[XPU] Get [ZE]: 0x78000011 on torch.compile with new driver #151898
Labels
module: xpu
Intel XPU related issues
triaged
This issue has been looked at a team member, and triaged and prioritized into an appropriate module
🐛 Describe the bug
The latest Intel GPU Driver introduces breaking changes that may lead to
torch.compile
failure with the error message ofTriton Error [ZE]: 0x78000011
on Windows.For instance, one might fail in the following cases:
imf
function. It means that Triton can't find the former libdevice name, so it will raise a load failure. For example: Change__imf_powif
->__imf_pownf
and__imf_powi
->__imf_pown
intel/intel-xpu-backend-for-triton#3936select_scatter
, and it might fail on interacting with theslice
op. For example: [Pytorch Upstream][Release 2.7][Regression] Triton Error [ZE]: 0x78000011 after update to latest public driver. intel/intel-xpu-backend-for-triton#3916Currently, the Intel GPU driver is working on the process to fix the failure. During the fixing stage, it can be mitigated by rolling back to the prior version. Please see the final
Solution
section for details.Error logs
Here is an example of the error message.
Solution
A temporary solution is to roll back to the previous version of the driver.
Affected Driver
Only on Windows did we detect such an issue. The
32.0.101.6734
and32.0.101.6737
are affected.Rollback method
One could download the previous driver from Intel® Arc™ & Iris® Xe Graphics - Windows* page. One could use the previous version, like
32.0.101.6647
for example.Please click on the "perform a clean installation" checkbox when installing.
Also, note that if you failed with other issues of Triton on XPU Windows, it might be because of the setting of
LEVEL_ZERO_V1_SDK_PATH
. The old driver will need to set this manually. Please refer to Windows.md for details.Normally, it requires two additional steps:
C:'\level_zero
.$env:LEVEL_ZERO_V1_SDK_PATH = "C:\level_zero"
.Then you could run normally.
cc @gujinghui @EikanWang @fengyuan14 @guangyey
The text was updated successfully, but these errors were encountered: