Frank Schlimbach
Frank Schlimbach
I guess your conda env-file is incomplete, it should add libblitz from conda-forge.
conda-forge also provides pythran, so the env-file wouldn't need to use pip at all.
Sadly, this doesn't seem to be maintained. CMake instructions are incomplete and even when working around it the code doesn't build. Lots of C++ errors.
Thanks @rgommers for your feedback. > After reading the protocol description, I think I understand the data shape/partitioning, but I'm less sure about the "locality" information. All it says is:...
> > Interchange between dask and ray could most likely be addressed by requiring ip-addresses and disallowing dask's more flexible identifiers. However, interchange with MPI is more difficult since MPI...
> I disagree. If I run something with a single process per socket or NUMA node or GPU then I need to know which chunk belongs to which process and...
For better clarity some background: Ideally we want this be as general as possible. It turned out to be not that simple to achieve generality so we approached it by...
Sorry for the late reply, I was out. > > As for providing a process ID, why is that not a granular enough. Can you elaborate? > > It requires...
> I don't think a process ID is granular enough. I'm also generally super weary of optional pieces to a protocol, because people then rely on those optional pieces instead...
> Ray allows to pass a list of `ObjectRef` objects in a remote function where you can call `get` method to materialize data in a worker. There will be no...