From NVIDIA Megatron-LM for visibility - #18

Open

RaymondLi0 wants to merge 7826 commits into

bigcode-project:multi-query-attentionfrom

This pull request is big! We're only showing the most recent 250 commits

Commits on Aug 12, 2026

Eliminate the ShardedObject all_gather_object on the FullyParallel load path (#5551)
asolergi-nv
authored
14346b6
chore: rotate oncall schedule
github-actions[bot]
committed
15c83d2
GTP + PartialCG: Classify latent projections for partial CUDA graphs (#6446)
JF-D
authored
81fe7c7
Balance LayerWise optimizer shards by Newton-Schulz cost, not parameter size (#6379)

deepakn94
and
claude
authored
8534490
Suppress noisy external-library log lines on non-rank-0 processes (#5590)
deepakn94
authored
fad6111
feat(ci): enhance triage script to assign on-call assignee dynamically (#6469)
balasaajay
authored
b2f2888

Commits on Aug 13, 2026

Skip attention mask materialization in RL training (#5885)
tdene
authored
5436289
Avoid unnecessary MoE router host synchronization (#6432)
JF-D
authored
a34880f
Refactor GDP dynamic inference onto the shared SSM interface (#6443)

authored
3cd8cf3
test(fsdp): mark default overlap check flaky (#6493)
wujingyue
authored
3ce4b4a
Paged stashing no longer assumes single layer config (#6419)
Phlip79
authored
301e0d6
Respect `profile_ranks` on log interval memory snapshots (#6354)
maanug-nv
authored
5714a20
Add unit tests for GDP inference (#6505)

authored
fecb29f
[fix] fix GTP+DCP ckpt saving/loading for GDP module (#6503)
fanshiqing
authored
8ba5c60
feat(fsdp): support expert parallelism in MFSDP v2 adapter (#6450)

wujingyue
and
svcnvidia-nemo-ci
authored
281075d
Complete clamped SwiGLU across expert paths (#5940)
FDecaYed
authored
2a75ac1
RL: unify rollout generation into one pipeline (#6495)
tdene
authored
87bf520
Avoid passing -1 sentinels into inference kernels (#6481)

tdene
and
jalbericiola
authored
487bb39
RL: bound per-env wandb keys to the char limit (#6477)

tdene
and
jalbericiola
authored
09921c0
Remove torch cuda empty cache (#5678)

shurkat-nvidia
and
claude
authored
2119719
[fix] GTP+DDP potential data-race: publish DDP params before GTP prefetch reads them (#6388)

fanshiqing
and
JF-D
authored
87cc590
Cover symmetric memory in CUDA graph test (#5987)

wujingyue
and
claude
authored
d7430d5
Fix tool call reasoning boundary (#6411)
nvcsathe
authored
f7b2209
Update Nemotron 3.5 Lightning golden values (#6445)
Phlip79
authored
0ee415b
Handle hidden chunked prefill during async overlap (#6442)
lmcafee-nvidia
authored
94939b8
Thread socket argument through MInf API (#6513)
tdene
authored
f9cc544
Update release notes page (#6520)
Phlip79
authored
69e282e
test(fsdp): verify optimizer CUDA graphs in v2 (#6498)
wujingyue
authored
3d0956f
Clean up SSM ops directory structure (#6444)

santhnm2
and
shanmugamr1992
authored
de4a82e
Add read/write ledger for bulky inference metadata (#6125)
tdene
authored
d918967

Commits on Aug 14, 2026

Commits on Aug 15, 2026

ci: AUT-1408 replace cicd PAT with GitHub App tokens (#6360)

svcnemo-autobot
and
ko3n1g
authored
669d6ed
fix: AUT-1501 abort when SSO list download fails (#6516)
svcnemo-autobot
authored
51f6a39
Fix missing golden values updates (#6528)
janEbert
authored
e79110c
Support HFSDP deferred DP-outer gradient reduction (#5972)

authored
e60e42f

Commits on Aug 17, 2026

Commits on Aug 18, 2026

Commits on Aug 19, 2026

Share prefill in sequence packing (#5887)

