Alexey Korepanov
Alexey Korepanov
Yeah thanks, it wasn't clear until i started to dig deeper. So I think warnings or errors are super necessary here
`.masked_fill` works both for float16 and bfloat16, so i still don't get an idea of `.float()` here. Is there any possibility that forward will get not float (16/32) type?
what are your plans on cartoon images? like in makeittalk
related https://github.com/pytorch/pytorch/issues/119360
@justinchuby
Is this operator best to use? https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md#com.microsoft.Irfft
Actually I still got question about `irfft`. For my case I can't use `dynamo_export` because lstm's currently not supported there, so I have to manually implement `aten::fft_irfft` to export my...
``` import torch from onnxscript import FLOAT, script from onnxscript import opset17 as op @script() def Irfft(X: FLOAT[481, 2]): x = op.Unsqueeze(X, [0]) x = op.DFT(x, 960, axis=1, inverse=1, onesided=1)...
So, am I understand correctly that currently there is no way to create irfft node using DFT and we should wait for fix?
Yeah, I've tried it already, but currently it is only available for cuda, but I need cpu implementation :c