Repository navigation
Conversation
…proofs The OoO-Spec-style drafting program is cut; both lines died on their pre-registered criteria (ARCHITECTURE.md Negative Knowledge has the numbers): - Free-run MTP drafting: held-out 76.1% first-token / 2.17-of-8 free-run vs the ~95% bar. The denoiser already self-drafts the whole canvas. - Step-simulator seeding: killed by the oracle control - seeding the EXACT real step-1 argmax converges slower than noise (8.60 vs 7.13 steps). Kept, per the standing caveat that disproofs are re-testable: - mtp_head / mtp_head_gpu (+ .metal): the 0.4B assistant head on CPU f32 and Metal, bit-exact against each other, 12.2 ms/token GPU. - Engine capture: per-step canvas hiddens/argmax in run_denoise_step, accumulate-across-chunks prefill hidden capture, capture_mtp_states (fresh capturing prefill + BackboneKv snapshot decode). - mtp-dump subcommand and the ignored mtp_* probe harnesses (draft probe, head-to-head, step dump, step1 seed probe with oracle/prior controls). - python/mtp/: rollout trainer, step-simulator/step1 trainers, OpenRouter teacher relay, HF importer, drift-test scripts. Gates: cargo test 789/789, smoketest 17/17, golden 8/8 byte-identical. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
`cargo fmt --check` fails on main. 881cae4 landed unformatted and no CI run exists for that sha, so the gate never saw it and every branch cut from main inherits a red fmt job. Reflow only, from `cargo fmt --all` on the pinned 1.96.0. `cargo build --release` is warning-free and the suite is 789/789. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Owner
Author
|
Not needed. |
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.
cargo fmt --checkfails onmain.881cae46(current main HEAD) landed unformatted, and there is no CI run forthat sha, so the fmt gate never saw it. Every branch cut from main since then
inherits a red fmt job, which is how this surfaced: it is the only failure on
#25.
Reflow only, from
cargo fmt --allon the pinned 1.96.0 toolchain. Thebyte-level changes are trailing commas and bracket placement where rustfmt
split an over-long line.
RUSTFLAGS="-D warnings" cargo build --releaseis clean andcargo test --releaseis 789 passed, 0 failed.Merging this first turns #25 green.
🤖 Generated with Claude Code