authored
3ad8c2a
Update copy-pr-bot.yaml [skip ci]
github-actions[bot]
committed
5f9b2bb
Add opt-in MFSDP unified communication stream (#6563)
wujingyue
authored
540f420
tests: allow trusted TE pickle extra state in affected tests (#6602)
maanug-nv
authored
1d69364
Reflect #4440 onto GPT models, not just hybrid (#6476)
tdene
authored
afa707e
ci: Remove scheduled CI workflow (#6640)
chtruong814
authored
0961f12
fix(ci): request pull request permission for queue comments (#6641)
ko3n1g
authored
e7d58db
Remove core teams from FSDP owners (#6623)
Phlip79
authored
29361d6
Add THD-format per-rank batch fetch with padding mask for sequence packing (#6625)

ilml
and
claude
authored
c452d5a
Optimize MIMO pipeline communication (#6606)
yashaswikarnati
authored
f481e63
chore: rotate oncall schedule
github-actions[bot]
committed
cab7116
Support no-shard, ZeRO-1, and ZeRO-2 in MFSDP v2 (#6137)

Autumn1998
and
wujingyue
authored
01dc20b
Add hybrid SSM state handoff (#6584)
wdykas
authored
731b791
Checkpoint shared storage when inputs are returned (#4497)
CarlosGomes98
authored
19deef6
Add script-transfer procedure to GPT-to-Hybrid migration skill (#6251)
guihong-nv
authored
7e3780f
Remove leading NaN iteration times from golden values (#6539)
balasaajay
authored
93ebf8d
Add varlen dataset schema-detection and message-converter helpers (#5908)

ilml
and
claude
authored
9691f47
MXfp8 Inference Fixes and clean up (#6588)
wdykas
authored
a1912b3
ci: centralize container builds and add cache population workflow (#6646)
balasaajay
authored
e7b92ef
tokenizers simplification (#6622)
dimapihtar
authored
e433171
ci: Change copy-pr-bot update schedule to 9 a.m. UTC (#6669)
chtruong814
authored
82894c4
Support mixed-precision optimizers in MFSDP v2 (#6506)
wujingyue
authored
fe1e8f8

Commits on Aug 20, 2026

ci: AUT-1606 bump release library to v1.8.10 (#6678)
svcnemo-autobot
authored
4e1df1b
Add CUDA graph support for GDP dynamic inference (#6595)
santhnm2
authored
3a56046
Pass process groups to GDP projections (#6687)
yashaswikarnati
authored
a1580eb
Fix fused MLA and MTP checkpointing for Megatron-FSDP. (#6271)

rapatel
and
gautham-kollu
authored
93924d5
fix: prevent local CUDA graph stream aliasing (#6643)
JF-D
authored
54b5e98
Share MFSDP gradient completion hook (#6695)
wujingyue
authored
cf3b8ee
Fix LayerWise buffer routing to respect use_muon = False (#6642)
kwangjunahn
authored
1d82259
remove unused dist_checkpointing code (#6663)
dimapihtar
authored
893eb05
Add HybridModel MTP loss control (#6119)
Phlip79
authored
db018b3
Add VLM support to the dynamic-batching inference server (#6260)

RPrenger
and
cspades
authored
881f489
ci: AUT-1616 assign root build files to CI owners (#6710)
svcnemo-autobot
authored
f300051
Scope activation offload summaries to model ranks (#6652)
yashaswikarnati
authored
28797d0
Add DP-balanced sequence-packing scheduler and wrap_data_iterator entry point (#6684)

ilml
and
claude
authored
a9381be
Fix Refit Cache for mamba (#6672)
wdykas
authored
fbbc142
Add GDN dynamic inference (#6499)

anlthms
and
codex
authored
606b9f4
[NCCL EP] Add MXFP8 dispatch support (#6524)
YangFei1990
authored
7afe031
rl: add durable rollout bank for completed groups (#6352)

lauradang
and
claude
authored
787649a
fix: ensure GTP params are ready before CUDA graph replay (#6525)
JF-D
authored
b3c9100
fix(moe): broadcast padding mask for expert bias counts (#6114)

