Skip to content

Pull requests: NVIDIA/Megatron-LM

New pull request New
850 Open 4,253 Closed
850 Open 4,253 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

fix(datasets): correct S3 object existence checks community-request
#6990 opened Aug 31, 2026 by TechCloud-x • Draft
1
ci: AUT-2153 sync cerno pin through 5d3107a9 Run tests
#6987 opened Aug 31, 2026 by svcnemo-autobot Collaborator Loading…
2
(feat) gtp+cpu-init support complexity: low nemotron
#6986 opened Aug 31, 2026 by fanshiqing Member Loading…
6 tasks done
1
@fanshiqing
5
[distrib_optimizer]: skip fp32 grad copy in order to save gpu memory on torch>=2.10.0
#6984 opened Aug 31, 2026 by shifangx Contributor Loading…
6 tasks
1
feat(fsdp): validate data-parallel sharding strategies complexity: low Run functional tests
#6983 opened Aug 30, 2026 by wujingyue Contributor Loading…
1
fix(optimizer): detach FP32 model params before creating shards in DistributedOptimizer community-request waiting-on-customer Waiting on the original author to respond
#6982 opened Aug 30, 2026 by Ultron09 Loading…
4 of 6 tasks
1
@wujingyue
6
fix: cache a copy of the GPT dataset loss mask complexity: low Run tests
#6981 opened Aug 30, 2026 by ZhiyuLi-Nvidia Contributor Loading…
4 of 6 tasks
1
[Dev] Fix DSA indexer metric accounting under Dynamic CP and CUDA Graph capture
#6980 opened Aug 30, 2026 by buptzyb Contributor • Draft
3
docs: update MFSDP public shard placement docs-only documentation only (docs or docstrings)
#6979 opened Aug 30, 2026 by wujingyue Contributor Loading…
1
Add block-atomic DBuffer placement
#6978 opened Aug 30, 2026 by wujingyue Contributor • Draft
1
Move raw multimodal Tensors to CUDA before inference request checkpoint recomp. complexity: low Final Review PR is in the "final review" stage
#6976 opened Aug 29, 2026 by cspades Member Loading…
1 of 6 tasks
@cspades
2
fix(checkpointing): load_biencoder_checkpoint passes release twice to get_checkpoint_name community-request
#6975 opened Aug 29, 2026 by Anai-Guo Loading…
@wujingyue
7
MFSDP v2: overlap the DP-outer gradient reduction on its own stream community-request
#6973 opened Aug 29, 2026 by Achyuthan-S Contributor • Draft
10
security(claude-review): AUT-2169 isolate analysis from publication
#6972 opened Aug 29, 2026 by svcnemo-autobot Collaborator • Draft
2
fix: propagate persistent async checkpoint worker failures community-request
#6971 opened Aug 29, 2026 by zupengwang Loading…
3 of 6 tasks
1
@wujingyue
11
ci: Add selective unit testing support using pytest-impacted
#6969 opened Aug 29, 2026 by balasaajay Contributor • Draft
6 tasks
2
Don't assert an address across DBuffer storage release and reallocate complexity: low
#6968 opened Aug 29, 2026 by wujingyue Contributor Loading…
1
fix(determinism): Separate the MoE aux-loss fusion flag so --deterministic-mode guards only the op that breaks it complexity: low Run CICD
#6967 opened Aug 29, 2026 by ZhiyuLi-Nvidia Contributor Loading…
4 of 6 tasks
2
fix(determinism): Pin the Triton autotune cache to one shared directory under --deterministic-mode complexity: low Run CICD
#6966 opened Aug 29, 2026 by ZhiyuLi-Nvidia Contributor Loading…
5 of 6 tasks
2
fix: Do not pass is_first_microbatch to TE when it is not maintained complexity: low
#6965 opened Aug 29, 2026 by ZhiyuLi-Nvidia Contributor Loading…
4 of 6 tasks
1
Guard FlashAttention 4 import by installed distribution community-request Final Review PR is in the "final review" stage waiting-on-customer Waiting on the original author to respond
#6964 opened Aug 29, 2026 by hershg Loading…
@wujingyue
12
fix(moe): preserve fused expert weight storage through DDP complexity: medium
#6963 opened Aug 29, 2026 by anlthms Contributor Loading…
1 task done
2
fix(fsdp): reuse persistent model-weight storage complexity: medium Run functional tests
#6962 opened Aug 28, 2026 by wujingyue Contributor Loading…
12
Allow MLA ('+') pre-layers in HybridStack overlap scheduling
#6960 opened Aug 28, 2026 by Connor-XY Contributor • Draft
1
Add shortcut-moe
#6959 opened Aug 28, 2026 by jiemingz Contributor • Draft
6 tasks
2
ProTip! Adding no:label will show everything without a label.