Create a conda-forge package for trackastra

[constantinpape]

Hi team,
it would be very helpful to have trackastra available as dependency on conda-forge. This would allow other conda-forge packages to directly depend on it, which is not possible with dependencies that are just available on tif.
Creating conda-forge packages is quite easy; I could also help out with this if you want.

[bentaculum]

Hi @constantinpape - on the off-chance that this is still helpful, trackastra is now finally on conda-forge.

[constantinpape]

Hi @bentaculum ,
yes, that is indeed still very relevant! Now we can integrate trackastra as a full-fledged dependency.
Thanks!!

[bentaculum]

Just tested out microsam + trackastra, it's super cool!
FYI, I just put batched inference into Trackastra #35, gets me a 2x speedup on mps, and probably similar or more on cuda.

[bentaculum]

There is a drawback to the current conda installation: It seems that pytorch in the conda universe is currently not shipped with accelerated BLAS, neither on conda-forge nor on the pytorch channel, which makes Trackastra predictions on CPU unusably slow.

[anwai98]

Hi @bentaculum,

Thank you so much for bringing trackastra to conda-forge.

It seems that pytorch in the conda universe is currently not shipped with accelerated BLAS

Should we move this to https://github.com/conda-forge/pytorch-cpu-feedstock/issues and discuss it with the team?

[bentaculum]

Thanks @anwai98 for the pointer. I filed an issue for this.
I also tried out installing pytorch-cpu as recommended but couldn't get it to install on MacOS.