Repository navigation
Xe2_Or_Later flag doesn't handle Xe3 architecture - it treats them as Unknown architectures. #5946
Description
Several test cases in PyTorch that use torch.xpu._sleep() are gated by @uniitest.skipIf(not Xe2_Or_Later). It makes sense as torch.xpu._sleep() is not supported for Xe1 architecture, for example PVC.
The problem is that effectively the Xe2_Or_Later treats Xe3 as Unknown and for Xe3, the Xe2_Or_Later flag evaluates to False, while the name suggests that it should be True. That is caused by missing entries for newer hardware in _DEVICE_ID_TO_CODENAME dict and _CODENAME_TO_ARCH dict. The mappings are only for PVC and BMG, while newer hardware is treated as Unknown, which currently evaluates always to false in Xe2_Or_Later flag.
This task issue is to consider solution for that problem. I suppose that all the supported hardware ids and codenames should be added and flag evaluation should raise rather than silently categorizing the unknown device type to not Xe2_Or_Later.
The device ids are taken from compiler header file here: FILE
The Xe2_Or_Later is defined in THIS FILE
Activity
[-]`Xe2_Or_Later` flag doesn't handle Xe3 or Xe4 architecture - it treats them as Unknown architectures.[/-][+]`Xe2_Or_Later` flag doesn't handle Xe3 architecture - it treats them as Unknown architectures.[/+]daisyden commentedon Nov 9, 2026
label-issue: #5946
Root cause:
torch/testing/_internal/common_xpu.py:26-56maps only PVC/BMG ids and codenames; unknown (Xe3) devices fall toXPUArch.Unknown(:69), soXe2_Or_Later(:72-74) is False on Xe3. Test-harness defect in pytorch.typeTaskTaskpreserved.testtest: utXe2_Or_Later; no e2e/oob signal.modulemodule: infracommon_xpu.py:26-74, not an op or runtime surface.priorityLowfunctionalityXe2_Or_Laterevaluates False on Xe3, wrongly skippingtorch.xpu._sleep()tests.Applied to this issue. Generated by a local
label-issuerun.