Skip to content

Pull requests: NVIDIA/Megatron-LM

New pull request New
399 Open 2,609 Closed
399 Open 2,609 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

Support GEMM + SwiGLU fused MLP (rebased from #3971)
#4636 opened May 5, 2026 by Connor-XY • Draft
5 tasks
1
chore: nightly sync main into dev (05_05_2026) Run functional tests Run MBridge tests Attach this for testing this PR against MBridge main
#4635 opened May 5, 2026 by svcnvidia-nemo-ci • Draft
3
Inference: Cache input + position ID views complexity: low
#4634 opened May 5, 2026 by mathemakitten Contributor Loading…
5 tasks
2
Minimal repro for Mamba's resharding
#4633 opened May 5, 2026 by wujingyue Contributor • Draft
1
moe: add DeepEP V2 ElasticBuffer support to MoE flex dispatcher community-request
#4632 opened May 5, 2026 by dmvevents • Draft
1
1
Refactor TE fused ops integration into mixin
#4630 opened May 5, 2026 by CarlosGomes98 Contributor • Draft
5 tasks
4
Disable MSC by default; opt in via --enable-msc complexity: low
#4629 opened May 5, 2026 by asolergi-nv Contributor Loading…
5 tasks done
5
dist-ckpt: add local-replica mode to eliminate cross-rank reads of replicated params on torch_dist load complexity: medium
#4628 opened May 5, 2026 by asolergi-nv Contributor Loading…
5 tasks done
4
Overlap logprob calculation by putting it on separate stream
#4627 opened May 5, 2026 by tdene Contributor • Draft
5 tasks
1
docs: add bump-dependency skill
#4626 opened May 5, 2026 by ko3n1g Contributor • Draft
5
ci: introduce L-tier scope vocabulary via parser
#4625 opened May 5, 2026 by balasaajay Contributor • Draft
5 tasks
4
[dev]: faster implementation of mHC fused kernels
#4624 opened May 5, 2026 by jingqiny-99 • Draft
5 tasks
8
Fix optimizer CPU offload for megatron-fsdp dtensor param complexity: low module: megatron-fsdp
#4623 opened May 5, 2026 by wplf Member Loading…
@wplf
4
feat(attention): Add rotary_base_per_layer for Step-3.5-Flash
#4622 opened May 5, 2026 by shifangx Contributor • Draft
5 tasks
2
[Dev] Fix single grouped weight when enabling MXFP8 primary weight
#4621 opened May 5, 2026 by zhongbozhu Contributor • Draft
5 tasks
2
[training migration] Migrate config validation
#4618 opened May 5, 2026 by maanug-nv Contributor • Draft
5 tasks
1
One single flag that determines if we are in inference
#4617 opened May 4, 2026 by tdene Contributor • Draft
5 tasks
1
chore: Update Docker image version to 26.04-py3 Run functional tests
#4611 opened May 4, 2026 by balasaajay Contributor • Draft
5 tasks
10
Fix gradient corruption with layerwise param all-gather overlap Approved All necessary approvals have been made complexity: low
#4609 opened May 4, 2026 by deepakn94 Contributor • Queued
1 task done
16
fix tokenizers in respect to newer transformers complexity: low Expert Review [deprecated] Apply this label to indicate that your PR is ready for expert review. Run functional tests
#4608 opened May 4, 2026 by dimapihtar Contributor Loading…
5 tasks
5
ci: Bump GHA versions Approved All necessary approvals have been made complexity: low
#4606 opened May 4, 2026 by chtruong814 Contributor Loading…
5 tasks
2
Enable async scheduling for decode-only inference Run tests
#4604 opened May 4, 2026 by lmcafee-nvidia Contributor • Draft
1
Inference: Speed up the moe_sum_kernel by capping number of blocks complexity: low
#4603 opened May 4, 2026 by sidsingh-nvidia Contributor Loading…
5 tasks
2
ci: replace local _release_library.yml with FW-CI-templates composables
#4602 opened May 4, 2026 by ko3n1g Contributor • Draft
1 of 4 tasks
25
add CLIP pretraining community-request
#4600 opened May 4, 2026 by andrea-pilzer • Draft
2
ProTip! Follow long discussions with comments:>50.