deepspeedai / deepspeedai/DeepSpeed

[QUESTION] How to do p2p send/recv a tensor with stride info in pipeline parallelism

Open
#1,609 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
43.1k
Forks
5k
Avg merge
4d 15h
Merged PRs (30d)
112

Description

I'm curious about how deepspeed pipeline strategy handles the stride information when do p2p communication. I've seen DDP handle these stride issues with 'bucket_view' and 'as_strided' api , so the tensor with specific stride could be reconstructed correctly. But I can't find the similar operations in p2p communication implementations.
Does anyone know the reason for this? I would be very grateful if anyone can give some answers or related ideas.

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

Review DeepSpeed's p2p communication implementation and compare it with the DDP behavior involving bucket_view and as_strided described in the issue. Done means documenting whether stride information is preserved or reconstructed during pipeline-parallel tensor transfer, and explaining the reason.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
distributed-systems, machine-learning
Issue type
Documentation
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
20/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.