[IPEX] Fix xpu generator

[Nuullll]

Description

• Checks whether torch.Generator supports xpu device before hijacking. "xpu" is supported starting from torch 2.1 and at the same time torch.xpu.Generator API is removed starting from IPEX 2.1.
• Properly handles the device arg when it's a str. Fixes issue [Bug]: Can't use IPEX with Token Merging #14434

Screenshots/videos:

Using token merging:

txt2img works as expected:

Checklist:

• I have read contributing wiki page
• I have performed a self-review of my own code
• My code follows the style guidelines
• My code passes tests

[zakusworo]

Hi, i'm succesfully running IPEX 2.1 from your repo here, But, when i'm use Vlad's System Info Benchmark, it's gives error about NaN in benchmark step 3,4,5 (16,32,64 batch i think)

But, when i'm downgrade to IPEX 2.0, it's fine, no error

[Nuullll]

But, when i'm use Vlad's System Info Benchmark, it's gives error about NaN in benchmark step 3,4,5 (16,32,64 batch i think)

Extensive benchmark tests batch size 1, 2, 4, 8, 16.
I didn't verify IPEX 2.1 much for SD workloads. What I know is that IPEX 2.1 perf is ~20% worse than IPEX 2.0, so I stick to IPEX 2.0 as well for now.

[zakusworo]

But, when i'm use Vlad's System Info Benchmark, it's gives error about NaN in benchmark step 3,4,5 (16,32,64 batch i think)

Extensive benchmark tests batch size 1, 2, 4, 8, 16. I didn't verify IPEX 2.1 much for SD workloads. What I know is that IPEX 2.1 perf is ~20% worse than IPEX 2.0, so I stick to IPEX 2.0 as well for now.

oh, okay. IPEX 2.0 is fine. with your IPEX extension, my A770 can pass 20 it/s on benchmark.

the only thing i concern about IPEX 2.1 (with torch 2.1 ofc), is to be able use FP8 Weight Storage form memory management. but it's still broken for IPEX 2.1, with "torch not compiled with CUDA" error.