Run matched training and compare the loss curves.
What it does
Guides training runs and loss comparisons between Megatron-LM and Megatron Bridge using matched configurations.
When to use it
Use when running training, comparing MLM and Bridge losses, translating arguments, or investigating diverging loss curves.
How to use it
Give it a training or correlation question for the project; it returns launch commands and configuration instructions.
What you provide
- An existing project
Good to know
- cannot be undone: nemo_experiments
Latest review
What you need · 4
Training launches must use uv run rather than bare torchrun or python.
The training launch commands require a Python runtime; no version is specified.
The launch wrapper uses the torch.distributed.run module.
Access · 1
This skill
nemo_experiments
Write
Deletes previous training checkpoints (irreversible)