[DataFlow runtime] Domino end-to-end + StepContext for schedule-dependent loss#629
Merged
Merged
Conversation
Contributor
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
maocheng23
added a commit
that referenced
this pull request
Jun 30, 2026
… W3′ naming Review fixes (verified against the files): - Status (confirmed): stop calling the in-review composable-launch stack (#627/#628/#629) "landed"/"DONE"/"done". Split the genuinely-merged spine from the in-review stack in §1; one consistent "in review" label in §1/Phase A/success table and across the roadmap (README, Phase A). Leave the spine's "landed" wording (it is merged). - Module placement (confirmed): Evaluator/EvalCache are top-level domain managers (specforge/eval/), not specforge/runtime/eval/ — fix the eval-and-breadth.md outlier to match plan.md §2.3 and domain-refactor.md. - W3′ naming (confirmed): SGLangServerEngine is ONE engine with two feature transports (capture-into-FeatureStore for W3/O1.3, inline-HTTP for the light W3′) — disambiguate in §2.2, the workload table and §G2 rather than overloading one name. - O1.3 spike (reviewer's premise refuted — it is already an explicit 🔴 gate): added the valid narrow point instead — the spike scopes only the sglang_server slice of Phase B; the de-EAGLE3 extraction and domain Trainer carry no engine risk. Additional contradictions found by a completeness sweep and fixed: - StrategySpec registry: plan.md said it "stays in runtime/training unchanged" but §6 + Phase E move it — clarify the per-step strategy seam stays, the registry converges into training/strategies/. - TargetEngine source: extracted from modeling/target/*TargetModel (adapters wrap it), not "absorbs runtime/inference adapters". - Draft package: models/drafts is the target layout; note today's modeling/draft/ + real filenames. - Dependency graph: align domain-refactor (E depended on {C,D}) with README (D→E, C parallel). - Drop the up-19/up-20 branch tags that only appeared in the online doc. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Contributor
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
Collaborator
Author
Code reviewNo high-confidence issues found. Checked for bugs in the Domino StrategySpec, StepContext threading, lambda schedule parity with the script, and offline/online launcher wiring. |
…dent loss
Domino is the third algorithm on the composable launch — a StrategySpec plus the
ONE genuine shared-contract extension the analysis predicted.
- strategy.py: StepContext{global_step, total_steps} threaded into forward_loss
(optional; eagle3/dflash ignore it). DominoTrainStrategy: reuses the DFlash
feature schema + adapter; its forward_loss reads ctx to compute the decaying
lambda_base that blends Domino's base loss (mirrors train_domino.get_lambda_base).
- trainer.py: TrainerCore.train_step / eval_step accept a StepContext; fit passes
StepContext(global_step, total_steps=max_steps). Backward-compatible.
- contracts.py: DraftStrategyName += "domino".
- registry.py: domino spec — reuses DFlash transform/collate/adapter, domino reader
(strategy tag) + DominoTrainStrategy. No new builder, no launch.py change.
- tests/_fixtures.py: build_domino (DFlash draft w/ projector_type="domino" head ->
OnlineDominoModel).
- tests/test_domino_launch.py (new): CPU lambda-schedule test + offline/online GPU
end-to-end.
Adding domino touched ZERO launch.py and reused the dflash data path — exactly the
"new algorithm = a spec + its loss" goal.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
maocheng23
force-pushed
the
dataflow-up-22-dflash
branch
from
6945bfe to
a6b8b7d
Compare
maocheng23
requested review from
FlamingoPg,
shuaills and
sleepcoo
as code owners
maocheng23
force-pushed
the
dataflow-up-23-domino
branch
from
64eb276 to
9c1c020
Compare
Collaborator
Author
|
Addressed review feedback (self-review pass) — includes one real bug fix.
Deferred: exposing Validated: full |
jiapingW
approved these changes
Jul 3, 2026
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
2 participants
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.
What
Domino — the third algorithm on the composable launch. Beyond a
StrategySpecit needs exactly one shared-contract extension:StepContext.Changes
strategy.py:StepContext{global_step, total_steps}threaded intoforward_loss(optional; eagle3/dflash ignore it).DominoTrainStrategyreuses the DFlash feature schema + adapter; itsforward_lossreadsctxto compute the decayinglambda_basethat blends Domino's base loss (mirrorstrain_domino.get_lambda_base).trainer.py:TrainerCore.train_step/eval_stepaccept aStepContext;fitpassesStepContext(global_step, total_steps=max_steps). Backward-compatible.contracts.py:DraftStrategyName += "domino".registry.py: domino spec — reuses DFlash transform/collate/adapter + a domino reader;DominoTrainStrategy. No new builder, nolaunch.pychange.tests/_fixtures.py:build_domino(DFlash draft w/projector_type="domino"head →OnlineDominoModel).tests/test_domino_launch.py(new): CPU lambda-schedule test + offline/online GPU end-to-end.Adding domino touched zero
launch.pyand reused the dflash data path — the "new algorithm = a spec + its loss" goal.StepContextis the one genuine, deliberate contract change (vs. leaking schedule state through ad-hoc kwargs).Testing
Part of the 197 tests OK suite run at this tip (sci-h200 / H200), incl. the new domino offline/online GPU tests + the CPU schedule test.
Stacked on the dflash PR. Part 3/3.
🤖 Generated with Claude Code