pytorch/builder

`numpy v2.0.0rc1` on pytorch/ conda channel problem

k-dominik opened this issue ยท 6 comments

Dear maintainers,

recently (around April 3rd, 2024) numpy v2.0.0rc1 was added to the pytorch conda channel.
Many users source pytorch from the pytorch channel in addition to whatever else they need from conda-forge. It is also common practice to prefer "strict" dependency resolution, making pytorch a higher priority channel than conda-forge. This hides package from lower priority channels if they are present in the pytorch channel. Effectively, this now means users cannot get any numpy==1 package anymore if pytorch is included with higher priority.

I'm not sure about the reason numpy was uploaded here (probably for pytorch-nightly), but it would be preferable to remove the main label from that package here and add it under some other label (or a different conda channel all together, light pytorch-nightly if this is for experimentation).

I reported a similar issue with libjpeg-turbo (#1679) in the past (which still persists).

It would be great to get this sorted :)

CC: @jakirkham

cc @rgommers (for awareness)

Another label makes sense to me, since numpy 2.0.0rc1 is needed at build time but not at runtime. This is also what conda-forge did, see conda-forge/numpy-feedstock#312 and https://anaconda.org/conda-forge/numpy/files.

@atalman does this plan sound reasonable to you? If so, what are the next steps here?

This should have been removed now as we don't need it in this channel (we used another channel for our build dependencies that require it)!
Can you confirm it is working as expected now?

Hi @albanD,

thank you so much for sorting this out - it works for me now as before.

Any chance on a similar fix for libjpeg-turbo on linux-64 (#1679)?

Cheers
Dominik

Thanks all! ๐Ÿ™

Let's go ahead and close this out as resolved then

We can discuss libjpeg-turbo in issue: #1679