NVIDIA-NeMo / NVIDIA-NeMo/RL

R3: narrow routed_experts wire dtype (int32 → uint8/uint16) to cut transfer bandwidth

Open
#2,826 0 comments 0 reactions 1 assignee Claimed by @terrykong View on GitHub
Feature Speed
Dominant language
Python
Stars
2k
Forks
561
Avg merge
4d 5h
Merged PRs (30d)
145

Description

## Problem
`routed_experts` (`[tokens, num_moe_layers, topk]`) is transferred from generation to
training every step as `int32`. Expert ids are small (`0..num_experts-1`), so ~3 of every
4 bytes are wasted. Narrowing to `uint8` (≤256 experts) / `uint16` (≤65536) is a future optimization plan.

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.