torch.random.fork_rng_state does not restore rng state about intel-extension-for-pytorch HOT 3 CLOSED

@garrett361 Thanks for reporting the issue. We will try reproducing it and come back with more details

from intel-extension-for-pytorch.

@tye1 fyi
@garrett361 Apparently torch.random.fork_rng is using the device_type as cuda by default (although the device is set as xpu). I found the rng to work as expected with the context initialized as below

torch.random.fork_rng(devices=(device,), device_type='xpu')

Output:

Results: [tensor([1.9484], device='xpu:0'), tensor([1.9484], device='xpu:0'), tensor([1.9484], device='xpu:0'), tensor([1.9484], device='xpu:0'), tensor([1.9484], device='xpu:0'), tensor([1.9484], device='xpu:0'), tensor([1.9484], device='xpu:0'), tensor([1.9484], device='xpu:0'), tensor([1.9484], device='xpu:0'), tensor([1.9484], device='xpu:0')]

from intel-extension-for-pytorch.

Ah thank you! That makes total sense; I didn’t realize there was a device_type arg. I probably should have known. Appreciate it!

from intel-extension-for-pytorch.