Conda install cuml failure about cuml HOT 7 CLOSED

For example the current README install process will cause anyone that wants to use CUDA 10 issues as it locks the CUDA version to 9.2. Also FWIW we do not lock CUDA versions in the cuDF conda package so I would prefer to stick with that standard.

from cuml.

Running conda install -c pytorch faiss-gpu cuda92 prior to the cuML install resolves the issue.

from cuml.

Thanks @beckernick , we've just updated the README to reflect the need to install these two packages before cuml.

@raydouglass do you know if it's possible in the cuml conda package, to make a dependency on these two pytorch channel packages?

from cuml.

I support the faiss-gpu as a dependency, but think we need to not make CUDA a conda dependency. Given that users can use any local version of CUDA 9.2+ or any of the conda packages that install the CUDA toolkit, locking it to a particular version seems too restrictive and will cause issues for most developers. Documentation should be updated for the requirement around CUDA and the options on how to install while the other dependency can and will be added to the conda package in the morning.

from cuml.

Running conda install -c pytorch faiss-gpu cuda92 prior to the cuML install resolves the issue.

Hi,

I think that the solution above should be detailed here:

https://anaconda.org/rapidsai/cuml

If not, some developers will not be able to install it.

Regards,
Miguel

from cuml.

Thanks @miguelangel . @mike-wendt do you think it makes sense to update the conda page given your concern about locking into cuda versions?

from cuml.

I believe this has been resolved

from cuml.