feat(diffusion): add SFT loss hub and pre-encoded data manager - #90

Merged

zhihengy merged 19 commits into

feat/encoder-hubfrom

feat/sft

Aug 8, 2026

Commits on Aug 7, 2026

feat(diffusion): add SFT loss hub and pre-encoded data manager
zhihengy
committed
de4c35a
feat(diffusion): add frame stride to SFT encode script
zhihengy
committed
4c9ff87
feat(diffusion): auto-build SFT cache via family encode hooks
zhihengy
committed
1616581
fix(sft): per-sample content-addressed cache, atomic writes, seeded VAE sampling
zhihengy
committed
e1e8710
refactor(sft): plug into RolloutManager via rollout-function/convert/log hooks with lazy encoder pool
zhihengy
committed
ce29764
feat(sft): image datasets for Wan (single-frame media branch)
zhihengy
committed
2fb0a6c
refactor(sft): pick dual-expert via uniform anchor index
zhihengy
committed
e5d7afc
refactor(sft): seat encoder pool via the rollout placement view
zhihengy
committed
0f026e8
chore(sft): encoder pool inherits the engine's 0.3 GPU share
zhihengy
committed
abcc73a
refactor(sft): reuse --diffusion-height/width/output-num-frames; add --fsdp-flow-shift

zhihengy
and
claude
committed
34f1cb2
refactor(sft): validate explicit plugin paths instead of auto-wiring defaults

zhihengy
and
claude
committed
d974b27
refactor: make --train-only an explicit flag; --debug-train-only becomes its alias

zhihengy
and
claude
committed
f1ceb9d
refactor(sft): move frozen-encoder logic to encoder_hub; TPC back to training-only

zhihengy
and
claude
committed
795e5d7
style: black-format the train_only exclusion assert

zhihengy
and
claude
committed
cd44bab
feat(sft): support disaggregated encoder GPUs

zhihengy
and
cursoragent
committed
68c88f8
fix(ci): lazy-load rollout manager from placement helper

zhihengy
and
cursoragent
committed
4d2450c
Revert "fix(ci): lazy-load rollout manager from placement helper"
zhihengy
committed
8f792a0
fix(ci): expose SGLang source package to CPU tests

zhihengy
and
cursoragent
committed
6ebacad
fix(sglang): defer diffusion-only imports until engine startup

zhihengy
and
cursoragent
committed
f049028

Navigation Menu

Uh oh!

feat(diffusion): add SFT loss hub and pre-encoded data manager - #90

feat(diffusion): add SFT loss hub and pre-encoded data manager#90
zhihengy merged 19 commits into
feat/encoder-hubfrom
feat/sft

Commits on Aug 7, 2026

feat(diffusion): add SFT loss hub and pre-encoded data manager

feat(diffusion): add frame stride to SFT encode script

feat(diffusion): auto-build SFT cache via family encode hooks

fix(sft): per-sample content-addressed cache, atomic writes, seeded VAE sampling

refactor(sft): plug into RolloutManager via rollout-function/convert/log hooks with lazy encoder pool

feat(sft): image datasets for Wan (single-frame media branch)

refactor(sft): pick dual-expert via uniform anchor index

refactor(sft): seat encoder pool via the rollout placement view

chore(sft): encoder pool inherits the engine's 0.3 GPU share

refactor(sft): reuse --diffusion-height/width/output-num-frames; add --fsdp-flow-shift

refactor(sft): validate explicit plugin paths instead of auto-wiring defaults

refactor: make --train-only an explicit flag; --debug-train-only becomes its alias

refactor(sft): move frozen-encoder logic to encoder_hub; TPC back to training-only

style: black-format the train_only exclusion assert

feat(sft): support disaggregated encoder GPUs

fix(ci): lazy-load rollout manager from placement helper

Revert "fix(ci): lazy-load rollout manager from placement helper"

fix(ci): expose SGLang source package to CPU tests

fix(sglang): defer diffusion-only imports until engine startup