mlcommons / mlcommons/storage

Turn off the CRC check in DLIO's use of numpy.load()

Open
#223 0 comments 0 reactions 1 assignee View on GitHub

@wvaske is already working on this.

Since Jan 9, 2026.

Dominant language
Python
Stars
203
Forks
67
Avg merge
20m
Merged PRs (30d)
8

Description

During v2.0, Jon Flynn reported a whole set of things that DLIO was doing that seemed to be not required by the benchmark and that consumed a lot of CPU cycles on the client. He said that with these changes, he got 3x higher benchmark scores on the same client hardware.

I went from:
Base Code: ~7 H100s passing
Removed CRC: ~11.5 H100s passing
Final Code: 24 passing at 95%
all in Buffered mode
The O_DIRECT pathway avoids the CRC check as well, which is one reason its faster.

Image

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.