deepspeedai / deepspeedai/DeepSpeed

[BUG] multi-GPU Training Frozen After Some Validation Epochs In NGC Container.

Open
#1,735 20 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

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

Description

Describe the bug
I run a validation epoch every 10 training epochs. Then after 1~50 validation epochs(500 training epochs), training was stuck(validation is finished): 1 GPU utilization is 0%, other 3GPU utilization is 100%(same as normal case) but power consumption is low.
If I do not run validation step, everything is fine.
If I do not use NGC Container, everything is fine.
I have tried single GPU validation(use engine.module in forward) or multi GPU validation(use engine, with torch.no_grad(), use DistributedSampler), multi GPU will stuck earlier than single GPU.
I have tried add barrier, but it does not work.

System info

  • 4 V100 (msrhyper itphyperdellcl1)
  • nvcr.io/nvidia/pytorch:22.01-py3

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

No repository files, tests, or entry points are named. Start by reproducing the multi-GPU freeze with and without validation in the specified NGC PyTorch container, comparing single- and multi-GPU validation and the barrier behavior; done means identifying and fixing the validation-related hang without breaking training.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
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.