Obsolete links in Google search index

[vadimkantorov]

https://www.google.com/search?q=onnx+dtypes

finds https://onnx.ai/onnx/api/mapping.html, but it's 404

Removing the obsolete links from Google search is non-trivial, but possible. These links in Google results are quite nasty and make it harder for novices...

[justinchuby]

It was just removed yesterday. Hopefully google will see the 404 soon and take it down

[justinchuby]

That said are you looking for anything about onnx dytpes? What info would be good to have in documentation?

[vadimkantorov]

What info would be good to have in documentation?

I was looking to see if anything has changed wrt complex / tuple / tile dtypes support...

[vadimkantorov]

Hopefully google will see the 404 soon and take it down

Unfortunately, Google can take arbitrarily long to detect this, especially if this page is not linked anywhere anymore, some related discussion in:

• [docs] Google finds docs pages that give 404 pytorch/pytorch#67237

[justinchuby]

I checked again and Google seemed to have picked it up. For future readers: https://onnx.ai/onnx/intro/concepts.html#element-type

[vadimkantorov]

Thanks! Interesting that complex64 dtype/2-tuple layout is there. So is it that it's supported, but only for a limited number of ops?

[justinchuby]

(Limited=0). It seems no actual work has been done to expand operator support.

[vadimkantorov]

I see. Might be worth adding a note to the docs about it e.g. even a new page "Complex numbers and complex operators support"?

Haha, I think people are often searching for the "currently working" way of exporting complex ops (dft/stft) to onnx/ort - playing nicely with dynamic shapes etc :)

[justinchuby]

Yeah. We just released a fix to improve dft support in https://github.com/microsoft/onnxscript/releases/tag/v0.2.5, but from testing we are still missing a scaling factor for the inverse case (microsoft/onnxscript#2209). Any help is appreciated!