NVIDIA / NVIDIA/TransformerEngine

[JAX] Upstream checkpointing of quantizations in TE/JAX to MaxText

Open
#2,658 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

field-request fp4 FP8 jax
Dominant language
Python
Stars
3.5k
Forks
831
Avg merge
3d 11h
Merged PRs (30d)
65

Description

MaxText upstream already supports TE quantization, but for optimal performance for NVFP4 and MXFP8, quantization checkpointing support should be upstreamed to MaxText to prevent 2x quantization in fwd and rematerialized in backward,

This work was initially started here but needs continuing: https://github.com/AI-Hypercomputer/maxtext/pull/2773

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.

Research direction

Start by reviewing the linked MaxText pull request #2773 and the existing TE quantization support in MaxText. Trace how NVFP4 and MXFP8 quantizations are checkpointed during forward and rematerialized backward passes. Done means the support is upstreamed so these paths avoid the reported 2x quantization.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
machine-learning, performance
Issue type
Feature
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.