8000 intmm is being compiled inconsistently with errors. · Issue #155838 · pytorch/pytorch · GitHub
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content
intmm is being compiled inconsistently with errors. #155838
Open
@HDCharles

Description

@HDCharles

🐛 Describe the bug

We've noticed a bunch of CI failures in TorchAO

e.g. the autoquant compile test errors here: https://github.com/pytorch/ao/actions/runs/15569349188/job/43841241599

note the subsequent commit doesn't have that error so this is inconsistent: https://hud.pytorch.org/pytorch/ao/commit/b6bb7dc240b9083d105b52ee8a0393496cdbc428

What seems to be happening (link to paste of error: https://gist.github.com/HDCharles/03903b2612c727c39cd11a47594c66b0) is that a kernel is being selected that is incompatible with the actual input sizes. quantization and compilation happen without issue and then when the compiled model is run, it throws an error.

This error only started showing up as of this weekend. I'm unable to reproduce this error locally but its showing up in CI a lot.

Any help would be appreciated. I'm not sure if there's an easy way to generate a tlparse or minifier repro from CI.

Error logs

https://gist.github.com/HDCharles/03903b2612c727c39cd11a47594c66b0

Versions

see CI settings

cc @jerryzh168 @jianyuh @raghuramank100 @jamesr66a @vkuzo @jgong5 @Xia-Weiwen @leslie-fang-intel @msaroufim @chauhang @penguinwu @voznesenskym @EikanWang @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @aakhundov

Metadata

Metadata

Assignees

No one assigned

    Labels

    module: inductoroncall: pt2oncall: quantizationQuantization support in PyTorchtriagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate module

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions

      0