NVIDIA-NeMo / NVIDIA-NeMo/RL

[mcore] Dynamic CP group for training with sequence packing

Open
#1,248 0 comments 0 reactions 1 assignee Claimed by @parthmannan View on GitHub
Performance
Dominant language
Python
Stars
2k
Forks
561
Avg merge
4d 5h
Merged PRs (30d)
145

Description

**Is your feature request related to a problem? Please describe.**
In long-context GRPO, the sequence length distribution in a batch can be very skewed, and always using a large CP to tolerate the longest sequence is inefficient for most short sequences.

This issue requests for dynamic CP group feature that can dynamically arrange DP and CP group size per batch to reduce exposed CP communication.

**Describe the solution you'd like**
A clear and concise description of what you want to happen.

**Describe alternatives you've considered**
A clear and concise description of any alternative solutions or features you've considered.

**Additional context**
Add any other context or screenshots about the feature request here.

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.