feat(distill): Engram run wiring — attach, Sinkhorn table routing, spec - #220
Merged
Merged
Conversation
attach_engram: one table on the encoder at block N, addresses computed from input_ids per forward (no dataloader change), zero-init projection = identity at step 0. engram_param_split routes the table to the Sinkhorn-balanced update (5x lr) and the projection to Muon. Wired at BOTH distill paths' student construction - main dispatches by spec mode, and the first launch proved the lesson: the attach only in distill() left the sequence-mode run training a vanilla replica (no [engram] attached line, identical param counts). The launch gate now demands the attach line as positive evidence before a run counts. Spec ara-diac-small-2-1-engram: single variable vs run-007 (2.1, 4.5701) - one 2M x 32 table at block 3, canonical r7 labels (e70ce991, pre-seeded after the lite-2.0 re-labeling lesson), 6ep sequence-KD Muon. 9 attach specs: identity at step 0, memory reaches logits, param split, addressing, export, ONNX gather.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The training-side wiring for TODO.impl/10's run (launched; verdict PR follows when the gate completes).
maindispatches by spec mode, and the first launch proved why that matters: the attach only indistill()left the sequence-mode run training a vanilla replica — no attach line, identical param counts. The launch gate now demands the[engram] attachedline as positive evidence before a run counts.ara-diac-small-2-1-engram: single variable vs run-007 (4.5701) — one 2M x 32 table at block 3, canonical r7 labels (sha e70ce991, pre-seeded), 6ep sequence-KD Muon.9 new specs (attachment identity, memory flow, param split, addressing, export, ONNX gather); full Engram suite 15 green.