Copy headers to prefix/include

[ELundby45]

pybind11 rebuild

Changes

• Bump build number
• Copy header files to %PREFIX%/include/pybind11 this matches the functionality of the pip package and allows me to remove a hack that was required here: CI SKIP Pkg 6001 win cf pytorch-feedstock#53 (comment)

[JeanChristopheMorinPerso]

@ELundby45 From what I see in https://conda-metadata-app.streamlit.app/?q=pkgs%2Fmain%2Fwin-64%2Fpybind11-global-2.13.6-py39h214f63a_0.conda, the headers should already be there no?

[ELundby45]

@ELundby45 From what I see in https://conda-metadata-app.streamlit.app/?q=pkgs%2Fmain%2Fwin-64%2Fpybind11-global-2.13.6-py39h214f63a_0.conda, the headers should already be there no?

I don't think so? I see:

Library/include/pybind11/attr.h

and 

include/pybind11_global/pybind11/attr.h

But what I need is:

include/pybind11/attr.h

[JeanChristopheMorinPerso]

include/pybind11/attr.h is "broken" though... Files on Windows should go in Library (in the conda ecosystem).

[danpetry]

a sub-system of pytorch is expecting pip-like include directories. alternative is to patch pytorch to expect more conda-like directory structure. I prefer this solution as it's more broadly applicable, but either will do imho (although patching pytorch will be harder)

[danpetry]

it occurs when you define custom pytorch operators, possibly specific to automatic gradient calculation. The error hasn't cropped up in the general case.

[JeanChristopheMorinPerso]

The thing is that all packages in uou ecosystem look for stuff in Library. So the problem is not with pybind11. If pytorch looks for stuff in $PREFIX instead of LIBRARY_PREFIX on Windows, then something is wrong and other stuff is probably broken.

[JeanChristopheMorinPerso]

Basically, my concern is that merging this change could potentially hide very important problems in our ecosystem.

[danpetry]

in the rest of the ecosystem or just pytorch?

Yea, probably best to make sure pytorch is looking in the right place at runtime. Otherwise the jit functions might not work somewhere else either.

[JeanChristopheMorinPerso]

in the rest of the ecosystem or just pytorch?

Saying ecosystem was intentional. If something in pytorch is not able to find the right files where they should usually be, it could well be an ecosystem problem or a pytorch problem. By "fixing" the problem in pybind11, you risk making the problem bleed into potentially the rest of the ecosystem.