chore: nightly sync main into dev (07_05_2026) - #4684

Closed

svcnvidia-nemo-ci wants to merge 113 commits into

devfrom

main2dev/07_05_2026

Commits on Apr 22, 2026

Fix nvtx_decorator to check _nvtx_enabled at call time (#4184)

authored
9a3c927
fix merges_file typo in megatron_hf_tokenizer (#4392)

chelseajohn
and
john2
authored
60f71e1
Enable NullTokenizer for pretraining to reduce I/O access (#4057)
asolergi-nv
authored
c9dfe34
docs: Add SECURITY.md (#4431)
chtruong814
authored
7073492
Mamba inference opt (#4414)

authored
40627d0
DDP refactoring: Extract parameter layout computation into optimizer classmethod (#3812)

deepakn94
and
claude
authored
55b8111
Update PR template with explicit request for issue (#4409)
Phlip79
authored
90e09b6

Commits on Apr 23, 2026

Misc inference fixes (#4397)
sidsingh-nvidia
authored
ab2b33d
Rename Mamba to Hybrid outside megatron/core (#4159)

Phlip79
and
claude
authored
60408d5
Include mtp layers in token per expert logging (#4412)
Mellonta
authored
a52014c
fix: NVRx async compatibility and defer resiliency import (#4420)

sbak5
and
dimapihtar
authored
32275b2
ci: add base_sha to codecov/codecov-action upload step (#4445)
ko3n1g
authored
9bb35a8

Commits on Apr 24, 2026

Commits on Apr 25, 2026

Fix multivalidation (#3388)
RPrenger
authored
241a5ca
Add missing knob for reduce_scatter_with_fp32_accumulation (#4410)

WanZzzzzz
and
asolergi-nv
authored
f2dcd42

Commits on Apr 26, 2026

Enable CUDA graphs for MTP inference (#4260)
santhnm2
authored
03f4111

Commits on Apr 27, 2026

Commits on Apr 28, 2026

Commits on Apr 29, 2026

Update copy-pr-bot.yaml [skip ci]
github-actions[bot]
committed
533dc75
[training migration] Add serialization features to config container (#4309)
maanug-nv
authored
1c4e537
Fix conflict with inference graphs (#4504)
tdene
authored
f4a49cf
chore: rotate oncall schedule
github-actions[bot]
committed
251c6e9
Add tools/prepare_cache.py for offline GPT dataset cache preparation (#4080)

asolergi-nv
and
claude
authored
c5201a0
[build] fix: move mamba-ssm and causal-conv1d to optional [ssm] extra (#4517)

ko3n1g
and
claude
authored
cb3d5d9
mamba: avoid redundant HBM reloads in causal_conv1d_update shift loop (#4460)

authored
4e208a8
Standardize misc graph interface (#4485)

tdene
and
mathemakitten
authored
3f59bbb
Fix inference graph override in RL flow (#4323)
tdene
authored
29864b2
Unify and refactor Megatron-FSDP documentation. (#4418)

cspades
and
megnvidia
authored
b23aa3f
Revert "ci: add base_sha to codecov/codecov-action upload step (#4445)" (#4526)
chtruong814
authored
51ea07e
Skills for running unit tests and working with slurm (#4502)
yashaswikarnati
authored
cfee04e
Reorganize order of operations in inference context and text generation controller (#2929)
tdene
authored
0d98cb8

Commits on Apr 30, 2026

Commits on May 1, 2026

Update copy-pr-bot.yaml [skip ci]
github-actions[bot]
committed
3460bba
Adding code for Flextron (#4429)

authored
2d862fe
Fix partial cudagraphs + HybridEP not properly triggering DDP hook (#4500)
jiemingz
authored
3b1521e
Ignore pytorch link anchors (#4582)
maanug-nv
authored
9776b58
MoE dispatcher fixes: size NVLS dispatcher buffers from actual tensor sizes (#4576)

mathemakitten
and
ericharper
authored
4e0f636

Commits on May 2, 2026

Finalize all builders in preprocess_data, not just the last key (#4573)

sayalinvidia
and
claude
authored
74c857b
refactor(skills): add when_to_use frontmatter, split ci-test-system, enforce skill workflow (#4574)

ko3n1g
and
claude
authored
a6cf566
Make last_token_logits graphable (#4552)

tdene
and
santhnm2
authored
396bee1
fix(ci): correct off-by-one in total_steps_evaluated formula (#4591)

ko3n1g
and
claude
authored
0031752
Add fault injection support via nvidia_resiliency_ext. (#4370)

hexinw-nvidia
and
ko3n1g
authored
0afcfbf
Guard vocab reduce_scatter on TP > 1 (#4565)
mathemakitten
authored
cf736dc

Commits on May 3, 2026

Commits on May 4, 2026

Enable shared expert overlap with allgatherv in inference (#4570)
sidsingh-nvidia
authored
442a936
Add vLLM grouped gemm backend for MoE inference (#4566)

authored
bb979dd
Move KD teacher loading to after Float16Module (#4394)

authored
99abdc8
ci: update gpt3_7b_tp4_pp1_memory_speed gb200 golden values (#4601)
ko3n1g
authored
0efa47a
Fix inference unit test (#4589)

maanug-nv
and
ko3n1g
authored
0c479ee
Checkpoint conversion between GPT_model and Hybrid_model (#4482)
guihong-nv
authored
cf21d70
ci: add cadence input for test filtering in CI workflows (#4561)
balasaajay
authored
c8fde51
Handle SSM sharded tensor merge OOM with CPU fallback (#4442)

returnL
and
maanug-nv
authored
fa9c714
Fix `mtp_use_repeated_layer` behavior for GPT models (#3965)

authored
2194f51

Commits on May 5, 2026

FlashInfer sampling (#2456)

tdene
and
santhnm2
authored
878228f
Fix main2dev workflow (#4610)

Phlip79
and
claude
authored
7924242
Add logic to enable chunked MLP during training (#3656)
pengdurice
authored
f4a0710
Inference bug-fixes: Re-enable EP syncs for the legacy A2A dispatcher and re-simplify ep_sync accidentally reverted by #4306 (#4587)

sidsingh-nvidia
and
claude[bot]
authored
c817dad
Remove invalid `timeout` argument for dist.barrier (#4512)

authored
0b2b572
Fix buffers in refit (#4580)
wdykas
authored
4397e07
Named validation sets (#4578)
RPrenger
authored
4858caf
Fix Hang in tests (#4575)
wdykas
authored
ae65776
Single commit for main2dev nightly (#4614)

Phlip79
and
claude
authored
b819ac7
convert tokenizer args to config (#4406)
dimapihtar
authored
40d024b
Siddharth/fix ep sync (#4607)

authored
190c833
mmiranda working on another set of broken links (#4534)
megnvidia
authored
6e5fb47

Commits on May 6, 2026

Fix gradient corruption with layerwise param all-gather overlap (#4609)

deepakn94
and
claude
authored
b25a76e
test: mark TestFusedApplyMLARope::test_forward_backward_for_q flaky_in_dev (#4639)
ko3n1g
authored
c325855
chore: rotate oncall schedule
github-actions[bot]
committed
fd443f2
remove legacy GPT code (#4322)
dimapihtar
authored
39ec5eb
ci: introduce L-tier scope vocabulary via parser (#4625)
balasaajay
authored
008c3e9

Commits on May 7, 2026

Commits on May 8, 2026

fix: post-CI corrections (round 2)
svcnvidia-nemo-ci
committed
e7c249b