libjpeg-turbo on pytorch conda channel shadows conda-forge with old version

[k-dominik]

Hello!

I came across an issue sourcing pytorch from the pytorch channel, along with a lot of other dependencies from conda-forge. I'm using strict channel priority to get more predictable environments. For linux there is an issue that can occur, if any of the packages depend on libjpeg-turbo. "Current" version, according to conda-forge-pinnings of libjpeg-turbo is 3. So any recent versions of packages built in the conda-forge ecosystem depend on that version. The libjpeg-turbo package on the pytorch channel is only available in version 2.0.0. As a result, environments might not be solvable (in strict mode).

I'm not sure what a good solution to this is. I assume this package is there for a reason, so removal might not be an option. For me it could work to add the 3.0.0 libjpeg turbo version (from conda-forge) to the pytorch channel. However, this could be a maintenance nightmare, as libjpeg-turbo would've to be kept in sync. Removal of the main label would be preferable...

Cheers
Dominik

[jakirkham]

It looks like defaults has libjpeg-turbo version 2.0.0

Would it be possible for the pytorch channel to use those? If so, could libjpeg-turbo be dropped from the pytorch channel?

[albanD]

cc @seemethere I guess this one is a depency from domains? Who is the most up to date about which version we need here?

[jakirkham]

IIUC this is connected to FFmpeg, which is also on the pytorch channel. Maybe it is worth auditing the channel for common dependencies like this and evaluating whether they can be dropped in favor of defaults packages

[k-dominik]

Hello dearest maintainers,

I'm terribly sorry to bump this issue again, but it's currently blocking some dependency updates for us. Any chance to get a resolution/decision on this package?