NVIDIA-NeMo / NVIDIA-NeMo/RL

Track R3 RouterReplay support and validation for Megatron VPP

Open
#2,740 0 comments 0 reactions 0 assignees View on GitHub
Feature
Dominant language
Python
Stars
2k
Forks
562
Avg merge
4d 5h
Merged PRs (30d)
145

Description

## Background

R3 RouterReplay currently assigns replay tensors by Megatron global `layer_number`, which should be the right invariant for virtual pipeline parallelism
(VPP). However, VPP support is not yet available in main, so PR #2590 cannot fully validate real VPP behavior yet.

## Plan

Create a follow-up PR after VPP lands in main to validate and fix any VPP-specific RouterReplay behavior.

Contributor guide

Open the contributing guide

Research direction

Wait for VPP support to land in main, then review PR #2590 and the RouterReplay handling of Megatron global layer_number. Reproduce or add validation for real virtual pipeline parallelism behavior, and use the observed results to identify whether any VPP-specific behavior needs fixing.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
machine-learning
Issue type
Feature
Difficulty
4/5
Estimated time
3-5 days
Activity status
Quiet
Clarity
Needs clarification
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.