chore: nightly sync main into dev (16_05_2026) - #4831

Closed

svcnvidia-nemo-ci wants to merge 51 commits into

devfrom

main2dev/16_05_2026

Commits on May 10, 2026

Create a Protocol for the MLP layer of TransformerLayer (#3435)

nschank
and
asolergi-nv
authored
5e31514

Commits on May 11, 2026

Revert "Add Python-side guardrail for HybridEP InfiniBand limit and rename seq_len (#4094)" (#4718)
ko3n1g
authored
a2ec5c1
chore(beep boop 🤖): Bump (main) (2026-05-11)
github-actions[bot]
committed
e93755e
Add Python-side guardrail for DeepEP IB limits (#4719)
janEbert
authored
ad58411
ci: revert bad uv.lock bump and label future bumps with `Run functional tests` (#4730)
ko3n1g
authored
5123f6a
[ci] fix: treat cancelled run-main-script step as failure (#4727)
ko3n1g
authored
33d47e0
ci: Major refactor of release-workflows (#4602)

ko3n1g
and
claude
authored
e42e2fa
build(deps): bump nvidia-modelopt to 0.43 (#4723)
ko3n1g
authored
434368c
fix(fsdp): recognize legacy GDN TP metadata (#4664)

Glitchfix
and
cspades
authored
74687fe
Fixes for Nemotron3 Super release test config (#4544)

maanug-nv
and
ko3n1g
authored
9718f7d
feat(gpt): add output postprocess hook (#4686)
Glitchfix
authored
97f3bce
Add bump-base-image skill and update golden value comparison (#4733)

balasaajay
and
cursoragent
authored
f744215
Guard omegaconf imports (#4685)
maanug-nv
authored
6486d52

Commits on May 12, 2026

Commits on May 13, 2026

Commits on May 14, 2026

fix: use no_mask in local ViT layer spec (#4395)

Phlip79
and
claude
authored
1ba0aa9
refit clean up and refactoring (#4762)

wdykas
and
William Dykas
authored
20bf831
Make weight and optimizer memory estimation take into account expert parallelism correctly (#4687)
YangFei1990
authored
a7c9e8c
Support recomputing in HybridModel (#4496)
xuantengh
authored
118933a
One single flag that determines if we are in inference (#4617)
tdene
authored
925422c
[main] feat(moe): Support packed sequence for gated delta net (GDN) (#2645)

authored
2d1fa8d
remove dead manual_release_grads code path in 1F1B overlap schedule (#4511)

Wohox
and
asolergi-nv
authored
58d45bf
Fix recompute checkpointing + training CGs (#3919)
tdene
authored
ada8dfe
Use Protocols to type-check linear_proj submodules of Attention (#3434)

authored
dbfc96b
fix tokenizers in respect to newer transformers (#4608)
dimapihtar
authored
d167123
Bump nvidia-modelopt>=0.44.0 (#4803)
kevalmorabia97
authored
98031e1
Update owners (#4794)
Phlip79
authored
266562f
ci: Update workflow to use same commit for building docker image and running tests (#4787)
balasaajay
authored
e9a0930
chore: Update nightly tests golden values (#4805)
balasaajay
authored
a53107c
Inference: Optimize Prefill Engine Steps for Nemotron (#4764)
sidsingh-nvidia
authored
9b4074b

Commits on May 15, 2026

Strengthen test_checkpoint to verify distributed checkpoint behavior (#4711)

authored
493e2be
Combine GEMM + SwiGLU fused MLP PRs (3890, 4071, 4095, 4219, 4311, 4324) → main (#4636)

authored
6815c0f
Disable MSC by default; opt in via --enable-msc (#4629)
asolergi-nv
authored
95bba94
[fix] Use MSC for checking checkpoint existence (#4251)

pavelgein
and
asolergi-nv
authored
f2070ee
ci: tolerate git-gc race in /home/runner chown after checkout (#4808)
balasaajay
authored
b3b6719
Reorder mtp_post_process after attention backward in 1F1B schedule plan (#4695)

gdengk
and
Gao Deng
authored
212f885
additional tests for nvrx (#4522)

dimapihtar
and
claude[bot]
authored
f9952cb

Commits on May 16, 2026

Commits on May 17, 2026

fix: post-CI corrections
svcnvidia-nemo-ci
committed
cd222aa