Skip to content

Pull requests: NVIDIA/Megatron-LM

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
Assignee
Filter by who’s assigned
Assigned to nobody Loading
Sort

Pull requests list

Make LRU prefix caching eviction policy only evict child blocks
#5822 opened Jul 15, 2026 by santhnm2 Contributor Draft
6 tasks
Fuse gated RMSNorm in GatedDeltaNet output path
#5821 opened Jul 15, 2026 by Wong4j Contributor Draft
6 tasks
Handle Full iteration CUDA graph data_iterator corruption Final Review PR is in the "final review" stage
#5819 opened Jul 15, 2026 by vasunvidia Contributor Loading…
6 tasks
Support LSQ (scale learning) in MCore
#5818 opened Jul 15, 2026 by rohansjoshi Draft
1 of 6 tasks
test: re-enable unit tests with TE release builds Approved All necessary approvals have been made complexity: low
#5817 opened Jul 14, 2026 by balasaajay Contributor Loading…
1 of 6 tasks
feat(docker): Add NCCL installation script and install NCCL 2.30.4 Approved All necessary approvals have been made complexity: low
#5815 opened Jul 14, 2026 by balasaajay Contributor Queued
1 of 6 tasks
Print important dependencies complexity: low
#5814 opened Jul 14, 2026 by gautham-kollu Contributor Loading…
6 tasks
[Example] Benchmark DCP on Nemotron Math v3 complexity: high
#5811 opened Jul 14, 2026 by ilml Contributor Loading…
Fix zero-loss indexer gradients for DSv4 CP
#5809 opened Jul 14, 2026 by kunlunl Contributor Loading…
[dev] support thd full iter graph under fixed GA
#5807 opened Jul 14, 2026 by HaochenYuan Contributor Draft
6 tasks
Gate expensive work behind log levels
#5805 opened Jul 14, 2026 by shurkat-nvidia Contributor Loading…
4 of 6 tasks
Set num_splits to 0 for FA4 inference complexity: low
#5804 opened Jul 14, 2026 by santhnm2 Contributor Loading…
1 of 6 tasks
Guard MLA YARN RoPE against cos/sin dim mismatch community-request Final Review PR is in the "final review" stage
#5799 opened Jul 14, 2026 by CodersAcademy006 Contributor Loading…
1 of 3 tasks
Inference: Extend default cuda-graph coverage to 512 tokens complexity: low
#5797 opened Jul 14, 2026 by sidsingh-nvidia Contributor Loading…
1 of 6 tasks
[main] DeepSeek-v4 full support on main
#5795 opened Jul 14, 2026 by hxbai Contributor Draft
6 tasks
Avoid FSDP unit terminology in MFSDP v2 docs-only documentation only (docs or docstrings) Final Review PR is in the "final review" stage
#5793 opened Jul 14, 2026 by wujingyue Contributor Loading…
ProTip! Updated in the last three days: updated:>2026-07-11.