[WIP] Controller on datapacker - #2132

Open

PengchengShi00 wants to merge 11 commits into

InternLM:mainfrom

PengchengShi00:controller_on_datapacker

Commits on Sep 20, 2026

support value head
PengchengShi00
committed
ff5283a
[Refactor] Move train data preparation into AgentLoop and TrainingController
YanhuiDua
committed
4bb5b80

Commits on Sep 21, 2026

[Refactor] Unify response_ids suffix convention and phase-split rollout conversion
YanhuiDua
committed
7c29de9

Commits on Sep 22, 2026

[Refactor] Write controller-facing train meta fields at canonicalization
YanhuiDua
committed
57c92a8
[Feature] Add RLTrainItem and worker-side rollout conversion
YanhuiDua
committed
da7970c
[Feature] Add index-only RLDataPacker with legacy schedule equivalence
YanhuiDua
committed
22e8746
[Refactor] Switch to worker-side pack materialization with per-DP plan dispatch
YanhuiDua
committed
aed49b5
[Refactor] Remove legacy controller packing path and sync tests
YanhuiDua
committed
35784b6
[Refactor] Move loss_ctx build into worker convert and pack stages
YanhuiDua
committed
5bb416e

Commits on Sep 24, 2026

merge branch pr-2099 add critic
PengchengShi00
committed
c3725bf

Commits on Sep 29, 2026

add ppo training
PengchengShi00
committed
ad163d4