GitLab

* first step, everything compiles

* fix rebuilds; skip cuda version check for rocm

* use macro for __shfl_up_sync __shfl_down_sync

* add BFloat16 support for ROCm and CUDA

* add USE_ROCM definition to setup.py

* flake8 fixes

* change other to other.size() for compatibility

* fix typo

* Add parameter to `scatter_[log_]softmax`

* Update softmax.py

* Format code.

* compile with half

* Fix

* fix

* rename

* update

* update

* update

* update

* typo
Co-authored-by: rusty1s <matthias.fey@tu-dortmund.de>

Hi,
I think the top index of the sum is not right in the documentation of segment_csr. I might be wrong though!

torch.version.cuda can still exists when it was installed CPU-only.

Debugging for the corner cases involving -inf.