[doc] fix: note the megatron commit MTP needs for recompute_granularity=full

[gaohongkui]

What does this PR do?

Follow-up to #7326, where the review outcome was that verl should document the required Megatron fix rather than carry a compatibility shim. This does that.

docs/advance/mtp.md pins megatron dev at 23e092f41 (2025-12-09) for MTP + CP. That commit predates ffd66a3e6 ("Roll input IDs for MTP labels", NVIDIA/Megatron-LM#3457, 2026-06-03) by about six months, and #3457 is what threads padding_mask through MultiTokenPredictionLayer._checkpointed_forward.

Without it, MultiTokenPredictionLayer.forward passes padding_mask= to a _checkpointed_forward that does not declare the parameter, so MTP with recompute_granularity=full raises a TypeError on the first step — i.e. following this doc as written still crashes for that combination. Released megatron-core 0.18.0 and 0.18.2 do not carry the fix either, tracked upstream as NVIDIA/Megatron-LM#4933.

Checklist Before Starting

• Search for similar PRs. Queries run: mtp.md megatron dev commit, mtp recompute_granularity, 3457 megatron — no open PR touches this.
• Format the PR title as [{modules}] {type}: {description}

Test

Documentation only, so the thing to verify is that the factual claim is right. I checked the signature of MultiTokenPredictionLayer._checkpointed_forward at each ref by parsing the file with ast rather than grepping — grep is misleading here because the call site a few lines below the signature also contains the string padding_mask, which is how I first got this wrong:

ref	date	padding_mask in signature
4c6360260	2026-05-22	no
ffd66a3e6 (#3457)	2026-06-03	yes — first commit that has it
23e092f41 (currently pinned in this doc)	2025-12-09	no
core_v0.18.0	2026-06-22	no
core_v0.18.2 (newest release)	2026-07-20	no
dev @43124b60c	2026-08-08	yes
main @6518b75ec	2026-08-09	yes

The bisect was done over the file's commit history via gh api "repos/NVIDIA/Megatron-LM/commits?path=megatron/core/transformer/multi_token_prediction.py&sha=main", then fetching the blob at each candidate.

Repo checks run: python3 tests/special_sanity/check_docs_time_info.py (✅), PR_TITLE=... python3 tests/special_sanity/check_pr_title.py (✅). Both added links return HTTP 200.

API and Usage Example

No API change.

Design & Code Changes

One bullet in docs/advance/mtp.md §1 "Scope of Support" gains the additional constraint for recompute_granularity=full, plus the Last updated date.

I deliberately did not move the pin itself. I have not validated MTP + CP on a newer dev commit, and I would rather add the constraint than silently replace a pin someone else verified. If you have a newer dev commit you trust for MTP + CP, bumping it would make this bullet simpler and I'm happy to do that instead.

Checklist Before Submitting

• Read the Contribute Guide.
• Apply pre-commit checks — pre-commit was not installable in my environment; I ran the repo's own sanity checks directly (above). Note that pre-commit (3.12) is currently failing on main for an unrelated file (tests/trainer/ppo/test_reinforce_pp_multiturn_on_cpu.py, a ruff-format diff from [algo] fix: carry running_return through observation spans in REINFORCE++ (#7278) #7300), so a red pre-commit here would not be from this change.
• Add / Update the documentation — this PR is the documentation change.
• Add unit or end-to-end test(s) — not applicable to a docs-only change.

---

AI assistance disclosure: prepared with AI assistance (Claude). I reviewed every changed line and verified each factual claim above myself.