Lightning-AI / Lightning-AI/pytorch-lightning

training with 500 gpus is almost 2x slower then training with 300 gpus in my case

Open
#17,715 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

bug distributed ver: 2.1.x
Dominant language
Python
Stars
31.4k
Forks
3.8k
Avg merge
6d 7h
Merged PRs (30d)
6

Description

### Bug description

In my case, training with 500 GPUs is almost 2x slower than training with 300 GPUs. Does anyone know what might be the causes and how to solve them? For more information, we use NVIDIA Nsight, and it has been observed that there are numerous broadcasts. Thank you in advance to anyone who can provide insights or information on these topics!

### What version are you seeing the problem on?

pytorch-lightning==1.4.2

### How to reproduce the bug

base on https://github.com/CompVis/stable-diffusion

### Error messages and logs

```
# Error messages and logs here please
```

### Environment

Current environment

torch==1.12.1+cuda113.cudnn8.3.2
pytorch-lightning==1.4.2

### More info

_No response_

cc @justusschock

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

Start with the stable-diffusion reproduction linked in the issue and compare the 300-GPU and 500-GPU runs using NVIDIA Nsight. Review the reported broadcasts and the supplied PyTorch and PyTorch Lightning versions; the issue is done when the scaling slowdown is explained and a verified remedy or actionable diagnosis is documented.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
distributed-systems, machine-learning, performance
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
15/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.