Pytorch2 improvements?

[anhstar]

I have installed the latest version of this fork with pytorch2 support but did not notice any improvement vs the older version.
Im using a RX 6800 with 16 gig VRAM and on both version, I get around 1.5 it/s to 1.8 it/s
Windering if anyone found a way to improve the performance ?
Thanks

[lshqqytiger]

There's no performance improvement.
I think performance-side things of PyTorch 2 are mostly related to SDKs like CUDA/ROCm/ipex.
Because DirectML is not a SDK, just a library, PyTorch is not compiled for/with it and then is not optimized for it.

[anhstar]

I see now, thanks for your answer :)