deepspeedai / deepspeedai/DeepSpeed

[BUG] I can't run fp8 with pipeline parallel

Open
#5,760 2 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

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

Description

Hi, I am trying to use fp8 with TransformerEngine. I am using a version of GPT-Neox repo, which uses deepspeed.

I can get fp8 to run in my MLPs with model parallel, but when I use pipeline parallel, it gets hung up with no errors, it's just frozen once it tries to train (and goes into the DeepSpeed library call).

I was wondering if anyone else ran into this issue. Thanks!

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

Reproduce the reported freeze with fp8, pipeline parallelism, TransformerEngine, and the GPT-Neox/DeepSpeed setup. Start at the DeepSpeed library call where training hangs and compare the working model-parallel path with the pipeline-parallel path. Done means identifying the cause and confirming that pipeline-parallel training proceeds without freezing.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
distributed-systems, machine-learning
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
20/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.