meta-pytorch / meta-pytorch/data

Add option to disable random seed sharding and syntonization for DistributedReadingService

Open
#1,048 0 comments 0 reactions 1 assignee View on GitHub

@ejguan is already working on this.

Since Feb 24, 2023.

Dominant language
Python
Stars
1.3k
Forks
179
Avg merge
6d 1h
Merged PRs (30d)
2

Description

🚀 The feature

This is a customer request for inference, no synchronization is needed because no all_reduce is needed. And, normally inference doesn't require shuffling and requires high throughput. Sharing random seed becomes redundant in this scenario.

We might want to add two more arguments to enable/disable synchronization.

Motivation, pitch

Provide better performance for inference using DataPipe + DataLoader2

Alternatives

Add a special ReadingService for distributed inference.

Additional context

No response

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.