NVIDIA-NeMo / NVIDIA-NeMo/RL

Investigate refit time discrepancy for Moonlight between Automodel and Megatron backends

Open
#1,743 0 comments 0 reactions 0 assignees View on GitHub
enhancement
Dominant language
Python
Stars
2k
Forks
561
Avg merge
4d 5h
Merged PRs (30d)
145

Description

Based on the following image, refit in dtensor v2 policy is significantly higher than Megatron. Investigate the difference and propose solutions to mitigate the difference.

Use grpo moonlight configs for investigation.

Image

Contributor guide

Open the contributing guide

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.