Skip to content

Pull requests: NVIDIA/Megatron-LM

New pull request New
331 Open 2,030 Closed
331 Open 2,030 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

Add common pile scripts complexity: high
#3902 opened Mar 17, 2026 by Phlip79 Loading…
5 tasks
Core 0.16
8
CI: add parallel GB200 integration test track Run functional tests
#3901 opened Mar 17, 2026 by ko3n1g • Draft
4 tasks
Core 0.16
3
Sbak/ckpt migrate
#3899 opened Mar 17, 2026 by dimapihtar • Draft
5 tasks
1
[main][fix] Add TE's attribute to check if offloading tensor
#3896 opened Mar 17, 2026 by lhb8125 • Draft
5 tasks
Core 0.16
3
Split merge-main-into-dev: docs and CI changes
#3895 opened Mar 16, 2026 by ilml • Draft
1 of 2 tasks
1
Split merge-main-into-dev: test updates
#3894 opened Mar 16, 2026 by ilml • Draft
1 of 2 tasks
1
Split merge-main-into-dev: essential runtime changes
#3893 opened Mar 16, 2026 by ilml • Draft
1 of 2 tasks
1
Support GEMM + Swiglu fused MLP mirror-to-main
#3890 opened Mar 16, 2026 by ksivaman Loading…
5 tasks
Core 0.16
15
Tokenizers simplification Run functional tests Run MBridge tests Attach this for testing this PR against MBridge main Run tests
#3889 opened Mar 16, 2026 by asolergi-nv • Draft
5 tasks done
Core 0.16
5
Enhance rotary positional embedding version checks
#3887 opened Mar 16, 2026 by huvunvidia Loading…
5 tasks
Core 0.16
3
Rename RL timers to be consistent complexity: medium
#3878 opened Mar 15, 2026 by tdene Loading…
5 tasks
Core 0.16
6
Add RL token throughput and packing metrics complexity: low
#3877 opened Mar 15, 2026 by tdene Loading…
5 tasks
Core 0.16
17
Parity with VLLM over the reasoning field Approved All necessary approvals have been made complexity: low
#3873 opened Mar 15, 2026 by tdene Loading…
5 tasks
Core 0.16
5
feat: CUDA graph support for packed sequence (variable-length) training
#3869 opened Mar 14, 2026 by seonjinn • Draft
5 tasks
2
fix cg acess issue by using dict instead of list to iteratively acces…
#3867 opened Mar 13, 2026 by ilml Loading…
5 tasks
Core 0.16
7
Patch EOD out of inference results Approved All necessary approvals have been made complexity: low Expert Review [deprecated] Apply this label to indicate that your PR is ready for expert review.
#3866 opened Mar 13, 2026 by tdene Loading…
5 tasks
Core 0.16
7
Merge main into dev complexity: high
#3865 opened Mar 13, 2026 by ilml Loading…
5 tasks
2
Allow untokenized messages when using the prevent tokenization mode. complexity: low Final Review PR is in the "final review" stage
#3864 opened Mar 13, 2026 by ArEsKay3 Loading…
5 tasks
3
Pass gracefully if token_id not found in message Approved All necessary approvals have been made
#3862 opened Mar 13, 2026 by i-riyad Loading…
5 tasks
6
Remove packed_attention_mask unused parameter complexity: low Final Review PR is in the "final review" stage
#3859 opened Mar 13, 2026 by tdene Loading…
5 tasks
Core 0.16
2
MFU tracking for inference complexity: medium
#3856 opened Mar 13, 2026 by tdene Loading…
5 tasks
Core 0.16
7
Separate save_checkpoint into per-type execution paths community-request needs-follow-up Issue needs follow-up
#3852 opened Mar 13, 2026 by Anmol202005 Loading…
5 tasks done
2
fix: Handle quantized CUDA tensors in async checkpoint writer complexity: low Final Review PR is in the "final review" stage
#3845 opened Mar 12, 2026 by sbak5 Loading…
5 tasks
Core 0.16
8
Add activation logging and tokens per expert logging
#3842 opened Mar 12, 2026 by Mellonta • Draft
5 tasks
1
[dev] fix no_shard training convergency and add unittest for no_shard
#3835 opened Mar 12, 2026 by wplf Loading…
1
ProTip! Adding no:label will show everything without a label.