Fix is_torch_xpu_available for torch < 2.3 #31573
Conversation
| @@ -754,11 +754,13 @@ def is_torch_xpu_available(check_device=False): | |||
| if not is_torch_available(): | |||
| return False | |||
|
|
|||
| import torch | |||
|
|
|||
| torch_version = version.parse(_torch_version) | |||
| if is_ipex_available(): | |||
| import intel_extension_for_pytorch # noqa: F401 | |||
There was a problem hiding this comment.
Previously, if is_ipex_available evaluated to False we would return False. #31238 enables using pytorch for XPU as well, if the installed version >= 2.4.
Here, we still return False if the installed version < 2.4 and IPEX is not available in the environment
There was a problem hiding this comment.
Thanks.
Could also be fixed by change except RuntimeError to except RuntimeError, AttributeError, but both works for me and leave you to make the call.
There was a problem hiding this comment.
I'll leave like this, as it makes the intended behaviour a bit more explicit imo :)
amyeroberts
force-pushed
the
fix-ipex-import-torch-le-2-2
branch
from
29a2593 to
2b2b4d0
Compare
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
What does this PR do?
A recent PR #31238 updated the
is_torch_xpu_availablecheck to reflect XPU support from Pytorch >= 2.4. However, this broke backwards compatibility with versions of pytorch < 2.4 whenis_ipex_availableisFalseandcheck_device=True.Technically, this fails if the installed torch in the environment is <= 2.2, and 2.3 will still pass, as torch 2.3 has the
torch.xpumodule. However, as the intended logic is to default to the pytorch available logic for >= 2.4 and previously this evaluated to False for 2.3, this is more backwards compatible.Fixes #31563