SpecDec: approve native heads together with their checkpoint

[arekay-nv]

Summary

Under v0.7, speculative decoding was [Limited Use]: permitted only for DeepSeek-R1, per MLPerf Inference v6.0 (#65). #112 replaced that gate with a curated approved-drafter list per benchmark, seeded by the benchmark task force (#79). §2.9.4 now also says that a drafter shipped with the canonical checkpoint but absent from the list may not be used.

So far only the Agentic Inference reference publishes a list (mlcommons/endpoints#494 and mlcommons/endpoints#519). DeepSeek-R1, the one benchmark v0.7 allowed speculative decoding for, has none, so its native MTP head is now disallowed for v1.0. The same happens to any future model that ships a native head before someone files a list entry for it.

This PR approves a head the model publisher ships inside an approved checkpoint together with that checkpoint. It is the "always OK" category #65 parked as [SPECDEC-0DAY] future work, narrowed to heads published with the target's own weights, and it keeps every §2.9.4 eligibility, verification and disclosure rule.

Section	Change
§2.9.1	The reference-implementation drafter bullet notes that native heads of approved checkpoints are approved with them and need not be listed
§2.9.4	New Native heads paragraph: approved with the checkpoint; approval recorded against the checkpoint's approval (model-list) cohort, so the two-cohort lead time runs from there; weight-identified by the checkpoint's model ID and checksum; everything else unchanged. The "v1.0 change", drafter-selection and Leaving the drafter unused sentences are reworded to match
§3.2	Native heads of listed checkpoints need not appear on the published approved-drafter list
§9.1	The Approved drafter check also matches a native head by its checkpoint; Drafter approval lead time counts a native head from its checkpoint's approval cohort

Rationale

Why native heads need no separate review. The curated list exists to check three things: open weights, no input-based optimization, and no training for benchmark performance. A head the publisher releases inside the model's own checkpoint, trained for the model in general, meets them for the same reasons the checkpoint itself is approved. Reviewing it separately adds nothing beyond the checkpoint's approval.

Consistency with §2.9.3. §2.9.3 already requires a derived checkpoint to keep every component of the canonical one, naming MTP and EAGLE-style heads explicitly. The head is already part of the approved, provenance-checked artifact; this PR lets a submission use it, not just carry it.

Lead time. The model list is published at least six weeks, three cohorts, before a round opens (§3.2). A head approved with its checkpoint therefore clears the two-cohort lead time for the round in which the checkpoint enters.

What does not change.

• A separately published drafter still needs an approved-list entry. Examples: a third-party EAGLE head, or the Kimi K3 DSpark heads.
• Exact verification still applies.
• A drafter may not be modified, apart from PTQ with disclosure.
• The per-point disclosure and same-drafter-across-the-curve requirements still apply.

Effect on current benchmarks. DeepSeek-R1's native MTP head (deepseek-ai/DeepSeek-R1) becomes approved. The agentic README's checkpoint-native entries (Qwen3.6-35B-A3B MTP, DeepSeek-V4.1-Flash DSpark) become redundant but remain valid. Benchmarks whose checkpoints ship no head are unaffected.

Notes

• §2.9.4 is [TENTATIVE — Subject to change after 2026-10-12], so this targets the open revision window.
• No headers are renamed, so existing anchors still resolve. git diff --check is clean.
• Related: Speculative Decoding Rules - Update for v1.0 #65 (parent), Speculative Decoding Rules Update for v1.0 - Approval Process Drafters #79 (approval process, seeding), SpecDec v1.0: curated approved drafter lists, drafter equivalence via existing accuracy-gate/input-based-optimization rules #112 (curated lists).

🤖 Generated with Claude Code

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[nvzhihanj]

Don't seem we need to expand here given these are already stated in other sections. I would recommend write it as a "blacklist", meaning what is not allowed with native heads drafters

[nvashutoshd]

Agreed.

@arekay-nv pls update and I can approve and merge this.

[nvashutoshd]

@arekay-nv

Pls also add the list of approved heads from the endpoints repo here
Preferably as a link, so we dont have two different list.s

[arekay-nv]

@arekay-nv

Pls also add the list of approved heads from the endpoints repo here Preferably as a link, so we dont have two different list.s

The approved drafters are maintained here. There is also an open PR to add legacy models there.