pytorch / pytorch/pytorch

image diffusion model output changes for the worse after 530ff92655

Open
#181,737 12 comments 0 reactions 0 assignees View on GitHub
bot-triaged module: cublas module: cuda module: numerical-stability triaged
Dominant language
Python
Stars
103k
Forks
29.6k
PR merge metrics
PR metrics pending

Description

I'm not sure if this is acceptable numerical drift or not. I've been bisecting a change where an image diffusion workflow in SGLang produces the following images from PyTorch 2.10 -> PyTorch 2.11. This is one of the issues blocking SGLang from upgrading to PyTorch 2.11.

I bisected this change down to 530ff92655. Claude tells me that this PR makes some more cases go through cublaslt, which we do know we have some problems with recently (@eqy) Any thoughts here? @ngimel @drisspg @nikitaved

GOOD: (approx PyTorch 2.10)

Image

BAD: (approx PyTorch 2.11)

Image

cc @ptrblck @msaroufim @eqy @jerryzh168 @tinglvv @nWEIdia @csarofeen

Contributor guide

Open the contributing guide

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.