feat: async colocated GRPO with Megatron inference - #2884

Merged

yuki-97 merged 13 commits into

NVIDIA-NeMo:mainfrom

tdene:tde/async_colocated

Aug 21, 2026

Commits on Aug 19, 2026

feat: async colocated GRPO with Megatron inference
tdene
committed
d6d5d93
Drop unnecessary synchronize

tdene
and
anlthms
committed
b469e9b
Defer colocated wake past checkpoint saves

tdene
and
anlthms
committed
deeac40
Allow inference_optimized during training

tdene
and
anlthms
committed
e33e49c
Add functional test
tdene
committed
67d8e1f
Handle nightly tests
tdene
committed
b1e36e5
Remove the NEED_REFIT carve-out
tdene
committed
48d05e5
Handle rebase conflicts
tdene
committed
7b3a56a

Commits on Aug 20, 2026

Address PR team review
tdene
committed
3fd09e2
Fix CI
tdene
committed
b3b51bb
Address reviewer comment
tdene
committed
8772e04
Fix CI
tdene
committed
ab0c137
Fix CI
tdene
committed
a53fc02