[DataFlow runtime] Domino disaggregated training (example + launcher + strategy params + capture/tracker fixes)#660

Open

maocheng23 wants to merge 24 commits into

dataflow-up-16-zerocopyfrom

domino-disagg

Commits on Jul 13, 2026

[DataFlow runtime] Phase B1 — TargetEngine ABC + de-EAGLE3 the target boundary

authored and
jiapingW
committed
[DataFlow runtime] Phase B2 — decouple the target engine from the sglang version

authored and
jiapingW
committed
[DataFlow runtime] E0 — layout consolidation: runtime/ is substrate-only (move-only)

authored and
jiapingW
committed
[SpecForge] E — draft-architecture registry (@register_draft)

authored and
jiapingW
committed
style: fix draft registry import ordering

maocheng23
authored and
jiapingW
committed
[SpecForge] Phase E — typed Pydantic config + specforge CLI

authored and
jiapingW
committed
fix(sglang): make the DataFlow capture backend work on sglang 0.5.14

authored and
jiapingW
committed
fix(sglang): port parallel-state init in patch.py to sglang 0.5.14

authored and
jiapingW
committed
support domino disaggregated training[bug]
jiapingW
committed
fix bugs
jiapingW
committed
1. dflash family draft model abstract

王佳平
authored and
jiapingW
committed
fix lint

王佳平
authored and
jiapingW
committed
[DataFlow runtime] Domino disagg: 1-server + DP=7 launcher + working --num-epochs

maocheng23
authored and
jiapingW
committed
support dspark training
jiapingW
committed
support train multi epochs
jiapingW
committed
disagg logging: DP-average loss/acc + log train/accuracy & train/lr

authored and
jiapingW
committed
examples: dflash 1-server + DP=7 disagg launcher (auto LR-schedule steps)

authored and
jiapingW
committed
examples: add two-node DFlash launcher

maocheng23
authored and
jiapingW
committed
refactor domino and dspark code
jiapingW
committed
fix some bugs
jiapingW
committed
fix lint and grad norm clip bugs[PR663]
jiapingW
committed
feat(data): add validated Qwen ShareGPT regeneration pipelines

jianuo-huang
authored and
jiapingW
committed
feat(data): explode multi-turn reasoning into generation events

jianuo-huang
authored and
jiapingW
committed
test(domino): add generic overfit and DFLASH serving gates

jianuo-huang
authored and
jiapingW
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[DataFlow runtime] Domino disaggregated training (example + launcher + strategy params + capture/tracker fixes)#660

[DataFlow runtime] Domino disaggregated training (example + launcher + strategy params + capture/tracker fixes)#660
maocheng23 wants to merge 24 commits into
dataflow-up-16-zerocopyfrom
domino-disagg

Commits on Jul 13, 2026

[DataFlow runtime] Phase B1 — TargetEngine ABC + de-EAGLE3 the target boundary

[DataFlow runtime] Phase B2 — decouple the target engine from the sglang version

[DataFlow runtime] E0 — layout consolidation: runtime/ is substrate-only (move-only)

[SpecForge] E — draft-architecture registry (@register_draft)

style: fix draft registry import ordering

[SpecForge] Phase E — typed Pydantic config + specforge CLI

fix(sglang): make the DataFlow capture backend work on sglang 0.5.14

fix(sglang): port parallel-state init in patch.py to sglang 0.5.14

support domino disaggregated training[bug]

fix bugs

1. dflash family draft model abstract

fix lint

[DataFlow runtime] Domino disagg: 1-server + DP=7 launcher + working --num-epochs

support dspark training

support train multi epochs

disagg logging: DP-average loss/acc + log train/accuracy & train/lr

examples: dflash 1-server + DP=7 disagg launcher (auto LR-schedule steps)

examples: add two-node DFlash launcher

refactor domino and dspark code

fix some bugs

fix lint and grad norm clip bugs[PR663]

feat(data): add validated Qwen ShareGPT regeneration pipelines

feat(data): explode multi-turn reasoning into generation events

test(domino): add generic overfit and DFLASH serving gates