NVIDIA/Megatron-LM

Ongoing research training transformer models at scale

76
/ 100
Verified

Provides composable GPU-optimized building blocks for transformer training, including advanced parallelism strategies (tensor, pipeline, expert, and context parallelism), mixed precision support (FP16, BF16, FP8, FP4), and custom training pipeline construction. Achieves up to 47% Model FLOP Utilization on H100 clusters while scaling from 2B to 462B parameter models. Integrates with Hugging Face via Megatron Bridge for checkpoint conversion and works with NVIDIA NeMo framework for production-ready recipes.

15,633 stars. Actively maintained with 226 commits in the last 30 days.

No Package No Dependents
Maintenance 25 / 25
Adoption 10 / 25
Maturity 16 / 25
Community 25 / 25

How are scores calculated?

Stars

15,633

Forks

3,689

Language

Python

License

Last pushed

Mar 13, 2026

Commits (30d)

226

Get this data via API

curl "https://pt-edge.onrender.com/api/v1/quality/transformers/NVIDIA/Megatron-LM"

Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.