Skip to content

[DataFlow runtime] DFlash end-to-end on the composable launch (offline + online)#628

Merged
jiapingW merged 1 commit into
dataflow-up-16-zerocopyfrom
dataflow-up-22-dflash
Jul 2, 2026
Merged

[DataFlow runtime] DFlash end-to-end on the composable launch (offline + online)#628
jiapingW merged 1 commit into
dataflow-up-16-zerocopyfrom
dataflow-up-22-dflash

Conversation

@maocheng23

@maocheng23 maocheng23 commented Jun 30, 2026

Copy link
Copy Markdown
Collaborator

What

DFlash trains end-to-end (offline + online) through the composable launch from the parent PR — via a StrategySpec entry + a DFlashAdapter, with ZERO launch.py changes.

Changes

  • registry.py: dflash spec — offline reader (OfflineManifestReader with dflash feature_keys, no aux/target swap), per-sample transform, padding collate; online via DFlashAdapter; supports_online=True.
  • specforge/runtime/inference/dflash_adapter.py (new): wraps generate_dflash_data, emits {input_ids, hidden_states, loss_mask}. verify_capture self-skips the eagle3 aux/target checks (different feature names + __aux_layer_ids__=None).
  • tests/_fixtures.py: write_offline_files_dflash + build_dflash (tiny Qwen3 target → DFlash draft + TargetEmbeddingsAndHead → OnlineDFlashModel).
  • tests/test_dflash_launch.py + test_dflash_online_launch.py (new, GPU): offline and online dflash train end-to-end through FSDP.

Note

DFlash is online-only in production today (no offline dumper exists — prepare_hidden_states.py is eagle3-only), so the offline path is exercised with synthetic fixtures while online is its real workflow.

Testing

Part of the 197 tests OK suite run at the stack tip (sci-h200 / H200).

Stacked on the composable-launch PR. Part 2/3.

🤖 Generated with Claude Code

@gemini-code-assist

gemini-code-assist Bot commented Jun 30, 2026

Copy link
Copy Markdown
Contributor

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

@maocheng23 maocheng23 mentioned this pull request Jun 30, 2026
Merged
maocheng23 added a commit that referenced this pull request Jun 30, 2026
@maocheng23 @claude
docs: address plan/roadmap review — status wording, module placement,…
fedb337 
… 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>
@maocheng23 maocheng23 marked this pull request as ready for review June 30, 2026 23:26
@maocheng23 maocheng23 requested a review from FrankLeeeee as a code owner June 30, 2026 23:26
@gemini-code-assist

gemini-code-assist Bot commented Jun 30, 2026

Copy link
Copy Markdown
Contributor

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

@maocheng23

maocheng23 commented Jun 30, 2026

Copy link
Copy Markdown
Collaborator Author

Code review

No high-confidence issues found. Checked for bugs in the DFlash adapter, offline transform/collate path, strategy registration, and online rollout-to-train wiring.

@maocheng23 @claude
[DataFlow runtime] DFlash end-to-end on the composable launch (offlin…
a6b8b7d 
…e + online)

DFlash now trains through the runtime via a StrategySpec + a DFlashAdapter, with
ZERO launch.py changes (the spec seam from the previous commit carries it).

- registry.py: dflash spec — offline reader (OfflineManifestReader with dflash
  feature_keys, no aux/target swap), per-sample transform, padding collate; online
  via DFlashAdapter; supports_online=True.
- inference/dflash_adapter.py (new): wraps generate_dflash_data, emits
  {input_ids, hidden_states, loss_mask}; verify_capture self-skips the eagle3
  aux/target checks (different feature names + __aux_layer_ids__=None).
- tests/_fixtures.py: write_offline_files_dflash + build_dflash (tiny Qwen3 target
  -> DFlash draft + TargetEmbeddingsAndHead -> OnlineDFlashModel).
- tests/test_dflash_launch.py + test_dflash_online_launch.py (new, GPU): offline
  and online dflash train end-to-end through FSDP.
- tests/test_strategy_registry.py: dflash-fully-wired assertions.

DFlash is online-only in production (no offline dumper exists yet — prepare_
hidden_states.py is eagle3-only), so the offline path is exercised with synthetic
fixtures while online is its real workflow.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@maocheng23 maocheng23 force-pushed the dataflow-up-21-composable-launch branch from 457b20a to 8faf111 Compare July 1, 2026 00:29
@maocheng23 maocheng23 force-pushed the dataflow-up-22-dflash branch from 6945bfe to a6b8b7d Compare July 1, 2026 00:29
@maocheng23

maocheng23 commented Jul 1, 2026

Copy link
Copy Markdown
Collaborator Author

Addressed review feedback (self-review pass).

  • dflash_adapter.py: dropped the redundant \"__aux_layer_ids__\": None emit (the RolloutWorker reads it via feats.pop(..., None), so an absent key is identical); made the loss_mask default lazy (only allocate the all-ones mask when a task omits it, not eagerly per sample).

Deferred: extracting the length-grouped batching shared with SGLangAdapter into a base helper — it touches the pre-existing SGLangAdapter / validated eagle3 online path, so noted for a follow-up.

Validated: full tests/test_runtime = 200 OK (2 skipped, 1 xfail), zero failures, on a 2-node H200 pod. Lint clean (black 24.10.0 / isort 5.13.2 / autoflake).

Base automatically changed from dataflow-up-21-composable-launch to dataflow-up-16-zerocopy July 2, 2026 05:41
@jiapingW jiapingW merged commit 14a18ba into dataflow-up-16-zerocopy Jul 2, 2026
1 check passed
@jiapingW jiapingW deleted the dataflow-up-22-dflash branch July 2, 2026 05:47
@maocheng23 maocheng23 mentioned this pull request Jul 4, 2026
Open
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@jiapingW jiapingW jiapingW approved these changes
@FrankLeeeee FrankLeeeee Awaiting requested review from FrankLeeeee FrankLeeeee is a code owner
@shuaills shuaills Awaiting requested review from shuaills
@sleepcoo sleepcoo Awaiting requested review from sleepcoo
@FlamingoPg FlamingoPg Awaiting requested review from FlamingoPg

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@maocheng23 @jiapingW