feat(e5): MTP-aux distillation rung — registered + implemented - #114
Merged
Conversation
EXPERIMENTS.md E5: control = run-006-r7-muon verbatim, single delta = a 3-step multi-token-prediction head (~1.1M params, 0.4%) on the student decoder, aux CE at beta 0.15, head discarded at inference so the shipped artifact stays vanilla and size-identical. Source: Hy4's native MTP mapped as a TRAINING auxiliary (TODO 07-hy4) — serving-side speculation stays parked, decode measured non-binding. gpu/mtp.py (MTPHead + build/mtp_named), distill_sequence wiring (head before optimizer so Muon sees its matrices; mtp_head.pt saved beside student.pt for resume; provenance-only copy under best/), spec ara-diac-small-2-mtp -> run-007-r7-muon-mtp. 3 unit tests (shift/mask semantics, uniform-logit baseline, naming). Gate (pre-agreed): adopt at <=4.5218 full-set windowed DER; honest report in [4.5218, 4.8218). Prediction: 4.5-4.75.
ronaldtse
pushed a commit
that referenced
this pull request
Sep 2, 2026
…still The F821 was a real latent bug: MTP checkpoint lines from the #114 wiring were pasted into the logit-mode distiller save block where mtp_head never exists — any logit-mode run would NameError at its first step checkpoint. The lines belong only in distill_sequence. The rest: import sorting, unused zipfile, quoted annotation, E501 signature wraps, zip strict=, context manager for models.yaml.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Pre-registered in EXPERIMENTS.md (gate before launch): control = run-006-r7-muon verbatim, single delta = a 3-step multi-token-prediction head on the student decoder (~1.1M params / 0.4 percent), aux CE at beta 0.15. The head is discarded at inference — the shipped student stays vanilla and size-identical; mtp_head.pt is saved beside student.pt for training resume and copied under best/ for provenance only. Source: Tencent Hy4's native MTP mapped as a TRAINING auxiliary (TODO 07-hy4; serving-side speculation stays parked — decode measured non-binding at our sizes). Spec: ara-diac-small-2-mtp -> run-007-r7-muon-mtp. 3 unit tests cover shift/mask semantics, the uniform-logit baseline, and Muon-facing param naming. Gate: adopt at <=4.5218; honest report in [4.5218, 4.8218). Prediction: 4.5-4.75.