NVIDIA / NVIDIA/TransformerEngine

[JAX] Move quantization to before AG in MoEBlock

Open
#2,998 0 comments 0 reactions 1 assignee View on GitHub

@jberchtold-nvidia is already working on this.

Since Jul 13, 2026.

Dominant language
Python
Stars
3.5k
Forks
831
Avg merge
3d 11h
Merged PRs (30d)
65

Description

Current MoEBlock initial impl follows the same A2Av flow in maxtext where for FSDP, weights are AG then quantized before groupedGEMM happens. This is wasteful as the quantization kernel has to work with fsdp times the amount of data needed, which makes it about fsdp times slower. Need to move it to happen before AG.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.