deepspeedai / deepspeedai/DeepSpeed

[QUESTION] Skipping examples with NaN loss

Open
#1,496 0 comments 2 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 trying to train a model on multiple GPUs with fp16 mixed precision and DeepSpeed + PyTorch Lightning. Certain pathological examples in the training set cause activations to overflow and ultimately the loss to become NaN. Is there any way to skip the backward pass for such examples? Alternatively, how is this problem usually addressed? My current hack is just to zero the loss in these cases, but that interferes with average loss computation and is generally inelegant.

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 files or tests are identified. Start by locating the DeepSpeed training path used with PyTorch Lightning, multi-GPU execution, and fp16 mixed precision, then reproduce the NaN-loss case described here. Done would require an agreed way to handle pathological examples without corrupting average-loss computation, with coverage for distributed training.

Written by the indexing model from the issue text.

Assessment

Tech stack
machine-learning, python, pytorch
Domain
distributed-systems, machine-learning
Issue type
Feature
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.