Skip to content

Pull requests: NVIDIA/Megatron-LM

New pull request New
963 Open 4,644 Closed
963 Open 4,644 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 first-class HybridModel config lists in Core
#7555 opened Sep 21, 2026 by Phlip79 Member • Draft
1 task
1
Qwen3.8-Flash-Next: gated-residual mHC, Qwen Sparse Attention, hashed n-gram memory, and mHC pipeline support
#7553 opened Sep 21, 2026 by BestJuly Contributor • Draft
1
fix: preserve FIM split tokens before empty fragments community-request
#7552 opened Sep 21, 2026 by Sunt-ing Contributor Loading…
4 of 7 tasks
2
Fix stale gradients in hybrid optimizer community-request
#7551 opened Sep 21, 2026 by adenzhou1350 • Draft
2 of 7 tasks
1
chore(beep boop 🤖): Bump uv.lock (main) (2026-09-21) Run functional tests
#7550 opened Sep 21, 2026 by svcnvidia-nemo-ci Contributor • Draft
1 task
3
chore(beep boop 🤖): Bump uv.lock (core_r0.19.0) (2026-09-21) Run functional tests
#7549 opened Sep 21, 2026 by svcnvidia-nemo-ci Contributor Loading…
1 task
2
Fix nonowner cleanup of shared checkpoint test directories
#7548 opened Sep 21, 2026 by Connor-XY Contributor • Draft
1
SHARP barrier uses the replicate DP-CP group
#7547 opened Sep 21, 2026 by michal2409 Contributor • Draft
3 of 6 tasks
1
Fix custom Hybrid test process-group creation order
#7546 opened Sep 21, 2026 by Connor-XY Contributor • Draft
1
[Dev] Perf: Improvements of Kimi-K3-style hybrid model context parallel
#7545 opened Sep 21, 2026 by yuzhongw-nvidia Contributor • Draft
7 tasks
1
Remove vestigial Megatron-FSDP optional-import guards
#7544 opened Sep 21, 2026 by shjwudp Contributor • Draft
1
Fix duplicate offsets in TikToken conversion and inverse vocabulary community-request
#7543 opened Sep 21, 2026 by Sunt-ing Contributor Loading…
4 of 7 tasks
2
feat(mlite): implement pipelined H2D/D2H overlapped fused CPU AdamW community-request
#7542 opened Sep 21, 2026 by shyoshyo Loading…
1 of 7 tasks
1
fix(mimo): omit encoder bridges from LLM-only provider Run tests
#7541 opened Sep 21, 2026 by yashaswikarnati Contributor • Draft
5 of 7 tasks
1
1
Rename experimental FSDP Flat placement to RowAtomic complexity: low
#7539 opened Sep 21, 2026 by wujingyue Contributor Loading…
1 task done
1
[dev] Restore the forward's FP8 autocast state in tensor_parallel.checkpoint recompute
#7538 opened Sep 21, 2026 by HaochenYuan Contributor Loading…
7 tasks
1
fix(training): pass model process groups to GPU sniff tests Run tests
#7537 opened Sep 21, 2026 by yashaswikarnati Contributor • Draft
7 tasks done
1
ci: Compare selected nightly tests on GB300 Run selective unit tests
#7536 opened Sep 21, 2026 by balasaajay Contributor • Draft
3
ci: check megatron.inference imports community-request
#7535 opened Sep 21, 2026 by priyadarshnisundararajan Loading…
1
[dev] docs: add combined 1F1B MoE A2A overlap guide docs-only documentation only (docs or docstrings)
#7534 opened Sep 21, 2026 by lhb8125 Contributor • Draft
7 tasks
1
[GTP] Symmetric Memory All-gather support in GTP
#7533 opened Sep 21, 2026 by prajwal1210 Contributor • Draft
5 of 7 tasks
@prajwal1210
3
ci: Configure Transformer Engine CUDA targets for GitHub builds Run tests
#7532 opened Sep 20, 2026 by balasaajay Contributor • Draft
1 task
2
test: centralize Megatron test-state cleanup community-request
#7531 opened Sep 20, 2026 by bitmeta69 • Draft
3 of 7 tasks
1
fix(core): mark learnable softmax_offset as tensor parallel for clip norm community-request
#7530 opened Sep 20, 2026 by dundysm Loading…
3 of 7 tasks
3
Add a PEFT pre-wrap hook to training model setup community-request
#7529 opened Sep 20, 2026 by Levius-Fubuki • Draft
1
1
ProTip! Adding no:label will show everything without a label.