ngmi / NVIDIA / Megatron-LM

NVIDIA/Megatron-LM

Python 15.3k stars

Ongoing research training transformer models at scale

✓ Synced 1h ago Share on X →
README badge: [![ngmi](https://ngmi.review/badge/NVIDIA/Megatron-LM.svg)](https://ngmi.review/repo/NVIDIA/Megatron-LM)
1.0k Merged PRs
9 days Avg Merge Time
0m Fastest PR
1 year Slowest PR
#849 Global Speed Rank

PR Size Analysis

Lines changed (additions + deletions) vs review outcomes. Re-sync to populate data for older PRs.

PRs by size
Avg review time (hrs)
Clean approval rate (%)

Top Reviewers

#1 @tdene
@tdene
255 reviews ✓ 49 approved
#2
@ko3n1g
230 reviews ✓ 193 approved ↺ 15 blocked
#3
@kvareddy
170 reviews ✓ 154 approved
#4
@jaredcasper
160 reviews ✓ 128 approved
#5 @chtruong814
@chtruong814
137 reviews ✓ 129 approved
#6 @ericharper
@ericharper
136 reviews ✓ 126 approved
#7 @yanring
@yanring
117 reviews ✓ 95 approved
#8
@deepakn94
115 reviews ✓ 66 approved
#9 @yaoyu-33
@yaoyu-33
112 reviews ✓ 79 approved
#10 @shanmugamr1992
@shanmugamr1992
107 reviews ✓ 91 approved

Recent Merged PRs

# Title Author Time Reviews Blocks
#3537 fix: skip non-tensor optimizer state entries in distrib_optimizer sav… @ahmadki 16.2h 8 ✓
#3527 Inference: Create finer grained cuda-graphs with better coverage of smaller batch sizes @sidsingh-nvidia 2 days 4 ✓
#3476 Remove redundant CUDA calls in the LLaVA dataloader @duncanriach 5 days 4 ✓
#3408 remove deprecated params from model parallel config @dimapihtar 10 days 3 ✓
#3411 remove deprecated mamba params @dimapihtar 10 days 4 ✓
#3543 [dev] `cp: Cherrypick CI changes` @ko3n1g 50m 1 ✓
#3412 remove deprecated async_grad_allreduce param @dimapihtar 10 days 3 ✓
#3413 remove deprecated get_te_version @dimapihtar 10 days 3 ✓
#3407 remove deprecated SampleListWebdataset @dimapihtar 10 days 2 ✓
#3536 Do not Slack notify for draft PRs @Phlip79 3.7h 1 ✓
#3533 chore(beep boop 🤖): Bump `uv.lock` (core_r0.16.0) (2026-02-23) @svcnvidia-nemo-ci 8.6h 1 ✓
#3526 do not add EoD @arendu 20.1h 2 ✓
#3250 Renable full_iteration cuda graphs for inference. Add them for the mamba block. @sidsingh-nvidia 16 days 7 ✓
#3510 Fix Megatron-FSDP fully_shard() optimizer state DCP checkpointing, and fix DTensor deepcopy bug from PyTorch 26.01. @cspades 17.2h 10 ✓
#3505 docs: Update docs for 0.16.0 @chtruong814 19.6h 2 ✓
#3449 Multimodal: fix argument checking @faradawn 3 days 1 ✓
#3484 ci: Also sync direct teams @ko3n1g 1 day 3 ✓
#3487 ci: Enable Dependabot Automerge @ko3n1g 1 day 4 1×
#3513 ci: MBridge testing branch name during merge-queues @ko3n1g 4.0h 1 ✓
#3030 Track off-policyness across RL steps @tdene 29 days 4 ✓