Skip to content

Pull requests: NVIDIA/Megatron-LM

New pull request New
210 Open 303 Closed
210 Open 303 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

Update pretrain_mamba.py
#1682 opened Jul 11, 2025 by vignesh1507 Loading…
[feat, moe] Add support for global aux loss
#1681 opened Jul 11, 2025 by Victarry Loading…
Issue 1672 fix: initializing the current pointed with int64 to avoid …
#1673 opened Jul 7, 2025 by sharanmayank Loading…
1
Support 1f1b a2a overlap
#1671 opened Jul 7, 2025 by lhb8125 Loading…
1
moe: remove unused variable scale_up
#1670 opened Jul 6, 2025 by WineChord Loading…
Speed up model parallel initialization
#1662 opened Jul 2, 2025 by alexqdh Loading…
Update README.md
#1660 opened Jul 2, 2025 by 21jun Loading…
Allow head_dim to override kv_channel calculation
#1655 opened Jun 27, 2025 by shuoyangd Loading…
bug fixed: wandb artifact requires the tracker file
#1654 opened Jun 27, 2025 by yezhengmao1 Loading…
Apply roll operation to position_ids in MTP
#1651 opened Jun 26, 2025 by iansheng Loading…
@iansheng @BoxiangW
fix twice allgather in moe distrib optimizer
#1645 opened Jun 23, 2025 by irobot2013-why Loading…
1
add fused_topk_softmax_without_capacity for topk router fusion
#1637 opened Jun 18, 2025 by AshOfCat Loading…
Add Anton's Megatron changes for hyena compatibility
#1636 opened Jun 18, 2025 by jwilber Loading…
Fix log-timer-to-tensorboard on logging
#1631 opened Jun 13, 2025 by wplf Loading…
Fix typos: vritual → virtual and decoeder → decoder
#1626 opened Jun 11, 2025 by EricLabile Loading…
Fix: Apply q_layernorm consistently in MLA LoRA path
#1624 opened Jun 11, 2025 by Flink-ddd Loading…
1
fix: when using moe parallel folding feature and set etp > 1 && ep == 1, the grad sync is incorrect and the loss curve is bad
#1622 opened Jun 10, 2025 by Louis-J Loading…
Set weights_only=False in optimizer
#1618 opened Jun 9, 2025 by zhic-mt Loading…
Fix mrope with context parallel
#1612 opened Jun 6, 2025 by liu-zichen Loading…
@BestJuly
use a cpu set to cache cuda tensor finished_request_ids
#1610 opened Jun 5, 2025 by ladyrick Loading…
Add DistTrain, Allow Encoder to Have Different DP Size
#1605 opened May 30, 2025 by zidanehuang001 Loading…
add node_rank argument for example scripts
#1604 opened May 30, 2025 by xylllllllll Loading…
CLIPViTModel support SP and CP
#1600 opened May 28, 2025 by Thaurun Loading…
Support Multiple Input Formats for checkpoint
#1599 opened May 28, 2025 by Thaurun Loading…
bugfix: cross_entropy inplace operations may cause backward error
#1594 opened May 24, 2025 by ChangWeiming Loading…
ProTip! Adding no:label will show everything without a label.