Skip to content

Pull requests: NVIDIA/Megatron-LM

New pull request New
799 Open 4,114 Closed
799 Open 4,114 Closed
Author
Filter by author
Loading
Label
Filter by label
Loading
Use alt + click/return to exclude labels
or ⇧ + click/return for logical OR
Projects
Filter by project
Loading
Milestones
Filter by milestone
Loading
Reviews
Filter by reviews
No reviews Review required Approved review Changes requested
Assignee
Filter by who’s assigned
Assigned to nobody Loading
Sort
Sort by
Newest Oldest Most commented Least commented Recently updated Least recently updated Best match
Most reactions
👍 👎 😄 🎉 😕 ❤️ 🚀 👀

Pull requests list

Route custom process groups through GDP and latent MoE checkpoints complexity: low nemotron Run functional tests
#6781 opened Aug 23, 2026 by yashaswikarnati Contributor Loading…
3 of 6 tasks
4
Add prefix caching support for GDP dynamic inference
#6780 opened Aug 22, 2026 by santhnm2 Contributor • Draft
6 tasks
2
[Inference] Fix deadlock when initializing ZMQ communication between inference engine ranks. complexity: low
#6779 opened Aug 22, 2026 by cspades Member Loading…
1 of 6 tasks
@cspades
2
fix(num_zeros logging with gtp padding): exclude GTP alignment padding from num_zeros accounting complexity: low nemotron
#6778 opened Aug 22, 2026 by fanshiqing Member • Draft
5 of 6 tasks
@fanshiqing
3
style: apply Google guide to MFSDP v2
#6777 opened Aug 22, 2026 by wujingyue Contributor • Draft
2
CP layout integration complexity: high nemotron
#6776 opened Aug 22, 2026 by Mellonta Contributor Loading…
1
Megatron-FSDP: Hybrid FSDP over replicated weights, single-rank reduction and bucket padding fixes complexity: high
#6773 opened Aug 22, 2026 by rapatel Contributor Loading…
1 of 6 tasks
3
Add PR-only selective unit testing
#6772 opened Aug 21, 2026 by Phlip79 Member • Draft
1
Mask MIMO MTP prediction paths that cross modality tokens complexity: medium nemotron Run functional tests
#6770 opened Aug 21, 2026 by kevjshih Contributor Loading…
4 of 6 tasks
@sancha @kevjshih @yashaswikarnati
7
RL: add per-step rollout bank metrics complexity: low
#6769 opened Aug 21, 2026 by lauradang Contributor Loading…
1
rl: persist incomplete rollout groups and replay them
#6768 opened Aug 21, 2026 by lauradang Contributor • Draft
3 of 6 tasks
1
Fix premature termination of persistent ckpt worker Approved All necessary approvals have been made nemotron
#6767 opened Aug 21, 2026 by shurkat-nvidia Contributor Loading…
1 task done
11
Select kernel backends through BackendSpecProvider, not HAVE_* flags
#6766 opened Aug 21, 2026 by guihong-nv Contributor • Draft
2
Support CP layout conversion in hybrid models complexity: high Final Review PR is in the "final review" stage nemotron
#6764 opened Aug 21, 2026 by Mellonta Contributor Loading…
15
Layer sharded Muon scope change to allow hybrid TP mode on expert and dense matrices
#6761 opened Aug 21, 2026 by wanyingw • Draft
6 tasks
1
Thread explicit process groups through the MTP path
#6760 opened Aug 21, 2026 by liding-nv Contributor • Draft
2
Document GitHub PR CI reruns
#6759 opened Aug 21, 2026 by wujingyue Contributor • Draft
1
chore: AUT-1623 nightly sync main into dev (2026-08-21) Run functional tests Run MBridge tests Attach this for testing this PR against MBridge main Run NeMoRL tests
#6756 opened Aug 21, 2026 by svcnemo-autobot Collaborator Loading…
3
[Main] Fix Numerical issue for GTP + Muon + MXFP8 param gather + use layerwise optimizer layout complexity: low nemotron
#6754 opened Aug 21, 2026 by zhongbozhu Contributor Loading…
6 tasks
6
[Dev] fix: account for DSA top-k sparsity and indexer cost in FLOPs complexity: low Run tests
#6753 opened Aug 21, 2026 by niyunsheng Member Loading…
4
Defer initial embedding synchronization during VPP model construction community-request
#6750 opened Aug 21, 2026 by SonglinLife Loading…
1
4
Fix MTP CP-boundary roll to communicate on the context-parallel process group complexity: low
#6746 opened Aug 21, 2026 by liding-nv Contributor Loading…
2
Support MIMO GTP checkpoint resume complexity: low nemotron
#6743 opened Aug 21, 2026 by yashaswikarnati Contributor Loading…
5
Integrate sequence-packing scheduler into training loop and wire varlen dataset into GPT pretraining complexity: high
#6742 opened Aug 21, 2026 by ilml Contributor Loading…
2 tasks done
5
[Dev] Add absolute row alignment for MTP sequence rolls
#6741 opened Aug 21, 2026 by buptzyb Contributor • Draft
2 of 4 tasks
1
ProTip! Exclude everything labeled bug with -label:bug.