cuichenx
and
jaredcasper
authored
723db5a

Commits on Aug 21, 2026

Reduce eval loss over the GTP-remat-inclusive data-parallel group (#6717)
deepakn94
authored
df197a8
GTP+Muon: support 'blockwise' and 'distributed' tp_mode (#6380)

deepakn94
and
fanshiqing
authored
bb47df5
ci: isolate tokenizer metadata writes (#6718)
chtruong814
authored
57f56d4
ci: prioritize baseline cache for merge groups (#6721)
chtruong814
authored
41943e8
Update copy-pr-bot.yaml [skip ci]
github-actions[bot]
committed
5b2a621
build: AUT-1615 add torch-memory-saver dev dependency (#6701)
svcnemo-autobot
authored
ab5e2c9
Apply precision configs to MIMO modalities (#6745)
yashaswikarnati
authored
b4a4c0c
[FSDP] Reject post-wrap load_state_dict(assign=True) in MFSDP V2 (#6203)

chengcuiping
and
svcnvidia-nemo-ci
authored
6a7bc99
Preserve pinned host memory for CPU-offloaded FP32 params (#4547)
Orange-Lv
authored
faecf6b
Forward rotary_base and rope_scaling_factor to LLaVAModel in the VLM builders (#6586)

adityaghai07
and
claude
authored
60e0396
Remove MFSDP DTensor availability shim (#6752)
wujingyue
authored
101966c
Use DTensor placements for DBuffer (#6744)
wujingyue
authored
0d970bf
Count MoE aux-loss contributors explicitly in hybrid models and repeated-MTP layers (#6748)

deepakn94
and
rogerwaleffe
authored
2e12e2c
Fix bugs in re-sharding `DistributedOptimizer` state in `torch_dist` checkpoint (#5077)

deepakn94
and
claude
authored
32a03b7
RL: refill all-placeholder groups (#6480)

tdene
and
jalbericiola
authored
24f9a65
feat(determinism): Make the channel-last conv1d deterministic (#6655)

ZhiyuLi-Nvidia
and
claude
authored
f101c47
Move stacked PR merge checks to AGENTS.md (#6758)
wujingyue
authored
3b0995d
add gigatoken support for hf tokenizers (#6514)

authored
889c909
Move MFSDP placements to the API boundary (#5971)
wujingyue
authored
b7d4760
fix: avoid requiring GTP for CUDA graph stream warning (#6765)
JF-D
authored
7423569

Commits on Aug 22, 2026

Commits on Aug 24, 2026

Commits on Aug 25, 2026

CP layout integration (#6776)

Mellonta
and
yuzhongw-nvidia
authored
aeaf7c6
fix: make local CUDA graph outputs pipeline-safe (#6168)
nvegesna-netizen
authored
377ad24
Update Q3 Roadmap Links (#6830)
Phlip79
authored
7020985
[Inference] Fix deadlock when initializing ZMQ communication between inference engine ranks. (#6779)
cspades
authored
19abd1f
Add test for layer-wise optimizer bucket params on single DP rank (#5257)

Rohithmatham12
and
svcnvidia-nemo-ci
authored
ff8c0f8
Batch invariant nvls (#6806)

authored
cc99d8e
M2n refit backend (#6712)
wdykas
authored
dc931ad
Remove Modelopt loading fn (#5405)
AAnoosheh
authored
a97b0e2
ci: Set provenance: false during docker build (#6849)
chtruong814
authored
4ac689b
ci: bump Cerno commit to f0cf5b7 (#6817)

thomasdhc
and
codex
authored
5d3107a

Commits on Aug 26, 2026

Commits on Aug 27, 2026

Commits on Aug 28, 2026

Commits on Aug 29, 2026

Commits on Aug 30, 2026

Support destination CP in non-colocated bridge communicator (#5720)
yashaswikarnati
authored
f2f0f7b
GDP CP Support (#6815)
Mellonta
authored
d7e06cf

Commits on Aug 31, 2026

Commits on Sep 1, 2026