ef6d75b3ee4d053c878995c7acbddf49937a37d9
* Make sure the target device is used for MSCCL
* Enable single process mode by default to use MSCCL in MT
* Create a per-rank state when GPUs share a thread
[ROCm/rccl commit: 03a3ef3c34]
Languages
C++
67.5%
C
20.6%
Python
6.6%
CMake
3.4%
Shell
0.6%
Other
1.1%