NVIDIA / NVIDIA/TransformerEngine

FSDP with FP8 is not working

Open
#1,247 2 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

bug
Dominant language
Python
Stars
3.5k
Forks
831
Avg merge
3d 11h
Merged PRs (30d)
65

Description

Image

layernorm needs to have reshape instead of view.
In pytorch base module, torch.get_autocast_gpu_dtype() is deprecated.
@timmoon10 @ksivaman @ptrendx @cyanguwa

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 by reproducing the FSDP with FP8 failure described in the attached example, then trace the LayerNorm path and the PyTorch base module mentioned in the issue. Check that LayerNorm uses reshape rather than view and that the deprecated torch.get_autocast_gpu_dtype() usage is addressed; done means the FP8 FSDP case works without the deprecation.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
distributed-systems, machine-learning
Issue type
Bug
Difficulty
3/5
Estimated time
1-2 days
Activity status
Quiet
Clarity
Mostly clear
Newbie friendliness
45/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.