5f-fi801d57a-cu132-20260715
• manifest: sha256:9b6f1ab6db4d3a7b7b786481eb32abe82e86d185648d62c3ac1cfa6d72a55e47

The release uses commit 05f50ae79c48835275f22f76e8dfb10b0024dec6 only to retain the exact v17 integration parent; the mergeable fix remains this single direct-FF PR.

[voipmonitor]

/review

[voipmonitor]

Superseded by d437a23cfd (Add reusable virtual TP sharding for odd TP sizes) already present on dev/fathomless-firmament. The canonical implementation is broader (MTP and DSpark) and includes the TP6 draft-before-validation regression coverage from this PR. Closing to keep the FF merge queue unambiguous.