feat(mllm-v2): Support single-controller for Megatron-Inference. - #4009

Merged

terrykong merged 19 commits into

mainfrom

cye/rl_mllm_omni_mm_nocolgym_rebase_pr2957

Sep 7, 2026

Commits on Sep 7, 2026

Pass pre-expanded media tokens to MLLM.
cspades
committed
5bb259b
Add tests and lint.
cspades
committed
62da428
Address more review feedback.
cspades
committed
2152f02
Support multimodal Megatron-Inference with NeMo-RL TransferQueue (V2).
cspades
committed
6cf8a44
WAR fix for multimodal request checkpoint embedding recompute.
cspades
committed
7bce7df
Address review feedback and incorporate code from #2957.
cspades
committed
2ee0d37
Add SC L1 and nightly tests.
cspades
committed
f7e892e
Fix rebase errors to reenable V2 multimodal with MCore.
cspades
committed
5568079
Get rid of rebase artifacts.
cspades
committed
e4ee67a
Restore Gym revision required by multimodal branch
cspades
committed
11f2eb7
Revert "Restore Gym revision required by multimodal branch"
cspades
committed
91286e1
Address more review comments.
cspades
committed
319aaa1
Strengthen reserved port handoff test
cspades
committed
f87affb
Lint.
cspades
committed
56acacf
Fix rebase error.
cspades
committed
81c99e8
Fix flaky unit tests and test coverage.
cspades
committed
2b064d2
Add a digest to the RL venv cache to automatically refresh stale venv's.
cspades
committed
32fbaef
fix(sc): give vLLM workers the nemo_gym extra by default instead of swapping the env at runtime
terrykong
committed
3e547fa
Revert the cache digest and refresh.
cspades
committed
82a0dc5