Enable training cudagraphs for RL - #2452

Merged

mathemakitten merged 25 commits into

NVIDIA:mainfrom

mathemakitten:helenn-training-graphs-rl

Jan 14, 2026

Commits on Dec 1, 2025

Turn on training cudagraphs for RL
mathemakitten
committed
8ae06df

Commits on Dec 2, 2025

Add data pointer comparison to tensors in ArgMetadata
mathemakitten
committed
97266e3
Merge branch 'main' into helenn-training-graphs-rl
mathemakitten
committed
8cd1808

Commits on Dec 5, 2025

persist cudagraphs
mathemakitten
committed
e59966f

Commits on Dec 8, 2025

Gate training cudagraph persistency via -–rl-persist-cuda-graphs
mathemakitten
committed
836bda6
Fix typo in args
mathemakitten
committed
c8903d9
persist cudagraphs
mathemakitten
committed
4185aac

Commits on Dec 9, 2025

Update to rl_training_cuda_graphs
mathemakitten
committed
b966303
Merge main
mathemakitten
committed
fd156c3

Commits on Dec 15, 2025

Commits on Dec 17, 2025

Include Peter's change on RNG state tracking safety for cudagraphs
mathemakitten
committed
b271f28

Commits on Jan 12, 2026

merge main
mathemakitten
committed
cae2cb9
Merge
mathemakitten
committed
52600fb
Flip training cudagraphs back on after refit MR
mathemakitten
committed
075530f

Commits on Jan 13, 2026

Fix RL sequence packing bin size (#2909)

tdene
authored and
mathemakitten
committed
3e6fafc
If empty bin, give it a default PackedSeqParams so the signature will match for cudagraphs
mathemakitten
committed
c0882a4
runs fast need refactor
mathemakitten
committed
3d71442
Cleanup
mathemakitten
committed
67b9f3b
Merge branch 'main' into rl-training-graphs
mathemakitten
committed
792310a
Maximize --rl-sequence-packing-max-sequences-per-bin for now
mathemakitten
committed
db1bae4