Markus Goetz

In this case you may simply add them locally as already implemented and working

That would be rule 3.) if both have a minor axis, than the one with the larger minor axis will be redistributed

Couple more things that came to my mind that have to be considered: * __binary_op should ensure that the comms of the two operands match. We cannot perform the operation...

Here is a great paper summarizing formats and operations on sparse tensors. I would go with a COO format first where the set of entries is roughly evenly distributed http://groups.csail.mit.edu/commit/papers/2016/parker-thesis.pdf

PyTorch currently implements fp16 support for GPUs only. This issue will have to be resolved after HeAT has device support. First implementation idea is to check the device the HeAT...

A first (perhaps incomplete) list of requirements: * There are pre-defined benchmarking test cases, measuring ** Strong-scaling compute/memory/time/speed-up ** Weak-scaling compute/memory/time/speed-up ** Different backends * Benchmarkings are performed continouosly on...

It was not done as part of the PR. It is still an open todo

It was not closed by #757. At the moment we state in individual algorithms that we have used/adapted/been inspired by code of other libraries. It would be generally good to...

More or less, it would be good to make a copy of that in the documentation as well. Moreover, it would probably good to use the split tensor figure from...

Essentially text writing - it would be an explanation of the (relaxed) bulk synchronous processing (BSP) model including the graphic from the paper. Mainly boils down to saying that there...