Skip to content

Investigate flex-attn-masks regression on March 15 #6376

Description

@whitneywhtsang

PVC:
Image
BMG:
Image

First bad 07cb4d2: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/23123761222

* 07cb4d25ea Revert "[SYCL][XPU] Use event-less SYCL APIs whenever possible" (#6363)
* dc827fd97b Update PyTorch pin (#6357)
* 3344bd79d8 [DOCS][CI] Update DLE upto 2025.3.2 (#6358)

Last good 3344bd7 + 07cb4d2: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/23272917130
=> Regression caused by PyTorch pin update.

First bad: pytorch/pytorch@6aaa37a:
PVC: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/23318228019
BMG: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/23355670694 (hang)
Last good: pytorch/pytorch@8f67583:
PVC: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/23320596007
BMG: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/23323953337
=> Regression caused by workaround a bug in Triton floordiv

Metadata

Metadata

Type

Projects

No projects

Relationships

None yet

Development

No branches or pull requests

Issue actions