Support consistent application checkpoint recovery and safe sampling replay - #749

Merged

lfengad merged 9 commits into

nvidia-cosmos:mainfrom

shuangwu:feat/controller-resume-adapter

Sep 21, 2026

Commits on Sep 18, 2026

feat(resume): add instance-owned controller metadata and sampler adapter
shuangwu
committed
ae1584a
test: verify application checkpoint optimizer and sampling parity
shuangwu
committed
be8ac93
test: cover epoch and worker agreement in external resume
shuangwu
committed
d6c865a
test(controller): publish distributed checkpoint restart canary
shuangwu
committed
fbe311e
docs: define deterministic checkpoint requirements and current limits
shuangwu
committed
ecdf089
docs: scope checkpoint recovery to consistent resumption
shuangwu
committed
184e302
feat: support committed application checkpoints and conservative replay
shuangwu
committed
7fe6af8
test: preserve resume agreement checks with fail-fast errors
shuangwu
committed
c00b577
style: format terminal drain test fixture
shuangwu
committed
1930038